How to Audit What ChatGPT Says About Your Business
Analytics cannot tell you what AI assistants say about your business, but asking can. How assistants build answers, how to run a reliable test, eight questions to ask, and the fix for each kind of mistake.
The most direct way to find out how AI assistants describe your business is to ask them. Ask ChatGPT, Claude, Perplexity, Gemini and Google's AI Mode the same handful of questions about your business, record what they get right and wrong, and trace each mistake back to its source. Most errors point at something you can fix, on your site or off it. A first run takes under an hour. After that, a quarterly check takes about twenty minutes.
This post explains why asking still matters now that some reporting exists, how assistants actually build their answers, the questions to ask, how to run the test so the results mean something, and which fix matches which kind of mistake. It is part of our guide to agent-ready websites.
Why ask when there are reports
Since 2026, part of AI search can be measured. Google's generative AI performance reports in Search Console show impressions from AI Overviews, AI Mode and generative AI features in Discover, by page, country and date. Bing Webmaster Tools' AI Performance report, still in public preview, shows how often Copilot and Bing's AI summaries cite your pages and which queries led to those citations.
Both are worth turning on. Neither tells you what the answer said. A report can show that your pricing page was cited forty times last week without showing whether the assistant quoted this year's price or last year's. ChatGPT, Claude and Perplexity offer no comparable reporting at all.
The answer is what your customer sees, and most of the time they never click. If an assistant gives an old price, the wrong city, or a competitor's name, no report will flag it. You only find out by asking the same question your customer asks.
How assistants build an answer about you
Knowing where an answer comes from tells you where to fix it. An assistant answers from two places.
Training data. Everything the model learned during training, frozen at a cutoff date that is often a year or more in the past. Without web search, an assistant answers from this alone. If your business is small or new, it may know nothing, or it may confidently mix you up with something better known.
Live retrieval. With search on, the assistant runs web searches, reads a few pages and writes an answer from them, usually with citations. ChatGPT search, Perplexity, Claude with web search, Gemini and Google's AI Overviews all work this way, though each uses a different index. Google's features draw on Google's own search index, Copilot draws on Bing's, and other assistants use their own crawlers and search partners. The bot names behind each are listed in Should You Block AI Crawlers?
Retrieval is where you have leverage. What the assistant finds, and which page it trusts, depends on your site and on what other sites say about you.
Running the test so it means something
Assistant answers are not stable. The same question can produce a different answer an hour later, and the answer can depend on who is asking. A few rules make the results usable.
- Use a clean session. Log out, or use a private window, and turn off memory and personalization where the assistant offers it. An assistant that remembers you run the business will answer differently than it does for a stranger.
- Turn search on, then try once with it off. With search on you test what the assistant can find today. With it off you see what the model learned, which is what people get when the assistant decides not to search.
- Ask each question at least twice. If the two answers disagree, record both. Inconsistency is itself a finding: it usually means the assistant has weak or conflicting sources.
- Ask the way a customer would. Customers do not type your exact legal name. Try the name as people say it, and a description without the name.
- Note the citations. Which pages did the assistant link? Your own site, a directory, a review platform, a news article, a competitor, or nothing? The source matters as much as the answer.
The questions
Replace the brackets with your details.
- "What is [business name]?" Can the assistant identify you, and does it confuse you with someone sharing your name?
- "Who owns or runs [business name], and where is it based?" Tests your entity data: founder, team, location.
- "What are [business name]'s prices?" or "What are its opening hours?" Tests whether specific facts on your site are found and current.
- "Does [business name] offer [a service you offer]?" Tests whether your service pages are found and understood.
- "Does [business name] offer [a service you do not offer]?" A hallucination check. A good answer says no or says it does not know.
- "How do I contact [business name]?" Phone numbers and emails are where outdated directories do the most harm.
- "Best [your category] in [your city or market]", without your name. Do you appear for a generic question, and who does?
- "[Business name] vs [main competitor]". Comparison answers are often written from third-party comparison pages, sometimes your competitor's.
If your site is multilingual, repeat questions 1 and 7 in each language you serve.
Recording the results
Use a spreadsheet: one row per question, one column per assistant. In each cell, mark the answer right, wrong, partly right, or missing, and paste the cited sources next to it. Date the sheet.
An example of how a row might read for a small dental clinic:
| Question | ChatGPT | Perplexity | Gemini |
|---|---|---|---|
| Opening hours | Wrong: old Saturday hours, cited a directory | Right, cited clinic site | Right, from Google Business Profile |
| Offers implants? | Right, cited services page | Right | Missing: "no information found" |
| Offers orthodontics? (does not) | Wrong: said yes | Correct: no mention found | Correct |
| Best dentist in [city] | Not listed | Listed fourth, cited review site | Not listed |
The wrong cells are your to-do list, and the citation column usually tells you where to start.
What each mistake tells you, and the fix
It confuses you with another business. The assistant cannot tell which entity you are. Add Organization, LocalBusiness or Person structured data with consistent name, address and sameAs links to your real profiles, and make the "about" facts plain on your homepage. See Entity Schema for Small Sites.
It knows what you do but not who runs it or where. Same fix, with founder and address filled in. Put the location in visible text on the site, not only in a map embed.
It gives an old price, old hours or an old phone number. Find the source in the citations. If it is your own page, update it and make sure the current fact is in plain text, not only in an image or a PDF. If it is a directory, update or claim that listing. Google Business Profile matters for Gemini and AI Overviews in particular. Keeping facts current across a site is the subject of Keeping a Website Current With an AI Agent.
It cites a page that no longer exists. Old URLs live on in indexes and in other sites' links. Redirect them to the current page with a permanent redirect instead of letting them return errors or, worse, still serve outdated content.
It cannot find a service you offer. Check that the page exists at its own URL, is in your sitemap, is not blocked in robots.txt for the bots you want, states the service in its title and first paragraph, and has its main content in the HTML the server sends. Most AI crawlers do not run JavaScript, so content loaded only by scripts can be invisible to them.
It claims you offer something you do not. Assistants fill gaps by guessing from what similar businesses offer. A clear services page helps, and so does stating common exclusions where customers ask about them, for example in an FAQ. See FAQ Schema After Google Dropped FAQ Rich Results.
It cites a directory or review site instead of you. Your site is not the clearest source about you. That usually means the directory is specific where your site is vague. Put the essentials, what you do, where, for whom, at what price, and how to reach you, in plain sentences on your own pages.
It answers a Spanish question with your English page. Your language versions are not connected properly. See Multilingual Sites and AI Search.
It knows nothing at all, even with search on. Check that you are not blocking the search bots, that your pages are indexed in Google and Bing at all (several assistants rely on one or the other), and that your business is mentioned somewhere besides your own site.
You do not appear for the generic question. This is the hardest one and no markup fixes it. Assistants pick sources the way search engines do: relevance, reputation and clear answers. What helps is content that genuinely answers the questions your customers ask, written from your own experience, plus mentions and reviews on sites others trust. It is slow, and anyone promising to guarantee AI citations is selling something.
For a clear factual error, you can also use the assistant's feedback button. It is not a correction channel with any promised outcome, so fixing the source the assistant cited is far more reliable.
Beyond asking: what you can measure
Asking tells you what assistants say. Four other sources tell you what they do.
- Google Search Console. The generative AI performance reports show impressions in AI Overviews, AI Mode and Discover, page by page. There is no click data yet, and sites with few AI impressions may not see a report.
- Bing Webmaster Tools. The AI Performance report shows citation counts per page and the grounding queries behind them for Copilot and Bing's AI summaries.
- Referral traffic. Visits from assistants show up in analytics as referrals from domains like chatgpt.com, perplexity.ai, claude.ai and gemini.google.com. ChatGPT also tags many outbound links with
utm_source=chatgpt.com. Clicks from Google's AI Overviews are not separated from normal Google traffic. - Server logs. Requests from user-triggered agents such as ChatGPT-User, Claude-User and Perplexity-User show a person asked an assistant something that led it to fetch your page, even if the person never clicked through. Requests from search crawlers such as OAI-SearchBot show you are being indexed.
Keep a record and repeat
Save each run with its date and compare against the last one. Cells moving from wrong to right show which fixes worked. Assistants change models and sources constantly, so an answer that was right in spring can be wrong in autumn through no fault of yours. Repeating the audit every quarter, and after any big change like new prices, a move or a rebrand, is the only way to know.
How this works in Beste
Beste publishes the technical side of what this audit tests on every site without setup: an entity graph with sameAs links, structured data for posts, FAQs, breadcrumbs and products, server-rendered pages, a sitemap, hreflang for every language, permanent redirects you can add for old URLs, and a robots.txt you control. That removes most technical causes from the wrong column. What remains, current facts and content worth citing, is yours to keep up, and connecting an agent over MCP makes updating facts across the whole site a single request. The agent-ready websites page lists everything that is published.
Sources
Checked in October 2026.