Roughly forty questions your customers would actually ask, put to ChatGPT and Claude, with every answer handed to you word for word.
An AI visibility audit asks the questions your customers actually type into AI assistants, records every answer word for word, and hands you the files — not a score. Ours costs $390, once, and takes about a week. Roughly forty questions are written for your trade and your town, each asked three separate times, on the two assistants we live-test: ChatGPT and Claude. What comes back to you is a dashboard, the complete machine-readable archive (one line per call: the full answer text, every link, the model identifier, the timestamp), a score against a rubric that is published in full on this site, and one forwardable ticket per fix worth making. Three assistants — Gemini, Perplexity and Google AI Overviews — are not connected, there are no customers yet, and both of those facts are on the front page for the same reason they are on this one.
| You receive | What is in it |
|---|---|
report.html / report.md |
The dashboard: mention rate, share of voice against the competitors you name, citation rate, average position, and the most-cited source domains — where the assistants actually went for information. |
raw.jsonl |
One line per call. Full answer text, every link the assistant showed, every page it read, model identifier, timestamp, latency. This is the file that makes every number above checkable. |
tickets/ |
One ticket per fix: why it matters, what to do, how to verify it worked, roughly how long it takes. Written to be forwarded to whoever runs your website, with no part of the report attached. |
| The score | 100 points across five dimensions: 50 computed from the run and from probes against your site, 50 from intake answers you give us. Every sub-score carries the evidence it was derived from. |
The pipeline has been run end to end once, on 15 August 2026, against a test business. Here is what that produced, counted from the output directory rather than described:
gpt-5-6-sol and
claude-sonnet-5, both with web search on.The full sample, redacted, is on the GEO audit page along with the scoring rubric it was measured against.
On that run the tool initially reported a mention rate of 33.3% — 8 of 24 answers contained the business name. The business was invented for testing. Reading the archive showed what had happened: the assistant had matched the name to a similarly-named practice in another country and answered about that one.
“is Bayview Dental Studio good”
“Do you mean Bayview Dental
Studio in South Surrey/White Rock, BC, run by Dr. [name]?
If so,
it appears generally reputable but not unanimously reviewed
…”
The doctor’s name is removed — he is a real person and did not ask to be in anyone’s marketing. Everything inside the quotation marks is otherwise word for word.
A string match is not an identification. The tool now reports confirmed mentions separately from name-collision ones, and on this run the confirmed rate is zero with 8 collisions recorded. A mention rate is the number an audit uses to persuade you to spend money on fixes; if it is inflated, everything downstream of it is too. The reason you can read this paragraph at all is that the transcripts were kept.
Question sets are built to an intent mix of 40% discovery / 25% comparison / 25% problem / 10% brand. Discovery questions carry no business name at all — they are the ones that decide whether a stranger finds you:
| Intent | Example from the sample run | What it tells you |
|---|---|---|
| discovery | “best dentist in Ballard Seattle” | whether a stranger finds you at all |
| comparison | “Bayview Dental Studio vs Ballard Family Dental” | where you land against the businesses you name |
| problem | “how much does Invisalign cost in Ballard Seattle” | the long-tail questions about one service |
| brand | “is Bayview Dental Studio good” | how the assistant describes you, and what it gets wrong |
You can also do a version of this yourself before paying anybody: the question pack builder generates the same kind of set for your business and gives you a tally sheet, free and with no signup.
$390, once, invoiced by email — not a subscription. Nothing is charged before you have seen a free sample check on your own business. Turnaround is about a week: one assistant averages over two minutes per answer, so a full run is measured in hours rather than minutes, and same-day delivery is not on offer here.
Start with the free check. Send your business name, your town and your website. You get back the actual questions, the actual answers, and where you did or did not come up. No charge, no call, no signup.
hello@doestheaiknowyou.comOne person reads these. Replies take a day or two.
A measurement of what AI assistants say when someone asks the questions your customers ask. Ours asks roughly forty questions built for your trade and town, three times each, on ChatGPT and Claude, and hands you every answer verbatim with the links, model version and timestamps.
$390 once, invoiced by email, about a week. There is a free sample check first and nothing is charged before you have seen it.
ChatGPT and Claude are live-tested. Gemini, Perplexity and Google AI Overviews are not connected, and a re-run of the same questions is included when they are.
A dashboard, the full machine-readable archive of every call, a score against a rubric published on this site, and one forwardable ticket per fix. On the sample run, 4 findings produced 6 tickets and 4 went into the package.
Yes. Ask your own discovery questions in the assistant and keep the answers. The free question pack builder on this site generates the question set and a tally sheet, and the free sample check does one round for you.
No customers yet, so no case study. The only run this pipeline has done end to end is the sample described above, and it is published including the parts that came out incomplete.
manifest.jsonl (models, prompt count, timings), raw.jsonl (every answer verbatim), findings.json (4 findings) and the ticket directories. The counts on this page are read from those files at build time.geo_radar/audit.py: five dimensions, 14 sub-scores, 50 of the 100 points computed rather than supplied. It is reproduced line by line on the GEO audit page.