AI visibility check: what ChatGPT says about your business, measured.
One test chat proves nothing. Here is what a real check measures, how to run one yourself in an afternoon, and the mistakes that make DIY checks fail their owners.
Jarvis Report · 2026-08-26 · built from live public data — we never invent a score
A customer asks ChatGPT for the best option in your category and your city, and the answer names 3 businesses as if that were the whole market. Every owner eventually asks the obvious question — am I in there? — and most answer it the wrong way: they open a chat window once, ask once, and walk away either falsely reassured or falsely alarmed. An ai visibility check done properly is a measurement, and measurements have rules.
Why one chat is not a check
Assistant answers vary between runs. The same buyer question, asked twice in fresh sessions, can return overlapping-but-different shortlists. That is not a bug you can wait out; it is how these systems behave. It has two consequences for measurement.
First, a single mention is not visibility — it may be a lucky draw. Second, a single absence is not invisibility — same reason. The unit of measurement is the pattern across repeated runs: in how many runs do you appear, in what position, described how? Appearing in most runs is a defensible answer. Appearing once in five is a coin toss that happened to land your way.
There is a third state people miss entirely: the assistant sometimes declines to recommend anyone. A refusal is a different result from a run where rivals are named and you are not. The second is the painful one — the assistant was willing to answer, had names available, and yours was not among them. A check that collapses refusals and misses into one bucket will misdiagnose you.
What a real check measures
1. The buyer questions, not your name
Asking "what do you know about Smith & Sons Plumbing" tests the wrong thing — customers do not ask that. They ask "best plumber near me for a burst pipe," "affordable emergency plumber in Mesa," "who should I call for a water heater install." A check runs the questions buyers actually ask, phrased several ways, because assistants weight phrasing more than owners expect.
2. Presence, position, and framing
Being named is one axis. Being named first, with a confident sentence of justification, is another. Being named with a hedge — "some customers mention slow response times" — is visibility working against you. Record all three, per run.
3. Who wins instead of you
The names the assistant returns are your real competitive set, and it is routinely different from the set owners assume. Once you know who actually gets recommended, you can compare their public evidence — review mass, listing consistency, crawler-readable sites — against yours and see the gap in concrete terms rather than vibes. Closing that gap is what the recommendation playbook is for.
4. What the machines can read about you
Half a check looks at answers; the other half looks at inputs. Can an AI crawler actually read your website, or does it receive an empty script shell? We measured our own site serving the same 5,034-byte empty frame on every page to AI crawlers while humans saw a full product — a failure invisible to everyone who only ever browsed it. Do your listings agree on your name, address, and phone? Is your profile complete enough for an engine to categorize you with confidence? Answer-side symptoms have input-side causes; our guide on the six reasons AI skips a business maps them.
Run it yourself: the afternoon version
The manual version is genuinely doable, and the table is the whole procedure. Budget a few focused hours the first time.
| Step | What to do | What to write down |
|---|---|---|
| Questions | Write 5 questions a real buyer would ask, in natural language, with your city in them | The exact wording of each |
| Runs | Ask each question in a fresh chat session, several times, across at least two assistants | Every business named, in order, per run |
| Refusals | Note when the assistant declines to name anyone | Refusal, separately from a miss |
| Framing | Copy the sentence the assistant attaches to each name | Praise, hedge, or neutral |
| Crawler test | Fetch your homepage with a bot user-agent, no JavaScript | Does readable body text come back? |
| Listings | Compare name, address, phone across your top directories | Every mismatch, however small |
| Tally | Count appearances per question across runs | Your rate, and who beat you |
Re-run the same protocol on a calendar — quarterly at minimum, monthly if you are actively fixing things. Keep the question wording identical between rounds, or you are measuring your prompts instead of your progress.
The mistakes that make DIY checks lie
Testing inside your own account. Assistants with memory of you will answer differently than they answer a stranger. Use a fresh session with no history — what the anonymous buyer sees is the only view that matters.
Asking about yourself by name. Recognition is not recommendation. An assistant can describe your business fluently when asked directly and still never surface it for a buyer question.
Stopping at one assistant. ChatGPT, Gemini, Copilot, and Perplexity retrieve differently. Owners are sometimes strong on one and absent on another; a one-assistant check hides that split.
Treating the run count as a score. Five runs is a signal, not a laboratory. Enough to distinguish "reliably present" from "reliably absent" — the honest resolution of a manual check — but do not read false precision into small differences.
Or let the measurement run itself
The manual protocol is real work, repeated forever. We built Jarvis Report to run exactly this check automatically: the live buyer questions for your city and category, repeated runs, refusals recorded as refusals, your rivals' numbers pulled live, your listings and crawler-visibility read the way the engines read them — in about a minute, free. Anything we could not measure is dropped from the weighting, never scored as zero; the pricing page shows what stays free.
Run the free AI visibility check on your business →Live public data only. We never invent a score.
Questions owners ask us
How often should I run an AI visibility check?
Quarterly if things look healthy; monthly while you are actively fixing causes. Models update, rivals accumulate reviews, and directories go stale on their own — the answer you verify this month can drift by next quarter without you touching anything.
My check shows I'm invisible. Now what?
Diagnose before spending. The causes range from a crawler-unreadable website to fragmented listings to thin review evidence, and the fix order depends on which applies. Start with the 20-minute diagnostic, then work the playbook.
Can an AI visibility check hurt me somehow?
Reading public data changes nothing about the data. Asking assistants buyer questions is what thousands of real buyers do daily; your checks disappear into that noise.
Do I need a tool for this or is manual enough?
Manual is enough to learn where you stand once. The value of tooling is repetition without drift: same questions, same method, every month, with the boring parts — listing comparison, crawler fetches, rival tracking — done for you. Use whichever you will actually keep doing.