How to check what AI models say about a business
By Johnny Lagneau · Updated September 8, 2026Originally published January 28, 2026
A single ChatGPT answer can reveal a useful question, but it is not a complete visibility audit. Start by defining the business and market, then record how several models answer the same buyer questions under comparable conditions.
The steps below produce a record you could hand to someone else and have them repeat. That is the whole point: a check you cannot repeat is an anecdote.
1. Define the business and market
Write down the website URL, the exact services, the location and the market context before you type anything into a model. If the business is a client, do not paste private client information into a public example — the questions should only use what a buyer could know.
2. Write buyer questions before checking answers
Three kinds, fixed in advance:
- Unbranded, for discovery. “Who does X near Y?” The business is not named; you are checking whether it appears.
- Named, for description. “What does business Z do?” You are checking accuracy.
- Comparative, for recommendation. “Z or W for X?” You are checking preference and the reason given.
The questions above are fictitious examples written for this guide.
3. Record the test conditions
For each answer, note the model and the version shown, the date, the exact question, the conversation context (a fresh conversation, or one with prior messages), and whether web search was enabled. Do not adjust the protocol between models to make a result look better; if you change a question, it is a new test.
4. Keep the answer and its available sources
Save the answer text as given, or an export of it, rather than a paraphrase. If the model provided links, record them and then open them: a cited page that does not say what the answer claims is itself a finding. If the model gave no source, write “no source given” — do not go and find one for it.
5. Classify the finding
One label per answer:
- Absent. The business is not in the answer.
- Named. Present, with no claim about it.
- Described incorrectly. Present, with a claim that contradicts a verified fact.
- Recommended. Put forward when a choice was asked for.
- Missing. The call failed or returned nothing usable. Never counted as absent.
6. Investigate before prescribing a fix
An incorrect description is a comparison between what the model said and what the business’s own pages say. Read the pages. If the page carries the wrong fact, that is the fix. If the page is right and the model is wrong, the fix may be elsewhere, or there may be no page-level fix at all. Propose a precise action only when the evidence supports one.
7. Repeat the check and document the limits
Repeat with the same questions, the same context and the same configuration, and note the model versions each time. Expect variation; models are not deterministic. And say what the sample is: a handful of questions on a date, not every buyer.
A record template
A blank form for one answer. It records nothing and sends nothing; copy the fields into your own sheet.
| Field | What goes in it |
|---|---|
| Business and market | URL, services, location, context |
| Question | Exact wording, and its kind: unbranded, named, comparative |
| Model and version | As displayed at the time |
| Date | When the answer was produced |
| Configuration | Web search enabled or disabled; fresh conversation or not |
| Answer | Verbatim text or export |
| Sources given | Links the model provided, and whether each was checked |
| Classification | Absent, named, described incorrectly, recommended, missing |
| Investigation | What the business’s pages say about the claim |
| Next step | A specific action, or “none warranted” |
What the useful output is
The useful output is not just a verdict about visibility. It is a dated answer, the evidence you could verify, and a next step you can explain. That is also the shape of a Brand Snapshot; the method behind it is on how we measure.