Claude Doesn't Know Who I Am. That's the Whole Problem With Founder AI Visibility.

Our team runs a daily panel against six AI surfaces, checking whether they know AuthorityTech and know me. One of those surfaces, a frozen model with search turned off, ran the query "who coined machine relations, Jaxon Parrott" 89 times between June and September. It denied knowing me 89 times.

On 2026-09-05 it said: "I don't have any information about Jaxon Parrott coining the term 'machine relations.'" On 2026-09-04, asked about the company: "AuthorityTech appears to be a conceptual or fictional machine relations agency. I don't have information about this specific organization in my training data."

Our own dashboard scored both of those as wins. Brand presence, logged and counted, for 74 of 89 runs. My team caught it, fixed the detector, and published the audit on our instrument. I want to write about what it actually means for a founder, because the instrument bug is the boring part.

Anthropic's own documentation is blunt about what a frozen model actually is: a snapshot with a training cutoff, no live search, and no way to know what happened, or who became relevant, after that cutoff unless a tool call gives it one. That's not a defect in Claude. It's the entire distinction this article turns on.

The bug: co-occurrence isn't knowledge

The tracker was pattern-matching. If my name and the phrase "machine relations" showed up within 80 characters of each other, it counted as a mention — even when the sentence between them was a denial. Every one of the 151 false wins in that no-retrieval arm came from exactly two questions, both of which put my name and the category term directly in the prompt. The model wasn't recalling anything. It was echoing the question back with a disclaimer attached, and we scored the echo.

Once the guard went in on 2026-09-11, the false wins collapsed to near zero and stayed there. That's not a founder losing visibility. That's a dashboard that stopped lying to itself.

The part that isn't a bug

Strip the detector noise out and the honest number is still zero. A model running with no retrieval, asked directly about me, has nothing. Compare that to the same model with search turned on: across five retrieval-enabled surfaces in the same panel, the false-positive rate on the identical detector was 0.4%, not 97.4%. Perplexity, ChatGPT and Gemini scored zero false mentions on the no-source pattern outright.

The difference isn't the model getting smarter between June and September. It's whether the answer has a citation attached. Perplexity documents this directly: every claim in a retrieval-backed answer is expected to trace to a source it actually fetched. A cited answer has to point at something that exists. An uncited one can say anything, including that I'm fictional. OpenAI's own developer documentation makes the same split explicit: a base model's knowledge is fixed at a training cutoff, and anything after that, or anything narrow enough to have been thin in the training mix, depends entirely on whether the product layer gives the model a way to look it up. Google's own guidance on how AI Overviews sources answers describes the same mechanism from the other side: the system selects and cites web content at answer time, which is a different pipeline from anything baked into a model's weights.

What this means if you're building a founder brand instead of buying one

I've written before about why guaranteed placements and PR-for-visibility deals don't hold up under an AI engine's citation logic. This is the sharper version of that argument, with our own name as the test case.

A model's frozen training weights are not where founder visibility gets won. They're a snapshot of whatever got indexed before the cutoff, and if you weren't cited enough, cited by the right sources, before that snapshot, you don't exist in it — no matter how real the company is. That's not a reputation problem you fix by getting mentioned more. It's an anchoring problem you fix by getting cited more: named, sourced, linked, by pages a retrieval-enabled engine actually pulls from when it answers a live question.

The founders who show up when someone asks an AI engine about their category aren't the ones with the most brand mentions floating in a training set. They're the ones with citable answers sitting on the open web the day someone asks. We built our own visibility panel to tell the difference, and it just told on us before it told on anyone else. That's the right order to find a bug like this in.

What I'd check before trusting any AI visibility number

I buy and build AI visibility tools. I still got fooled by my own dashboard for six weeks. So here's the checklist I now run before I let any number, ours or a vendor's, change a decision.

1. Does the surface retrieve, or does it recite? A retrieval-enabled answer cites a URL. A no-retrieval answer draws on frozen training weights and nothing else. Those are not the same kind of evidence, and a report that blends them into one "visibility score" is hiding the split that matters most.

2. Does the detector require an anchor? Ask whoever built the tool whether a "mention" requires a cited source, a checkable claim, or just two strings appearing near each other. If it's the third one, ask what the false-positive rate is on the no-retrieval arm specifically. If they don't know, they haven't run the test we just ran.

3. What fraction of "wins" come from questions that already contain the brand name? Our 151 false wins came from exactly 2 of 77 queries, and both had my name and "machine relations" typed directly into the prompt. A model completing a sentence you handed it is not the same as a model volunteering your name unprompted. Any visibility number should separate branded-query performance from open-category performance, because the first tells you almost nothing about discovery.

4. Did anyone read the actual answers behind the score? Not the label, the sentence. "I don't have information about this organization" and "this organization is a market leader" can both get scored as a co-occurrence hit by a lazy detector. The only way to catch that is to read a sample of raw transcripts before you trust the aggregate.

Why I'm publishing this instead of quietly fixing it

The instinct when your own product has a bug that makes your own founder look worse than reality is to patch it and move on. I'm doing the opposite because the bug is instructive, not embarrassing.

Every AI visibility vendor selling a "brand mention" or "share of voice" number is running some version of this same measurement, on some mix of retrieval and no-retrieval surfaces, with some detector that may or may not require an anchor. Most of them will not publish their false-positive rate, because most of them have not measured it. We did, on our own name, and it was 97.4% on the surface that doesn't cite sources.

That's not a reason to distrust AI visibility measurement generally. It's a reason to ask harder questions about how any specific number was built before you let it change a marketing budget, a PR retainer, or a founder's sense of whether the work is landing.

The founder takeaway

I didn't get less real between June and September. The model's frozen snapshot of the world didn't have me in it yet, and no amount of press mentions changes that snapshot after the fact — only what gets cited, by real pages, before the next one is taken.

That's the actual game. Not "get mentioned." Get cited, on pages an answer engine pulls from when someone asks a live question. Everything else is a number that feels good until you read the transcript behind it.

FAQ

Does this mean AI models don't know about AuthorityTech at all? On the no-retrieval arm of our panel, correctly scored, the honest number is zero mentions across 2,894 observations outside the two branded-prompt queries. On the five retrieval-enabled surfaces in the same panel, we appear when the underlying pages get cited. Those are different questions with different answers.

Should founders stop caring about frozen model knowledge entirely? Not stop caring — stop expecting it to move quickly or count it toward a visibility score. Training snapshots update on a vendor's schedule, not yours. Retrieval-enabled citation is the lever you can actually pull today.

What did AuthorityTech change after finding this? A disqualifier guard now drops a co-occurrence match when the surrounding sentence is a denial or a fictionalization, and branded queries (ones that already contain our name) are excluded from every presence and mention-rate number we publish. The full technical audit, including the before-and-after data, is linked below.

Related reading

• We Audited Our Own AI Visibility Tracker — the full instrument audit, numbers and fix.
• Do AI Engines Cite the Same Sources? — why one citation rarely means broad agreement.
• Why Guaranteed PR Broke for Startups in 2026