Gemini audited my website — and reviewed a different business entirely

Ask an AI to audit something you built and you are making one assumption before you read a word: that the model is looking at your thing, and not someone else’s. I asked Gemini Flash to review my website over the holiday. It reviewed a B2B consultancy called dixonai.com instead, praised a framework that isn’t mine, and never once flagged that it might have the wrong company. The report read like expert analysis. It was expert analysis. Just not of my site.

Here is the moment it broke. Gemini’s second recommendation was to add “vertical-specific playbooks” to my “GEFIE framework.” I have never heard of GEFIE. I do not have a framework with playbooks. What I do here is work out how to get reliable answers out of AI — real data, not hallucinations (the industry word for AI making things up) — and test the methods on real decisions, my own investing included. There is no GEFIE in any of it. So before I read any further, I checked which website Gemini was looking at.

The audit

The prompt was loose, and that’s part of the story. I’d asked, by voice, for feedback on “the website is Dixon Dixon AI” — almost certainly the transcription mangling “dixon.ai” as I said it out loud. Gemini took “Dixon AI” and resolved it to the entity with the bigger search footprint: dixonai.com, an AI consultancy run by a Rob Dixon, selling corporate AI training to CEOs and department heads. Different person, different business, different planet from what I do here. Then it wrote me a thorough, confident audit of that business.

It read well. Here’s the core of it, verbatim:

GEFIE is real — it’s a genuine dixonai.com framework. The “AI Hub” is what Gemini called a centralised portal product on that site. The corporate audience is theirs too. None of it is mine. The unnerving part isn’t that Gemini got facts wrong; the facts are mostly right. It got the subject wrong, and then reported accurate facts about the wrong subject with total confidence. That’s a harder failure to catch than a fabricated number, because nothing in the output looks broken. It doesn’t read as garbled.

It reads as a competent analyst who walked into the wrong meeting and kept talking.

The second session

Later in the holiday I went back with a tighter prompt and named the site outright — dixon.ai, spelled out, no ambiguity. This time Gemini found the right site. What came back was a mix — one thing right, one thing out of date, one thing wrong — and you couldn’t tell which was which from the output.

The out-of-date one first, because I owe Gemini some fairness here. It described dixon.ai as “leveraging frameworks like your ‘Filter Method’.” My methodology is called the Prompt Stack — but Filter Method was my early working name for it, from old conversations, before the site existed. The name never made it onto a single published page. Gemini hadn’t invented it; it had remembered it. What it didn’t do was flag that the name might be stale, or check it against the site it was supposedly auditing — which says Prompt Stack throughout. A model quietly correct about your own history and quietly wrong about your present is a strange thing to read. It also described the site as “practical developer-level prompt utility,” which misses who this is written for — anyone who wants reliable answers out of AI, not developers.

And then it caught something real:

That flag is correct. My own Google Search Console data backs it up: “dixon ai” searches reach this site at around position four, which means the first results those searchers see are mostly the other site. It’s genuine, useful information for anyone building under a contested name, and I’ve logged it as a real catch in the /catches record and as a documented error on /lessons. The problem is the framing it arrived wrapped in. The brand-collision risk is true; the “Filter Method” is a name from another era, presented as current; the “developer-level prompt utility” audience is wrong.

All three came in the same fluent voice, and nothing in the response told me which to trust.

The hallucination that doesn’t look like one

Search for a Gemini hallucination example and you’ll mostly find invented citations and made-up numbers — failures that announce themselves once you check. This one didn’t. The first session failed at the only step that has to be right before any other step matters — which thing it was reviewing — and produced no signal at all that it had. The output was thorough and well-organised the whole way down. It just happened to be about a business that isn’t mine.

So I now do one thing before I read any AI audit of my own site or work: I ask the model to state the exact URL it’s describing and one specific fact about it I can check in under thirty seconds. If either is wrong, I start over. If both are right, at least the entity question is settled and the audit is about the correct subject — which is the bare minimum before any of the rest is worth my time. None of this makes Gemini bad at everything; on calendar-sensitive research it’s been the most useful tool I’ve tried — that test is documented in the multi-model comparison. This is a different failure mode, on a different task.

Field Report

The most accurate thing Gemini told me about my own site across two sessions was that another business owns my name in search. It’s the one thing I already knew.