A significant part of my education has come from listening to podcasts while I’m walking, running errands, or doing work that doesn’t need my full attention. I listen to shows that are about current AI tools, marketing, search engines, how people find a business when they ask an AI for a recommendation. When someone says something I want to return to, I take a screenshot. The idea is to collect source information that I expect to refer back to. Back at my computer, I use it to find the episode’s transcript, capture that with a browser clipping tool, and send it into Obsidian.

You may have heard this kind of collection called a second brain. At it’s most basic level, Obsidian is an information database, a place to bring together what I’ve written about, what I learn, what I’m working on, and the connections between them. The collection lives in a folder on my computer that Obsidian calls a vault. Its linked notes use Markdown, ordinary text files with a few conventions for headings, links, and formatting, which means other applications can read them too. I also keep spreadsheets, PDFs, Word documents, and other files in that folder structure. They contain information I want available when I build the wiki, the organized, linked account of what I’ve collected and learned. Obsidian itself supports other file types on a limited basis but more importantly, the AI tools I use to process the collection can also draw from XLSX spreadsheets, DOCX documents, and other sources when I give them access and the tools needed to read those formats.

What belongs there

Podcasts are one source. I also save useful web pages, meeting transcripts, session notes, and my published work. A session note can preserve something I noticed while working. The collection also includes results from my AEO tools, along with Google Search Console and analytics data. Reports can sit alongside the conversation in which I interpreted it and the decision I made afterward.

Conversations with AI deserve a spot not only because they reflect my thought process, but also because they’re so easy to lose. I can spend an afternoon working through a problem, reject several approaches, settle part of it, and leave another part open. Returning to the chat later requires remembering which conversation held the useful reasoning, and chat titles are often just a step above useless. Add in that some things live in projects, some are ‘local’, and the need for a way to recall work becomes obvious.

I use a tool I’ve built called Project Memory Generator to turn those conversations into documents I can keep. It records the problem, the reasoning worth preserving, decisions I actually made, and questions I haven’t resolved. A suggestion I discussed with an AI needs to remain a suggestion if I never agreed to it, otherwise it gains undue importance in the wiki.

Capturing these sources and incorporating their contents into the wiki are separate steps. In the second step, known as ingesting, I use AI with instructions about how to read and classify the material, identify useful relationships, and distinguish it from what’s already there. The source might be a Markdown transcript, a PDF report, or a spreadsheet of results. I want the useful information from each available in the wiki, with enough reference to the source that I can go back and examine it. I also want cross-referencing between related data.

In Obsidian’s human-directed workflow, I link notes to one another as I recognize relationships. Bringing ChatGPT, Claude, or another AI tool into that work expands what I can ask of the collection. I can compare sources, proposed connections, and articulated relationships even if yet explained my thought process to the chatbot.

The relationship still needs a reason, of course. Some of that context comes from chat summaries, others come from my explicitly saying this connects to that in this way because of ‘reasons’. Without that, two (or more!) sources mentioning AI may have almost nothing useful in common. A marketing conversation and a note about how I evaluate evidence may have a great deal to say to each other, even though I saved them for different reasons. I build the vault to explore what connects those ideas.

Obsidian is great for machine access, but when it comes to human searching, I found it lacking, so I built a simple search interface. That’s the thing with systems that other people built; they can be a great foundation, but they can and should be modified to make them personally useful, otherwise it is at best, an underused resource and at worst, a waste of time.

Where Notion fits

I also keep documents in Notion, which gives me pages I can maintain and make available to connected AI tools. My writing guide lives there, along with the document that establishes basic facts about me and my business. I refer to these as canonical sources, meaning I’ve identified which document should govern a particular kind of information. If an AI needs to know how I write, I want it to read the current writing guide. If it needs a fact about my business, I want it to consult the document I maintain for that purpose. I don’t want an old draft that happens to sound confident to become the authority, and I really don’t want any AI tools to extrapolate hard facts about my business. These canonical source files are an important way of ensuring that doesn’t happen.

Some authoritative documents live in Obsidian, and I publish read copies of them to Notion. My vocabulary reference, the vault digest, and the decision log work this way. The digest gives an AI a compact account of active work and open questions; the decision log preserves choices and their reasoning. Copying everything would leave me with another collection to sort through, and more opportunities for an outdated copy to disagree with the document I actually maintain. In this arrangement, I use Notion for a selected set of accessible references, while the vault holds the broader material from which I can develop and examine ideas.

So why bother with Notion? It all comes down to how I use AI. Oftentimes, that’s at my desk, but I also use my phone or my laptop. Notion gives me another access route with the most important and most frequently used information. AI tools with access to the vault on my computer can read the material there. Tools connected to Notion can read the pages I’ve made available through that connection. I still need to direct them to the relevant sources. Having a document somewhere in my system doesn’t establish that the AI will read it for the task at hand.

A day’s work on how I decide

I recently spent a full day building a guide called How I Decide. It describes how I evaluate evidence, tradeoffs, uncertainty, and the point at which an interesting possibility becomes something I’m committing time to. The guide draws on AI records of earlier decisions made. I also shared philosophical context with ChatGPT Sol during it’s creation. One principle tells an AI to check how a result was produced before recommending that I change strategy. The supporting records include an AI-search query that returned painters when I was trying to assess photography visibility, and a monitoring change that made the numbers before and after it unsuitable for a direct comparison. Those records now support a general instruction in the decision guide. I can give a future AI the principle, the evidence behind it, and the limits of its application. I also keep the reasoning made that possible.

Deciding what to carry forward

Some podcasts had been excluded from the material carried forward as “not photography-specific” when I’d been using the pre-built ingest instructions (the Karpathy method). I’d saved them because something interested me, and photography alone was too narrow a description of why they might belong. I needed to make those reasons more legible to the AI. A source can be worth retaining because I want to challenge it. Something can deserve a place before I know which project it belongs to. And once it’s in the collection, I need the AI to distinguish the speaker’s position from my own.

Further, not everything deserves to receive equal weight. An article I’ve published represents a more considered position than a content pipeline full of things I might write. The pipeline can still tell an AI what I’m considering. A recent note might record why I’m reconsidering something I published. The question is what each source can establish, and whether I’ve actually adopted the position it describes.

On September 4, I began piloting the decision guide with a corresponding set of ingest instructions. Those instructions tell the AI to distinguish facts, observations, proposals, and decisions, and to request review when material appears valuable but has no clear home. They also call for a report of what it inspected, what it changed, and what it withheld and why. Those reasons are important to understand if the system is working as I intend and they become helpful down the road when I want to consider how my decisions and processes evolve over time.

This is part of what I mean by persistent identity: the voice guide, project knowledge, and feedback loop working together to carry accurate context across sessions. I supply the material, review how the AI handles it, and make corrections specific enough to inform later work. The day spent on the decision guide belongs to that work, just as capturing a useful conversation does.

For your own business, one saved conversation may be a useful place to begin. Keep the reasoning behind a decision, along with the decision itself, and give it to an AI when the subject comes up again. Then examine what it does with the distinction between what you considered and what you chose.

I want mine to understand my work well enough to help with the next thing.

The photograph was Napoleon Sarony’s portrait of Oscar Wilde, and the objection probably sounds more familiar now than it should. A photograph came from a machine. The camera performed the mechanical act. How, then, could the person standing behind it claim authorship?

Sarony had an answer. He had selected Wilde’s pose and expression, arranged the costume and accessories, chosen the lighting, and composed the scene. The lower court found that the photograph came from his “original mental conception,” and the Supreme Court concluded that photographs could receive copyright protection when they represented the “original intellectual conceptions” of the author. The camera made the photograph possible. Sarony made the decisions that produced this photograph.

We seem to be having a version of that argument again, except this time the machine is considerably more complicated. Anthropic recently began marking content created with newer Claude models, including Content Credentials for supported files and a separate watermarking approach for text. The details matter, and some of them are still being developed, but the basic purpose makes sense. If AI-generated material can move through the world looking like material that wasn’t generated by AI, giving people some way to know that a model participated is useful.

I don’t think Anthropic is wrong to do it. I think we’re stopping too early. We’re building increasingly sophisticated systems that can tell you which machine touched a piece of work while doing nothing to tell you which person stands behind it, and that is where I think they went wrong.

We’ve done this before

Open a Word document and look at its properties. Mine will say that Michael Kloth created it. It may show who last saved it, when it was created, when it was modified, how many revisions it has been through, which program created it, and other pieces of a record that most of us never look at unless we have a reason to. None of that proves I wrote every (or any) sentence. The field establishes a useful baseline that I’m the person associated with it. The application that created the file is another fact about the document, but Microsoft Word doesn’t become the author because I used it.

Photographers already work with a richer version of this. When I make a photograph with my Canon camera, the file carries information about the camera, lens, exposure, ISO, focal length, date, and other details of the capture. When I bring that file into Lightroom Classic or Photoshop, I can add my name, copyright information, contact information, and other IPTC metadata. Adobe’s Content Credentials work goes further, allowing creator identity and editing activity to travel with supported files in tamper-evident form.

That’s useful to me even before anyone else sees it. If I’m trying to remember how I handled an image five years from now, the record has value. If I’m working in photojournalism, where the difference between an acceptable adjustment and an alteration that changes the meaning of a photograph can matter enormously, the record has another kind of value. If I’m making fine art and want to replace half the sky, nobody needs the metadata to stop me. A provenance system can preserve enough information that a person can make the judgment appropriate to the work.

The strange direction we’re taking with AI

Generative AI should make that kind of record richer. Consider the process behind this article. I didn’t wake up, type a one-sentence prompt into ChatGPT, and receive the thing you’re reading. The idea started when I learned about it while I was walking my dog, then I started doing research to understand the specifics. Further, I worked through it in conversation which included rejecting some offered directions. I changed the central argument because I am a photographer. I brought knowledge that I gained in graduate school, then reinforced when I taught the history of photography at Washington State University. I could do that because that’s the field I know well enough to recognize the parallel. Neither Claude, nor ChatGPT suggested that line of reasoning at all, but if I’d done any part of this writing in Claude, it would have made that a part of the record.

A system capable of participating in all of that can preserve a record far more useful than “Revision 12.” Storage isn’t the meaningful constraint it once was. The private record could contain drafts, notes, source relationships, substantial revisions, tool participation, decisions that disappeared from the published version, and enough information for me to reconstruct how a piece developed six months or six years later An author writing a novel might want exactly that. Why did this character make that choice in chapter nine? The answer may have disappeared from the final manuscript, but it might still exist somewhere in the development record.

That doesn’t mean I want to publish the record. Photography already gives us the model for that too. I can strip EXIF metadata from an exported image. I can preserve only copyright and contact information. I can share more when provenance matters and less when it doesn’t. Word lets me remove document properties before a file leaves my computer. The richest possible creation record and the richest possible public record don’t need to be the same thing. I want the first one for myself. For the second, I want control.

Text makes the second part harder than photography does, and that matters because Word only gets us so far. A photograph usually leaves my workflow as a file, and the file itself is what I publish. The Word document holding this article won’t be the thing you read on Substack. The text will leave that container and become markup on a web page. The author field, revision history, editing time, and anything else attached to the .docx stay behind on my hard drive. Signing the authoring file is easy. Keeping that identity attached when the writing leaves the authoring file is the problem and any serious system for text provenance has to cross that boundary.

The missing assertion

The public version could still be much simpler. Michael Kloth stands behind this work. These tools participated in making it. Here is the credential that connects those statements to this version of the work. That isn’t a fantasy standard waiting for somebody to invent it. The Creator Assertions Working Group has developed an Identity Assertion specifically designed to let a named actor prove control over a digital identity and document that person’s relationship to a C2PA asset independently of the software or hardware generating the C2PA claim. CAWG also has a Metadata Assertion for binding things photographers already understand, including EXIF, IPTC, and XMP metadata, into a tamper-evident record. Its Identity Assertion 1.2 became a ratified specification in December 2025. The word independently matters.

The ordinary C2PA claim signature comes from the claim generator, the vendor’s software or hardware. CAWG separates the named person’s identity claim from that vendor-generated claim. It doesn’t eliminate certificate authorities or every other dependency in a trust system, but it means the application company doesn’t have to be the only party capable of saying who stands behind the work.

Nikon supplied an unexpectedly good demonstration of why that distinction matters. In August 2025, Nikon launched its Authenticity Service for the Z6III. A vulnerability surfaced almost immediately, the service was suspended, and Nikon subsequently told early users that the digital certificates issued during the affected period would be invalidated. The Content Credentials already attached to photographs under those certificates could no longer be used as proof of provenance. Nikon eventually restored the service, which is available today. The photographs hadn’t changed, the photographers hadn’t changed, something upstream in the trust system had.

That’s the danger in treating vendor-controlled attestation as sufficient. I don’t expect any provenance system to be indestructible, and revocation exists for good reasons, especially when somebody finds a vulnerability. But my claim that I made my work shouldn’t depend entirely on the continuing validity of a software or camera company’s claim about it. The infrastructure is moving toward something better. The individual creator still feels like the missing link.

Copyright has always been about the person

Copyright sits surprisingly close to the beginning of the American system. Article I of the Constitution gives Congress the power to secure exclusive rights to authors and inventors, and Congress passed the first federal copyright law in 1790. The details have changed repeatedly since then, but Congress hasn’t exactly developed a reputation for getting ahead of technological change.

Streaming had already remade the economics of music by the time Congress passed the Music Modernization Act in 2018. The law created a blanket licensing system for digital music services and a new Mechanical Licensing Collective, with formal roles for publishers, professional songwriters, digital services, and songwriter advocacy organizations. Technology created the problem, but a lot of organized interests had to agree on enough of a solution to get something through Congress.

Photographers have their own example that’s probably more useful here. For years, suing over a relatively modest copyright infringement meant going into federal court, which isn’t much of a remedy when the cost of enforcing the right can exceed the value of the infringement. PPA, ASMP, and other creator organizations spent years pushing for a small-claims alternative. PPA described a push lasting more than a decade and, during Senate meetings in 2020, said it was speaking on behalf of 30,000 photographers. ASMP documents more than 38,000 emails and tweets to members of Congress, testimony before Congress, Hill meetings, and coalition work with PPA, NPPA, the Copyright Alliance, the Authors Guild, and other groups. Congress finally passed the CASE Act in 2020, creating the Copyright Claims Board inside the Copyright Office.

One photographer in Tucson wasn’t going to create that system. Thirty thousand photographers represented by an organization had a better chance, and it was still a heavy lift for the people involved. That matters now because copyright law is again trying to describe authorship while the technology underneath creation changes quickly.

As of August 2026, the Copyright Office still points to its January 2025 report as its current analysis of copyrightability and generative AI. Its position isn’t that using AI poisons a work. AI-assisted work can receive copyright protection when a human determines sufficient expressive elements, including through creative selection, arrangement, or modification. Material determined entirely by the machine doesn’t receive the same protection, and prompts alone generally don’t establish authorship. There are unresolved questions inside that standard, but the basic principle should sound familiar to photographers. A machine can participate in making an image without becoming the photographer.

The trust problem

We’re trying to answer those authorship questions while dealing with a second problem at the same time. People don’t trust a lot of what they see. Some of that distrust predates generative AI by years. Some of it comes from synthetic photographs, synthetic video, fake quotations, fabricated news, automated social accounts, and enough low-effort generated writing that “AI slop” has become ordinary vocabulary.

LinkedIn, which has no particular incentive to discourage people from creating content for LinkedIn, said in May that it was seeing more low-effort AI-generated material that could sound polished while lacking a unique perspective or substance. LinkedIn’s own framing is close to the distinction I’ve been making: using AI to help write is fine, but the post still needs to represent the person’s voice and perspective. The company says its early testing correctly identified generic content 94 percent of the time. That sounds encouraging until you start asking the questions a creator has to ask. LinkedIn didn’t publish a false-positive rate in that announcement. It didn’t tell us how often an original piece gets caught by the same system, or what recourse a creator has when a classifier decides incorrectly.

This is the same underlying problem showing up from another direction. AI detectors ask whether a model probably produced something because that’s a question they can attempt to measure. Provenance systems tell us that a generative tool participated. Those are useful facts and I’d rather have them than not have them, but neither fact identifies the person responsible for what I’m reading. If the public record says Claude, OpenAI, Gemini, Grok, Photoshop, or Firefly, we’ve identified a production tool. We’ve done roughly the equivalent of identifying my Canon camera while leaving the photographer field blank. It’s no wonder that people treat the work as though a machine made it.

Photographers already know ISO

Photographers have an accidental head start on understanding why the standards work matters. ISO is the International Organization for Standardization. We interact with one of its standards constantly whether we think about that organization or not. The ISO speed setting on a digital camera comes from a standardized system for assigning and reporting camera sensitivity. The current digital-camera standard is ISO 12232. We don’t need to know the committee structure every time we change from ISO 100 to ISO 800. We need different cameras, software, labs, manufacturers, journalists, archivists, and everyone else in the chain to have enough shared agreement that the number means something.

Content Credentials are already much further into that process than “somebody should come up with a standard.” ISO’s current project page lists ISO 22144, *Authenticity of information, Content credentials*, under ISO/TC 171/SC 2 at stage 30.99, approved for registration as a Draft International Standard. C2PA participates with that subcommittee as a Category A liaison. The process isn’t finished, and an ASIS&T standards review voted no with comments during an earlier fast-track review in 2025.

The distinction inside that standards work matters. ISO 22144 is carrying the core C2PA Content Credentials architecture toward an international standard. CAWG’s creator Identity Assertion exists alongside that work as its own ratified specification, designed to plug creator identity into a C2PA manifest. The creator layer isn’t imaginary and it doesn’t require starting over. It just isn’t the thing the whole system is organized around.

OpenAI and Google are already C2PA steering-committee members, alongside Adobe, Microsoft, Meta, Amazon, Sony and others, so the companies building generative systems aren’t absent from this conversation. That makes the omission more interesting, not less.

Who gets to be the author?

Sarony’s camera didn’t diminish his authorship because the camera was never where authorship lived. It lived in the decisions. That’s messier with generative AI because the machine contributes more. I can ask for alternatives I wouldn’t have produced on my own. I can accept a sentence substantially as generated. I can reject twenty others. I can use one model for research, another for a first pass, rewrite the result myself, bring it back for criticism, and publish something that no single participant in that chain could accurately claim to have produced alone. That makes provenance harder and it also makes it more important.

“AI was used” is probably going to become less informative as AI disappears into ordinary software. Phones already make computational decisions every time we make a photograph. Photoshop has had algorithmic tools for years. Word checks grammar. Lightroom masks subjects. Cameras assemble exposures. The line between tool and maker has never been as clean as we sometimes pretend it was, and the vision was never about the tool anyway.

I want the record. I just don’t want the record to erase me. A useful system would let the private provenance become as rich as the technology makes possible, then let the creator decide which parts belong in public. It could retain the tools and meaningful edit history without requiring me to publish every conversation, abandoned draft, or note that got me there. When the work leaves its original container, the creator assertion needs some way to make that trip too. That last part may be the hardest problem in the whole argument, but hard isn’t the same as impossible.

Ssomewhere in that transition, before we get lost in which model produced which token, the finished work should preserve the thing that copyright law, photography, publishing, and creative practice have been trying to identify for a very long time. The person willing to put their name on the work.

He asked whether she’d read the terms of service first. She hadn’t.

I can’t remember the last time I read every word of a terms of service agreement before clicking whatever button stood between me and the software I was trying to use. Zusman pointed out that she’d handed photographs of her face to a platform she knew little about, with no idea whether they might be used to train models or end up in a facial recognition database. She deleted the account that night.

I kept coming back to the privacy part, partly because I’d never looked at my own software that way.

What the file becomes

My first concern was unfounded. Photographs themselves are not treated as biometric data.

Illinois’ Biometric Information Privacy Act, probably the best-known biometric privacy law in the country, explicitly excludes photographs from its definition of a biometric identifier. It includes scans of face geometry. It also includes voiceprints.

The photograph or recording is the source material. Software can derive something else from it.

Note taking has never been a skill I excelled at, so about a year ago I subscribed to Otter.ai. Otter is transcription software that helps me improve my client experience. With permission, as I talk to a client, Otter records the conversation so I don’t have to interrupt to take notes or ask someone to repeat themselves because I missed a word. It’s hard to overstate how much better it is to have a searchable transcript instead of whatever notes I managed to scribble while also trying to pay attention to the person in front of me.

The terms describe another process. Otter calls the result Speaker Identification Information. Its Data Processing Agreement is less delicate about the terminology: the service may process voiceprints so it can recognize users and automatically tag their names in a transcript.

That isn’t something I need it to do, partly because I was in the meeting and partly because Otter doesn’t even do a great job with that task. I can’t tell you how many meetings I’ve been in with a few people and ended up with a transcript that assigns twice the number of speakers as were actually in the meeting.

A transcription service can hear that one person stopped talking and another one started and label them Speaker 1 and Speaker 2. It doesn’t need to know either person’s identity, and it doesn’t need to remember the voice after the meeting ends.

Otter can remember enough about a person’s voice to recognize that person later. I had agreed to that. The client whose voice I recorded hadn’t agreed to Otter’s terms at all.

I had been asking the wrong question (it seems I do that a lot)

My first concern was model training, mostly because AI companies have trained the rest of us to look for that language now.

Otter does use some customer material for that purpose. Its current privacy documentation says it can train its proprietary AI using de-identified audio recordings and transcriptions. The transcripts may still contain personal information, something Otter itself acknowledges. Of course they can. Otter doesn’t control what people say during the recording.

The voiceprint appears to be a separate operation. Otter’s public documentation doesn’t say that it trains models on the voiceprint itself. It uses Speaker Identification Information to recognize speakers.

“They train on your data” turned out to be too blunt a description of what was happening. The same product makes several decisions about the same conversation: what gets recorded, what gets retained, what gets derived from the recording, whether that derived information identifies someone later, and whether any of the original material enters a model-training process.

Otter’s Enterprise controls made that easier to see. Enterprise customers can disable speaker learning. When they do, Otter can still separate speakers in a transcript, but it stops creating or updating the persistent voice information used to identify them later. Otter describes the setting as useful for organizations with biometric, privacy, or compliance requirements. The switch lives in Enterprise.

I had also been trying to understand how all of this squared with Otter advertising HIPAA compliance. The answer is less contradictory than it looked at first. HIPAA-compliant Otter isn’t the ordinary service with a privacy badge attached. It requires an Enterprise account and a signed Business Associate Agreement. Enterprise customers are opted out of global AI model training by default, and they have privacy controls that ordinary accounts don’t. That resolved my HIPAA question.

What I was left with was the feeling that if it costs more to protect my, and my clients’ biometric privacy, then it is that data that covers the rest of the cost. We’ve seen that playbook before.

Three transcription tools

I had already been looking at Fireflies.ai as a possible replacement for Otter, so I read those documents next.

Fireflies also processes characteristics of a person’s voice. Its Privacy Policy says service providers may derive voice information so the software can distinguish one participant from another, and it acknowledges that some jurisdictions may treat that information as biometric data.

Fireflies draws the line earlier than Otter does. Its policy says the voice information isn’t used to identify or authenticate an individual. Fireflies says it doesn’t receive or process that biometric representation on its own servers, and it publishes a destruction schedule for biometric information where Illinois law applies.

The service needs to know that Speaker 1 and Speaker 2 aren’t the same person. It doesn’t need Speaker 2 to still be Michael next Thursday.

The Terms deserve a little more attention than that comparison alone suggests. Section 5(c) says Fireflies won’t use User Content to train, retrain, fine-tune, or otherwise improve generative AI models, but it also reserves the right to derive usage, statistical, and other data from that content for internal business purposes, including service improvement. Its Data Processing Addendum makes a similar allowance for de-identified data. I don’t read that as nothing.

The difference for the question I was trying to answer is that Fireflies gets much more specific about the voice data itself. Its service providers use those characteristics to tell speakers apart, not to identify or authenticate them, and Fireflies says that biometric representation never reaches its own servers.

Fireflies also says meeting content isn’t used to train AI models and that the vendors processing that material operate under restrictions against using it for their own model training. That was considerably clearer than I expected from the first pass through its Terms.

I use Zoom too, which supplied a third version of the same decision. Zoom’s terms explicitly exclude the audio, video, chat, screen sharing, attachments, and similar customer communications in a meeting from training Zoom’s or third-party AI models. Zoom can create a voiceprint. The interesting part is when.

Its ordinary transcription functions can distinguish speakers without permanently knowing who they are. Features that actually need a persistent voice identity, such as personalized audio isolation or automatic smart name tags, require the person whose voice is being learned to enable or enroll in that feature.

I kept expecting the three services to converge once I read far enough into the legal language. They didn’t. They all need audio to transcribe a conversation. They don’t all make the same decisions about what else to extract from it.

The consent chain

Photographers aren’t new to handling other people’s data.

I have redundant systems for image files because losing a client’s photographs would be unacceptable. I use payment systems designed to keep credit card information out of my hands as much as possible. Galleries have access controls. Contracts and client records live in systems chosen partly because those systems are supposed to be safer places for them than a random folder on my computer.

I’d put transcription in a different mental category. It was a convenience and it was customer service. A design consultation can cover where the finished photographs will hang, what size a client is considering, details about the session, names, dates, scheduling, pets, family members, sometimes things that don’t matter to the photography at all but come up because two people are having a conversation. I started using transcription because it let me listen instead of dividing my attention between the client and my notes.

Once I add Otter to that conversation, there are at least two consent questions. One is easy to see. Does the client know I’m recording? The answer is always yes, because I ask for permission before we start. The second took me longer. Does the client know what the service receiving that recording does after it gets there? Here, I dropped the ball.

Recording-consent law varies by state. Biometric law varies much less, because most states do not have biometric privacy laws on the books today. Illinois requires notice and written consent before a private company collects covered biometric identifiers and gives individuals a private right of action when companies violate the statute. Other states regulate biometrics differently, often through their Attorneys General rather than private lawsuits. Arizona doesn’t have a general BIPA-style biometric privacy statute. Lawmakers introduced another biometric privacy bill in 2026, so that’s subject to change.

I don’t think photographers need to become privacy lawyers. But I also think “my state doesn’t prohibit this” is a remarkably low standard for deciding what I’m comfortable doing with somebody else’s information.

What changed

I started this research because I read about someone I don’t know that uploaded photographs to an AI headshot service on LinkedIn. I ended up looking much harder at software I already use.

Otter isn’t secretly doing something it never disclosed. The voiceprint language is there. So is the model-training language. Fireflies has its language. Zoom has its language. I hadn’t read enough of it.

I use transcription because I want to pay better attention to clients. That part hasn’t changed. What changed is whether I’m willing to hand their conversations to a service that can create a persistent representation of their voices when another service can do the job without making the same choice.

My Otter subscription will expire next week and it will not be renewed.

That was supposed to answer the question I left open in the last issue. The July 17 changes had gone live, the second run had arrived, and I expected the score to tell me whether the web had learned anything.

It did not.

The problem was not that the score stayed flat. The problem was that the question had never been measuring my market closely enough for the score to mean what I thought it meant. I still do not know whether the July 17 fixes changed anything. The instrument I planned to use for that answer was surveying painters.

Another monitor, running during the same period, reported fine art pet photography as one of my strongest positions. It queried the same answer engines and evaluated the same business during the same week, yet one score said I did not exist while the other placed me near the top of everything it tracked.

The questions were different. The commercial monitor included the word “photographer” and asked about Tucson. Mine asked who creates fine art pet portraits in Arizona. The engines answered my question correctly. The artist taking most of the top positions is a real fine art pet portrait artist working in Arizona.

She paints.

The geography was doing more work than I had noticed too. Ask someone from Phoenix about Arizona and there is a decent chance the answer will concern Phoenix, Scottsdale, Tempe, Mesa, Chandler, or someplace else in the Valley. Tucson may eventually enter the conversation, along with Flagstaff and whatever part of Sedona they visited last winter, but Phoenix has enough people to mistake its part of the state for the state itself. It seems the models made the same assumption.

I had spent two months reading that output as a verdict on my photography in Pima County.

The plumbing was fine

I built the tool, so I knew where to look when a result seemed wrong. I checked whether the search fired, whether sources came back, whether the provider stayed grounded, and whether the output parsed correctly. I also checked for the failures I have learned to expect: malformed JSON, missing citations, silently degraded calls, and information that sounded convincing until I tried to verify it.

None of those things had happened. The engines received a clear question and returned an accurate answer, which meant the tool had done exactly what I asked it to do. The problem sat one layer earlier. I had audited the plumbing and never examined the premise.

A tool cannot flag a well-formed question about the wrong thing. There is no error state for that. The output looks exactly like a valid measurement because it is one, and the mistake occurs before the search begins. That makes it harder to catch than the failures that announce themselves with missing fields or broken responses.

I apply more skepticism to other people’s numbers. When a vendor publishes a statistic about AI search, I treat it as a hypothesis until I understand what was measured, how the sample was built, and what the company has to gain by presenting the result that way. My own tool’s output entered my files as a finding. I did not trust the underlying engines more because I had built the interface around them. I trusted my own question without noticing that I was doing it.

The second tool is not what helped

The obvious response is to add another monitor, but that is not what caught the mistake. Had I copied my prompt from one system into the other, both would have surveyed painters. I would have received the same zero twice, from two different tools, and treated the agreement as confirmation.

Two tools agreeing on a bad premise is one measurement billed twice.

The disagreement existed because the questions had been composed separately. I instructed Claude to write the prompt inside my tool while thinking about category banks, geographic variables, and keeping weekly runs comparable. I wrote the commercial monitor’s prompt by hand in a different sitting, thinking about how someone looking for my work might phrase the request. Same person, two working postures, two questions that were not actually about the same market.

That is useful, but it does not solve the problem by itself. Another person can still share my assumptions. An AI system drafting blind will probably rebuild a lot of the same category language from the same business context. Even I, a week later, am still bringing the same model of my own market to the page.

I already had a source farther upstream.

Earlier this year I did two rounds of prospect research, one for pet photography and one for headshots. I asked six AI systems to map the questions people ask while they are still deciding whether professional photography makes sense for them. I was looking for the language on the other side of the transaction, before someone has read my site long enough to adopt my vocabulary.

Those questions do not sound much like my category labels.

Pet owners ask whether a photographer can work with an anxious dog, whether the whole thing is worth the money when they already have thousands of phone photos, or whether it is too late to make good photographs of an old dog. Headshot clients ask what happens if they are not photogenic, whether they will look awkward, and how to find someone who is good with people who hate being photographed.

The important part is not that AI generated those questions. It is that I generated them months ago for a different purpose, before this measurement problem existed and without any reason to shape the wording around what I would later want a visibility tool to prove. Working with these tools are like that; it’s not always a straight line but the work, when done right, always builds on itself.

That research was sitting in my files while my monitoring tool asked about “fine art pet portraits in Arizona.”

It is still mediated evidence, and my inquiry history gives me another source that began outside my own category system entirely. But both are more useful here than asking myself to invent one supposedly neutral prompt after I already know what result I am trying to test.

Read more

I had no plan for it. I opened Gemini, typed give me something fun, which is less a prompt than an abdication, and handed over the photo.

What came back was a character sheet. Birdy as the protagonist of a fantasy role-playing game, stats and all. Charisma boosted by the forehead wrinkles. Perception maxed out, on the reasoning that she can hear a cheese wrapper open from the next room. Stealth close to zero, because she is too striking to blend into anything.

We passed the phone back and forth from there. One of us would read the output, laugh, and prompt the next turn. The sheet asked for Birdy’s name and her real-life superpower, so we told it: Birdy, and mind control, since there is no version of events in which the hand does not eventually produce the treat. The model took that and kept building. Birdy became a sorcerer whose subclass is Absolute Ruler of the Living Room. It asked for her highest-value target. We said cheese. It returned a doctrine, the Cheese Tax, under which the sound of a wrapper travels on a frequency that bypasses ordinary physics.

I installed a phone update that night, and that is how I learned the phone now makes music. Lyria 3, Google’s music model, builds a track from a text prompt. I told Robin I could not wait to send her my first song. She was skeptical there was much use in a phone that writes music. I already knew what the song would be.

I left the update alone until late the next morning before I made anything with it. The play is real. It does not get to cut in front of the work.

The result is called The Cheese Tithe. It runs two minutes and thirty-eight seconds and came with its own cover image. It is a rock anthem about a dog holding a household hostage for dairy.

The lyrics describe the actual room.

One blue, one brown, they’re locking in the static,
Birdy’s on the blanket, keeping it emphatic.
The wire crate is open, the brick wall is a blur,
There is no escaping the gravity of her.

I never told it what the room looked like.

The chorus is the part that gets stuck in your head, which is the job of a chorus:

You feel the voltage, the master command!
There is no stopping the reaching of the hand!
The tithe must be paid, the willpower is gone!

I sent it to Robin. She laughed. She had doubted there was any use in a phone that writes songs, and the use turned out to be this: two and a half minutes of arena rock about our dog’s dairy racket, useful in no sense a spreadsheet would recognize.

None of it was for anything. That is the whole point, and it took me years to trust it.

In graduate school I started making photographs that had no assignment behind them. Macro-close work on train cars parked near a mall, paint and scratches and weather shot so tightly the train stopped reading as a train and became color and line. Ink dropped into water and caught as it fell. Glass lit until it looked like deep space (okay, that one came later). I didn’t think of it as a body of work. I thought of it as play, and I hadn’t yet learned that play could justify itself.

I understand it now. That work became a series I still print and sell, seventeen years on, and one of the glass photographs spent a season on a gallery wall in Los Angeles. I only gave the series a name recently, while rebuilding my website, when I noticed I had been filing the images by how I made them instead of by what they were.

The Birdy session and those grad-school photographs are the same move, seventeen years between them. Each one starts with something in front of me that has no particular use, a gratified train car or a dog waiting on cheese, and each one follows the play without settling the destination first. Sometimes it arrives at a pigment print on cotton rag. Sometimes at a rock anthem about a tithe. The destination does not need to justify the beginning. It doesn’t need to justify the final product either.

That does not make the song an outreach strategy. It was not made for that. But a song about the Cheese Tithe travels in a way a workflow essay never will. Someone forwards it to a person who had decided the whole subject was either frightening or beneath them, and for two and a half minutes it is neither. It is just funny.

Birdy has no idea any of this happened. She is on the blanket by my feet while I write this, snoring gently away.

The cheese will wait.

P.S. I finished this, and then Robin sent me hers.

The work begins before the draft. I start with a problem I have encountered in my own business, form a hypothesis, run tests, collect evidence, and work through what the results mean. That often includes long written exchanges, dictated notes, original tools I have built, outside research, and repeated attempts to clarify the argument before I ask for a full draft. The models work from a detailed writing guide, established business facts, documented terminology, audience context, and lessons carried forward from previous articles.

Once a draft exists, I edit it, question it, add examples, remove language that weakens the argument, check externally verifiable claims, and decide what remains. I add the photographs, captions, artwork, structure, and final context. I also review what the process taught me and update the writing system so the next article reflects those decisions more accurately.

AI usually writes most of the first complete draft because it can turn a large body of directed material into prose faster than I can type it myself. It does not decide what I believe, what the evidence supports, what the article argues, or what I publish.

Pangram may describe the surface of that workflow as “AI 100%, Human 0%.” I do not. That score can identify the tool that produced much of the sentence-level wording. It cannot measure the research, judgment, experience, direction, revision, or responsibility behind the finished work. Treating those two things as equivalent mistakes a production method for an account of authorship.

I do not think my own score proves how well or badly Pangram works across Substack. It does expose a question worth testing. I am considering a small study built around writers with documented workflows, comparing fully hand-written work, corrective AI use, directed AI drafting, and unattended generation. The useful question is not simply whether Pangram detects machine-produced wording. It is whether the label it gives readers accurately describes how the work was made.

If you publish on Substack and can document your process candidly, I would be interested in hearing from you.

The tools are part of how I make this. They are not the reason it exists.

Ask five AI engines whether anyone in Tucson brings a mobile photography studio to a client’s home or office, and my business appears in five out of five answers, ranked first in four.

Ask those same engines who creates fine art pet portraits in Arizona, and I appear in zero. Several recommend painters instead.

Same brand. Same credentials. Same afternoon.

In the Note, I called the hard half “discovery.” That word covers less of the problem than I thought. Discovery is whether the engines can resolve you: your name goes in, correct facts come out.

Distinguishability is whether the engines choose you: a buyer’s question goes in, your name comes out.

They are different problems. The second does not follow automatically from the first, no matter how much of the first you build. I started calling this condition in June: known but not recommended.

That still describes part of the problem, but not all of it. The more disappointing failure happens after the engine has found you, verified you, and assembled the facts correctly. It still has to decide what those facts mean.

Read more

What came back looked right. Business name, correct. Categories, correct. Price range, correct. Address: 6440 N Campbell Ave, Suite 210, Tucson, AZ 85718.

That’s not my address. It’s Vivace, a Northern Italian restaurant in the Catalina Foothills, established in 1993, and to the best of my knowledge has no relationship whatsoever to my photography business.

I caught it in about four seconds because I know where my mail goes, and this wasn’t it. Gemini apologized and corrected it, the way it apologizes for anything once you catch it, which isn’t the same as being trustworthy. I’d handed it a task that sat just past what I could verify myself on sight: structured data, schema fields, the layer I understand in principle but can’t proofread the way I proofread a sentence. That’s the bot-sitting territory from the last issue. The tools go furthest wrong precisely where you’re least equipped to catch them.

I only caught this one because the address happened to fall inside what I know cold.

I’ve been running this business through AI tools for more than a year now, and I still don’t trust Gemini with anything I can’t verify myself. I don’t know that changes.

Nothing published, no real damage. But I kept thinking about the version that doesn’t get caught in four seconds. I recognized my own address wasn’t the one in front of me. What happens with a fact I’d recognize a little less certainly? A credential date that’s off by a year. A service category that used to be right and quietly stopped being right somewhere along the way?

I built two documents for future work. Then I asked an AI tool to hold both of them against pages I’d already published.

What the Card Actually Is

The first document is what I’ve started calling my entity card. It gives an AI tool one place to check who I am instead of reconstructing the answer every time.

Legal name, founding date, the two cities I moved through before Tucson, the credentials with the years they were actually earned, the service categories in the language I’ve decided wins over the language that doesn’t. None of the information is secret. Most of it is already sitting in my ORCID record for anyone who wants to look.

None of it is new, either. Putting it in one dated document gives me something concrete to check everything else against, rather than asking an AI tool to reconstruct me from whatever it half-remembers every time I need something written.

A model without a document like this doesn’t lie to you, exactly. It encounters a gap and supplies something plausible enough to fit, and a business address in a business’s own city is plausible right up until it turns out to belong to someone else’s restaurant.

Give it a dated source of truth, and at least the address no longer depends on a guess.

Who It’s For, Not What I Sell

The second document took longer to admit I needed because it’s less obviously about me. It’s a set of profiles, one for each real kind of client or organization I work with, built in detail and in writing before I ask an AI tool to help with anything that touches marketing.

The shape is worth describing because the shape is the useful part.

One profile starts with the obvious markers: age range, income bracket, whether they rent or own. None of that does the actual work. What matters is three lines down: what they’re worried about before they ever email me, and which detail in my first response would tell them I already understand that worry without my having to say so out loud.

Demographics tell an AI tool who might buy something. Decision drivers tell it what might move that specific person from considering the service to contacting me. Most client profiles stop after the first part.

The same principle applies across the rest of my practice. A corporate team needs different language than an individual replacing a ten-year-old LinkedIn photo. A private pet client needs something different from a rescue trying to get an overlooked dog adopted. The subject might still be a dog, but the audience, purpose, emotional stakes, and useful next step are not the same.

Whatever I ask an AI tool to write, from a corporate service page to a caption for a shelter dog’s Instagram post, it’s supposed to know which situation it’s writing for and who needs to read it. Most of the time I don’t have to rebuild that context from scratch. It’s already in the reference material the tool is working from.

The Part I’d Already Solved

There’s a third piece of this that I wrote about a few issues back, the writing guide that tells whatever tool I’m using how I actually sound, not how a generic photographer sounds.

I won’t re-walk that ground here. By the time I built the other two documents, I already had proof this approach worked because the writing guide had been quietly doing its job for months.

The newer documents answer two different questions: is this aimed at the right person, and is it even true?

Checking the Work That Already Existed

I assumed the main benefit would be retrieval. Having the facts and the audience written down meant I could stop re-explaining myself every session. That part’s real, and it’s most of what I wrote about a few issues back.

Then I stopped treating the documents only as background material living in a knowledge library. I asked an AI tool to check my existing work against them.

I picked three older pages on my site, ones I hadn’t touched in a while, and asked whether they still matched what the entity card and the writing guide actually say. They didn’t, in two different ways.

One page, built for corporate clients, still sells something called an Executive Session. That’s the word I ruled out months ago, in writing, after my search data showed that people find me for “corporate team headshots,” not “executive headshots.” I didn’t deliberately put it back. The page was built before I’d made that decision, and I hadn’t gone back to check it against a decision that came later.

The tool caught it in one pass because for the first time it had something concrete to check against instead of an instinct that something read a little off.

The other page, a fine art series from my years in Kentucky, gives two different date ranges for when the work was made, one in the metadata and a different one in the paragraph underneath it. Read past that, and it describes the choice behind the work as something I made “before relocating west in 2011,” which skips the fact that I’d already left Kentucky for Washington three years earlier.

Claude helped me build that page. It was created before I had any of these persistent identity documents in place, which meant the tool was working from whatever context I supplied during that session and whatever it could infer around the edges.

I didn’t set out to invent anything. I’m sure Claude didn’t make that choice either. A compressed sentence had turned into an error in my biography, sitting on my own website, waiting for whatever eventually reads it and decides it’s true.

Both pages are getting rewritten this week. That’s not really the point.

Neither error would have surfaced from a normal proofread because a normal proofread checks whether the sentence works. What caught these was checking the sentence against something written down and dated, the same discipline that caught the restaurant’s address before it ever went live.

I got lucky with Gemini because I happened to know my own address cold. I don’t know every fact about my own business that well, and neither does any AI tool I hand a task to unless I’ve already told it what to check against.

The address was obvious. The next error might not be.

Over the past year I’ve gotten faster at my work than I’ve ever been. I’ve also been more tired at the end of the day than the work alone explains. For a while I filed that under getting older, running a one person shop, the general weather of the year. I started paying attention in March, when Julie Bedard of Boston Consulting Group went on the podcast Hard Fork to talk about a study her team had just put in Harvard Business Review. They surveyed about fifteen hundred full-time workers, gave the exhaustion a name, AI brain fry, and tied it to a number: self-reported productivity climbed as people added AI tools, up to around three, then fell off a cliff past four. I run five pretty regularly. By their count, I’m the cautionary tale, and I’ll come back to that.

A separate institute survey has been sitting with me more recently. This one comes from Glean’s Work AI Index, six thousand workers rather than fifteen hundred, and it isn’t asking how many tools fry you. It’s asking where the time goes. Eighty seven percent using AI. Seventy three percent saying it makes them more productive. On average they report saving about thirteen hours a week, close to a third of a workweek. And only thirteen percent say their organization is actually performing significantly better because of any of it. The gains are real. The gap is real. The report exists to ask where the thirteen hours went, since they clearly didn’t all land where they were supposed to.

One answer is bot sitting: the tracked and untracked labor of making the tool usable. Feeding it context. Watching its output. Debugging the parts that come back broken. Cleaning up after it. The report puts that work at roughly six and a half hours a week, about half of everything the AI just saved you. Not all of that is wasted time, some of it is iterative and healthy, but a lot of it is tedious, invisible, and tiring.

The single most draining piece is feeding the tool context, the part it feels like it should already know. Which documents matter. Where they live. Which ones are authoritative. Second is debugging, because the thing is probabilistic rather than predictable. You fix something and you can’t quite tell what you fixed or whether it will stay fixed. It adds up to a quiet tax that shows up as fatigue long before it shows up as a performance problem.

Then comes the finding underneath the finding. Forty percent of workers admit to shipping AI work they couldn’t explain if someone asked them to. The report gives that its own cruder name. The path from bot sitting to that failure mode is short, and it runs straight through exhaustion. You sit the bot until you can’t anymore, and somewhere in there “good enough” quietly becomes permission to ship.

Researchers have a polite term for settling at good enough: satisficing. Creatives have heard the slogans that make that feel reasonable. Done is better than perfect. If you’re not embarrassed by the first version of your product, you’ve launched too late. Document, don’t create. All of them were originally answers to a real problem, perfectionism that kept people from shipping anything at all. They work when the rough thing is still yours, when you can explain what you tried, what you learned, and what you’ll change. They stop working the moment the rough thing is something the tool wrote and you can’t defend it. That’s the edge where iteration turns into surrender and where “good enough” stops being a stepping stone and starts being a ceiling.

What the report describes is satisficing scaled into a workplace norm, work going out the door that the person who sent it could not defend. That’s the surrender of agency written as a survey statistic. It’s the generic photographer seen from the inside: competent, smooth, and indistinguishable from anyone else. It doesn’t get fixed by a better prompt or a faster model, because the variable was never the tool. The variable is whether you’re still in the loop or just clearing the machine’s throat for it until you stop checking.

Rebecca Hinds, who runs the Work AI Institute, said the line that I’ve been pondering since. Most people are too exhausted to be curious. I believe her. I also know that’s not the only way this goes, because it isn’t the way it went for me, and the difference between the two is where this issue earns its keep.

Read more

I checked it against something. The same grader gave a friend’s business eight percent too. And another. The number wasn’t a measurement. It was a default.

This is the story of what I found when I went looking for where that eight percent came from, and what it took to build a tool that would stop handing me numbers like it. I rebuilt the thing twice. The first rebuild lied to me in exactly the way the original had, which turned out to be the most useful thing it could have done. The second one stopped lying, and the price of that was watching most of my scores fall.

The grader that started the journey

The tool I started with came out of HubSpot, demonstrated on a podcast, the kind of free instrument a company builds to show you a problem it would like to sell you the solution to. You gave it a brand and it gave you back a composite score out of a hundred, built from five dimensions: brand recognition, market position, presence, sentiment, and share of voice. It ran across three AI engines and averaged them.

It looked like measurement. It had the furniture of measurement, the weighted dimensions and the provider columns and the tidy final number. What it didn’t have, in the one dimension I went looking inside, was data. The share-of-voice figure was an estimate the language model was asked to produce, a guess at what percentage of the market each competitor held, and when the model couldn’t produce one for the subject brand, the code filled the hole with a literal fallback value. Eight percent. Written into the source as a default for when the real answer was absent.

That is worth sitting with, because it is the whole problem in one line of code. When the tool had nothing, it did not say it had nothing. It said eight percent.

Read more

No spread. The engines agreed.

Where Michael Kloth Photography pulled fifty points of disagreement out of five systems, the newsletter pulled near-consensus. Perplexity, OpenAI, Gemini, and Grok landed within a few points of each other, low. Claude went lowest, 5 to 8, noting that a search for the exact title “returned zero direct results.” One engine read the Substack profile closely enough to report the subscriber count: seven, the day I ran it. The number was correct, and two months in, it wasn’t the part that stung.

Read more

I’ve been running AI through my photography business since June of last year, which is long enough to have formed opinions about where the convenience stories hold up and where they leave something out. This is one that leaves something out.

The old memory was passive. It wrote down what you explicitly told it to remember, and not much else. The new system is active. In OpenAI’s own description, it synthesizes context from your chat history on its own, including information that comes up naturally in conversation rather than through a “remember this” instruction. It also acts on what they call implicit preferences, the things you never stated but that it inferred from how you work.

Read the announcement and the example they lead with is a vacation. You plan a trip, ChatGPT remembers your travel style, future trips get easier. That’s the version of the worry most people land on, too: is my vacation going to end up somewhere it shouldn’t.

It’s the wrong example. Not wrong exactly, just not the business example.

Here is the one that should concern anyone with clients. You write a proposal for Client X in ChatGPT. Next week you write one for Client Y, in the same account, because that’s simply how a person uses one login. The system isn’t storing a note that says “this pricing is confidential to X.” It’s building a working model of your projects and surfacing what’s relevant. And the moment relevance peaks, the moment the model is most ready to volunteer what it knows about your proposal work, is the moment you sit down to write the next proposal.

What crosses over necessarily isn’t a name you’d catch on a read. It’s a number. A scope assumption. A piece of positioning you developed for X, arriving fluent and on-topic in the draft for Y.

This is a different failure than the one I usually write about. Fabrication gets caught eventually, because some part of you stays alert to the invented detail that was never real. I’ve called that plausible fiction. This isn’t that. The detail here is real. It’s just real about the wrong client. Your accuracy radar never fires, because nothing is inaccurate. The content is correct. The context is the breach.

That’s what makes it slip. Bleed that reads as competent is the kind that clears review.

And the stakes clear a real bar. If X and Y compete, and in a narrow market they often do, two firms in the same city, two organizations chasing the same grant, this stops being a privacy curiosity and becomes a confidentiality exposure with NDA-shaped edges. The embarrassing version is your dinner preferences showing up in a work email. The serious version is one client’s terms shaping another client’s deal.

The defense isn’t clever, and that’s the point. Keep client work in separate, governed spaces. Use Temporary Chat, which doesn’t read or write memory, for anything you don’t want synthesized into the general model of you. Read the memory summary page now and then, so you actually know what the tool thinks it knows. None of that is hard. It’s just deliberate, and deliberate is exactly what the convenience pitch invites you to stop being.

That’s the whole tension in a line. The casual user hears Dreaming and thinks, good, now I don’t have to bother. Run a business on these tools and the conclusion runs the other way. The bother was never overhead you could finally shed. The bother is the discipline that keeps one client’s work out of another client’s draft. It’s the moat.

The setting worth reviewing isn’t in the menu. It’s the assumption that convenience comes free.

I got a spread from 28 to 78.

A control isn’t supposed to do that. And the reason it does is the whole point of this issue, because once you understand why a brand with this much real signal scatters across fifty points, you stop reading the headline number as “your AI visibility” and start reading the thing underneath it, which is where the actual information lives.

One note before the numbers. These runs come from a version of the grader I’d already adjusted since last week’s issue. I was in its guts before I’d decided to write about what’s in there. More on why next week.

Read more

The tool lets clients upload a photo of their wall, select a print from my gallery, choose a mat color and frame style, and see roughly what the finished piece would look like in their actual space. The visualization part worked from early on. The “Send to Michael” button was a different story. Clients could build something and then it disappeared. I debugged it, made partial progress, and then set it aside when other things needed attention.

I came back to it recently using the Fable 5 preview of Claude Mythos, which Anthropic has made available to all account holders through June 22. The difference was immediate. Mythos found the problem, explained exactly what was wrong, and fixed it in one session. Two months of intermittent frustration, one session. I’m not going to oversell it, but the model matters, and that’s worth saying plainly.

The send problem turned out to be two issues compounded. The endpoint wasn’t registered correctly, and a pre-existing bug was re-triggering the export panel and hiding the contact form every time a step completed. Both fixed.

Once the email was actually arriving, I noticed it was reporting pixel dimensions of the art file rather than the physical print size on the wall. The pixel count of the source image tells me nothing useful during a client conversation. The print size in inches does. Fixed that. Then I reworked the resize behavior: the original tool locked art to its native aspect ratio, which isn’t how print ordering works. The corner handle now moves freely, and when the dimensions come within about half an inch of a standard print size (8x10, 11x14, 16x20, 20x30, and others, either orientation), it snaps to that size. Anywhere else it stays custom and labels itself as such. The email I receive now includes each print’s image area, whether it landed on a standard size or something custom, and the full framed dimensions.

The tool is live at visualizer.michaelklothphotography.com/visualizer/ if you want to see it. All that’s left is to finish populating the gallery with my artwork. Pet clients will get a customized version that shows portraits from their session.

I have ideas about where it could go from here. As it stands, it does what I built it to do: gives clients a concrete way to think about print size and wall placement before we get to the viewing appointment. That’s enough.

Thread closed.

On June 5th, the argument landed in Google’s formal developer documentation. Not one page. Three, coordinated. That’s a different kind of signal.

The first is the “Do You Need an SEO?” guidance page, which lists the services worth paying a professional for. The updated list now ends with this: “Optimizing for generative AI.” Lower on the same page, in a new section on evaluating an SEO’s tools, Google names the discipline directly: “advice on optimizing for AI experiences (also known as ‘AEO’ ‘GEO’ services).”

Not a conference talk. Not a blog post from a marketing agency. Google’s official hiring guidance.

The second is a new page dedicated to evaluating third-party SEO tools and advice. It covers the range of services claiming to help with AI search visibility, and on the tooling specifically, it says: “Third-party tools don’t have access to our internal ranking data. They can’t guarantee performance. Any predictions are their own and like predictions generally, may not happen.”

The third is the one that completes the move: a dedicated official guide on optimizing for generative AI features, which both of the other pages now link to. So the picture as of June 5th is three coordinated parts. The service is worth paying for. Here is how to evaluate the tools claiming to do it. Here is our own guidance on how to do it.

Read together, the updates make the same argument Sullivan made in Toronto, but now it’s in the documentation: AEO is real enough to pay for, and the tooling claiming to measure it is unreliable enough to scrutinize carefully.

I’ve been sitting with that second point for a while.

After the May issue I started running my own brands through the HubSpot AEO Grader, which is the tool Kipp Bodnar used for his live demo in the Marketing Against the Grain episode I referenced. The grader produces scores across multiple dimensions: brand recognition, share of voice, presence quality, sentiment. It generates competitor comparisons.

It has problems. I found them. And then I built something else. It’s not pretty, but it is effective.

The next two issues cover what I found and what I built. They’re specific, they’re mine, and they’re behind the paywall, which is where the methodology work in this newsletter lives.

The arc over the next three weeks:

Issue 2 (June 17, paid): What the tool found on a mature photography brand.

I ran seven subspecialty queries against Michael Kloth Photography, the business I’ve operated since 2006, with 20 years of credentials, two published books, and a Wikidata entity. The signal web behind it is real: PPA, ASMP, the TMC Foundation collection, Getty Images.

The scores ranged from 28 to 78 depending on the AI provider and the subspecialty. That spread isn’t a calibration error. It’s structural, and understanding why Grok and Perplexity score the same brand 50 points apart on the same query is the first thing you have to understand before any of this is actionable.

Issue 3 (June 24, paid): What the tool found on a three-month-old newsletter brand, and why I stopped using the HubSpot grader.

I also ran the tool on artofphotographyai.com, this newsletter, three months old, scoring in the 40s across the board. A newer brand has a cleaner failure profile. The gaps are diagnosable and the remediation path is explicit.

That issue also covers the HubSpot grader’s architecture problem, what I found when I looked inside it, and why a client acquisition funnel dressed as a diagnostic tool produces a specific kind of unreliable output that Google’s June 5th documentation is, whether intentionally or not, describing exactly.

The May issue ended with this: the strategy question isn’t whether to build non-commodity content. It’s recognizing which decisions have been optimizing for the wrong thing.

The June documentation adds a layer: the measurement question isn’t whether to evaluate your AI visibility. It’s recognizing which tools are built to answer your question versus which ones are built to answer a different question entirely, and produce your score as a byproduct.

Both problems have the same shape. That’s not a coincidence.

Issues 2 and 3 are for paid subscribers. If you’ve been reading this newsletter on the free tier and the direction this is going sounds useful, this is a reasonable moment to consider upgrading.

Read more

In early May, I ran four queries through Gemini the way a potential client might. Not my name. The kind of questions someone types into an AI assistant before they know whose name to type.

“I’m looking for a pet photographer in Tucson that is experienced with shy pets.” “I’m looking for a headshot photographer to come photograph my team in my Tucson office.” Two others in the same range.

All four came back with accurate information about my business.

The shy pet query was the strongest. Gemini named Pima Animal Care Center and the Humane Society of Southern Arizona, got the volunteer history right, described the mobile studio as something that eliminates the stress of unfamiliar environments for anxious animals. The team headshot query came back with the correct pricing, the ten-person minimum, the full-day format. Someone who had read my site carefully could have written both responses. In a real sense, someone did. The synthesis came from content that exists, documented specifically enough to be cited accurately.

This is what a site doing the content work looks like from the outside. Accurate synthesis doesn’t come entirely from your own site, though. Directory listings, published features, institutional affiliations, and other third-party sources carry significant weight in what gets cited. The queries don’t always tell you which lever is being pulled. Both matter.

I didn’t run these queries to audit my site precisely. I ran them because I was looking for what chatbots know about my site. But you already know this.

I also look for the actual questions my ideal clients ask when they’re talking to an AI assistant, phrased the way they phrase them in conversation, not the way a keyword tool formats them. Those are different things, and the gap between them is getting wider. That second kind of query will be what I talk about next week.

A search bar gets keywords. “Pet photographer Tucson anxious dog.” An AI assistant gets a sentence. “I’m looking for a pet photographer in Tucson who has experience with shy and anxious dogs, ideally someone who comes to your home.” The intent is identical. The language is completely different. And increasingly, the AI assistant is where the research goes when a surface answer isn’t enough, or where it starts altogether. That’s how I use these tools as often as not: a Google search that surfaces an AI Overview, then a deeper follow-up in Perplexity or Claude rather than clicking through to individual results. I know others who work the same way. The search bar and the chatbot aren’t separate behaviors anymore. They’re sequential steps in the same research session. They’ve also drastically lowered click through rates.

The People Also Ask research habit some of you have built is still useful. Run the questions people are asking, build content around the answers, show up when someone searches. The limitation is that PAA captures search bar language, and search bar language is only part of the picture now. Clients talking to Gemini or ChatGPT before they call you are using sentences. Run the queries yourself, in client language, and you get two things back: the actual vocabulary your clients use when they describe their problems, and a live measure of whether your content is producing accurate synthesis.

The same ten minutes of work. Two different kinds of useful information.

The queries worth running fall into a few types.

Discovery queries are what someone runs before they know who to call: “pet photographer in Tucson experienced with anxious dogs,” “headshot photographer who comes to offices in Tucson.” Intent-driven queries are specific service requests, phrased as a client would phrase them to an assistant: “I need someone to photograph my team at our location in Tucson, headshots, facility, some action shots.” Specialty match queries surface your niche credentials: “photographer in Tucson experienced with photographing veterinary practices.”

Run them across multiple tools. Gemini, ChatGPT, Perplexity at minimum. What one synthesizes about you and what another synthesizes may not match, and the variance tells you something about which sources those tools are drawing from and where your content is actually landing.

Screenshot what comes back, with dates. This changes. What Gemini said about me in early May is not what it said six months ago, and won’t be what it says in six months. The only way to track drift is to document it. I run this at least monthly, usually in a gap between other things. About ten minutes once you have the query set built.

What to do with what you find, how to use the client-language data to inform content, how to influence what gets synthesized, is a longer conversation. That’s a future issue.

What I can say now is that the queries your clients are running through AI tools are the most direct signal available for what they actually need answered before they call you. The PAA equivalent for conversational search isn’t a formal tool yet. The queries are free. The methodology takes ten minutes a month. The information is real.

This issue complicates it slightly.

Brian Halligan, the co-founder of HubSpot, posted something on X on May 4th that I’ve been thinking about since I read it. He’s been one of the louder voices arguing that companies need to make themselves “legible” to AI — meaning structured, documented, queryable, so agents can actually navigate them. He’s been right about this. I’ve been making a version of the same argument at the scale of a solo photography practice.

But the May 4th post introduced a counterweight he’s calling strategic illegibility. The idea: as you race to make your operation readable to AI tools, you may be translating proprietary knowledge into a format those tools can learn from, pattern-match against, and eventually surface to everyone else using the same platform. Your edge becomes their feature.

He offered three tests for deciding what to keep out of the legible layer. The vendor test: if a vendor sold an agent tomorrow that did this specific thing, would my competitive position weaken? The departure test: if the person carrying this knowledge left, would the company lose a capability that takes months to rebuild? The replication test: could a smart competitor read the legible version and rebuild the underlying capability?

These are useful questions. They’re also aimed at companies with teams, trade secrets, and enterprise AI contracts. Applying them to a solo photography practice requires some translation.

Run the vendor test against a voice guide. If every photographer in your market had access to your voice guide, would your competitive position weaken?

Probably not much. A voice guide captures how you write, what you won’t say, how your credentials translate into client language. It’s specific to you, which means it doesn’t become a generic tool in someone else’s hands. The generic photographer problem runs in the other direction: a photographer without voice documentation produces output that could belong to anyone. Your voice guide is the solution to that problem, not a new version of it.

Now run the vendor test against your pricing intuition. Against the specific read you’ve developed for when a client is genuinely interested versus being polite. Against the sequence you use in a viewing appointment when you sense hesitation about the larger collections. Against the thing you notice in the first five minutes of a session that tells you how the next hour is going to go.

Different answer.

There’s a body of work in philosophy and cognitive science on why certain forms of expertise resist documentation. The foundational version is Michael Polanyi’s observation, made in 1966, that we know more than we can tell. He wasn’t making a mystical argument. He was making a practical one: a significant portion of what makes an expert an expert is knowledge that lives in pattern recognition, in physical disposition, in the gap between what a rule says and what a specific situation requires. When you try to articulate that knowledge fully, something changes. The articulation isn’t the thing itself.

This argument has been updated for the AI context. A scoping review published in 2025 in AI & Society, covering practitioners across creative industries, described it this way: embodied practices resist technological substitution because they encompass contextual judgments, sensory discriminations, and improvisational responses. The Photographic History Research Centre has an entire 2026 conference organized around this question for photographers specifically, framing it as the unspoken, embodied, and often invisible forms of expertise that shape photographic practice.

There’s also a counter-argument worth taking seriously. Subbarao Kambhampati, an AI researcher at Arizona State, has argued that the history of AI progress is largely a history of conquering domains that were once considered tacit and therefore AI-proof. Face recognition. Motor control. Pattern detection in medical imaging. Each of these was supposed to be beyond systematization. Each yielded.

So the honest version of the argument is this: right now, in 2026, the judgment that comes from twenty years of sessions is yours in a way that documentation can’t fully capture and AI can’t replicate. How long that remains true is an open question. Methodology built for today, with eyes open about tomorrow, is the most defensible posture available.

That’s not a guarantee. It’s a hypothesis with good supporting evidence and an important dissenting voice. Science operates that way too: a hypothesis becomes a working theory through accumulated evidence, and the accumulated evidence usually surfaces complications along the way. This is where we are.

So what belongs in the documented layer, and what belongs in the other one?

The documented layer is the infrastructure for making AI useful: your voice, your prohibitions, your credentials and what they mean in plain language, your process as it exists on paper, the context that lets a tool operate in your register rather than the generic one. This is what the last three issues covered. It’s worth building and worth maintaining.

The other layer is harder to describe precisely, which is part of the point. It includes the judgment calls that don’t have rules: which images to show first in a viewing appointment and which to hold back, what you notice about a client in the first exchange of emails that changes how you prepare, when to stop pushing for the frame you planned and make a different one, how you read an animal that’s been in a kennel for three weeks and hasn’t had meaningful human contact since intake. These aren’t secrets you’re protecting from competitors. They’re not things you’ve decided to keep proprietary. They’re knowledge that exists in the doing, and writing them down in any form usable to an AI tool would require simplifying them past the point where the simplification is accurate.

The test Halligan offers is whether a competitor could read the legible version and rebuild the capability. I’d reframe it slightly for this context: could you write it down without making it smaller? If the answer is no, it belongs in the other layer. Not because documentation is dangerous, but because the documented version wouldn’t be the thing you’re trying to protect.

This distinction between the two layers isn’t unique to photography, or to AI. Any craft practice has knowledge that lives in the practitioner and knowledge that can be transmitted through instruction. What’s changed is that the documented layer now has a powerful collaborator who can work from it across sessions, maintain consistency at scale, and handle the output that doesn’t require judgment. That’s genuinely useful.

What hasn’t changed is that the undocumented layer is still where the work comes from.

The voice guide doesn’t make the photograph. It helps the writing that surrounds the photograph sound like the person who made it. That’s the right job for it. Knowing what job you’re asking it to do, and what jobs it can’t do, is what keeps the whole system calibrated.

What the shelter taught me, though, is that being patient and appearing patient to an animal are not the same thing.

You can be waiting in a way that transmits urgency. The posture is right, the pace is right, but something underneath is impatient, and animals receive that signal with a directness that humans are too polite to acknowledge. The dog doesn’t ignore it. The dog goes further into the corner.

I’ve been volunteering my photography services to animal welfare organizations since 2005. That’s not a credential the way MFA is a credential or CPP is a credential. It’s closer to what happens to your eyes when you spend enough mornings photographing animals who have no stake in cooperating with you. They can’t be coached before the session. They can’t be reasoned with. “We’re just going to do a few quick shots for the website” is not a phrase that lands with a dog who arrived three days ago and has spent most of those days unsure about the nature and intentions of every person who walks through the kennel door.

What you learn, over enough of those mornings, is that comfort isn’t something you can explain into existence. You build it through behavior, through specific behavior, over specific time. The animal watches what you do, not what you intend. The shutter sound either becomes routine or it doesn’t, based on how many times you’ve fired it with nothing bad following, not based on whether you told them it was going to be okay.

This is one of the places where mirrorless cameras changed something real. An electronic shutter fires in complete silence. No mechanical sound, no auditory signal, nothing. The first time you photograph a cat using silent mode and the cat doesn’t move because, from the cat’s perspective, nothing happened, you understand immediately why that matters.

At some point, I started noticing something in executive portrait sessions that I had only previously encountered in shelter work.

The senior partner who sits down in front of the camera with the expression of someone waiting for a moderately unpleasant medical procedure. The real estate agent who has been photographed for fifteen years and has learned, over those fifteen years, to produce a version of her face that reads as friendly without actually being present. The physician who is technically sitting still and technically looking at the camera and is, in every frame, obviously somewhere else.

You might be able to fool some people with that. After enough hours studying faces on a screen, you don’t fool me. And in an era when everyone has been photographed more, through video calls and social media and LinkedIn profiles and every other surface that requires a face, the performance has gotten more practiced. More practiced performance is still performance, and a portrait needs a person, not a performance. That problem has gotten, if anything, harder to solve.

I don’t know exactly when I made the connection between the two. It wasn’t a revelation so much as a gradual recognition, the kind that arrives on the drive home after a session when you realize you’ve been solving the same problem in a different species. The nervous executive is not nervous in the way a frightened dog is nervous. But the underlying fact is similar: they’ve learned to produce a performance of being photographed rather than to actually be photographed.

The tools for addressing it are, I’ve come to understand, essentially the same.

Work slowly, not to seem patient but because working slowly is patience. Move the camera before it’s technically ready, because waiting for perfect stillness transmits your own tension back. Let silence do work. Don’t fill every quiet moment with direction. Ask questions that require actual thought, not the performance of thoughtfulness. Watch for the moment when the person in front of you forgets, briefly, that a camera is present.

That moment exists in most sessions. It doesn’t always arrive. It arrives more often when you’ve built the conditions for it than when you’ve tried to manufacture it by asking someone to relax, which is an instruction no one in the history of photography has ever successfully followed.

People occasionally ask whether I use treats in executive portrait sessions.

No, but the squeaker works great.

This lands differently in different contexts. To someone who doesn’t know the shelter work, it reads as a photographer being glib. To someone who does, it describes something real: the skill set I’ve built over twenty years photographing animals who can’t be coached translates, more directly than I initially expected, to photographing professionals who are performing being comfortable rather than actually being comfortable. The squeaker is shorthand for a longer argument about where the expertise lives and what it was built from. And it’s unexpected. Sometimes unexpected is the right tool in the camera bag.

This is the penultimate issue of what I’ve been calling the AI foundational arc. Over the past several months we’ve covered voice infrastructure, the generic photographer problem, persistent identity, recognizable output. The argument underneath all of it has been that AI tools work differently when there’s something real for them to work from.

What the patience transfer has to do with that is this.

The voice guide doesn’t capture my writing by analyzing my sentence patterns. It captures my writing because I know what I believe about photography and why, and most of that knowing came from mornings at shelters, from sessions that went nowhere useful, from dogs who spent ten minutes watching me before they decided I was worth approaching. The methodology, the voice infrastructure, the project knowledge, the feedback loop, those are a system for preserving something. The something came first.

I’ve been building that something since 2005. Some of it came from the MFA. A different part came from the shelter. And a specific part of it, the part that shows up most clearly in executive portrait sessions, came from the realization that the skills I was building by photographing animals who couldn’t be reasoned with were exactly the skills I needed for photographing professionals who had learned to reason themselves into a performance.

The squeaker is a joke about a real thing.

That’s usually where the useful insights live.

Schema markup is confirmed to help Google and Bing’s AI systems understand and represent your business more accurately. Both companies have said so directly. Google’s Ryan Levering described structured data as playing a “critical role” in grounding their generative AI systems. Microsoft’s Fabrice Canel stated explicitly that schema helps their LLMs understand content. A controlled experiment published by Search Engine Land found that only the page with well-implemented schema appeared in a Google AI Overview, while the page without it wasn’t surfaced at all.

What’s not confirmed: whether schema has any direct influence on ChatGPT, Claude, or Perplexity. Those systems haven’t said. The likely mechanism is indirect — schema feeds Google and Bing’s knowledge graphs, which then feed retrieval pipelines that AI generators draw from. Your JSON-LD doesn’t have a direct line to ChatGPT. It works by making your business more legible to the systems that supply the AI, not by talking to the AI itself.

That distinction matters because it changes what you’re building. Schema isn’t a shortcut to AI recommendations. It’s infrastructure that makes everything else you’re doing more interpretable to the systems that matter.

One more honest thing before the implementation details: what others write about you matters more than what you write about yourself. Research analyzing AI brand citations found that roughly 85% come from third-party sources — press mentions, directory listings, venue blogs, review sites — not from the brand’s own content. Schema makes your first-party content more legible. No amount of it replaces the third-party signal you earn by doing visible work in the world.

With that context established, here’s what I built and why.

Read more