What Does AI Say About Your Company? How to Find Out

An AI answer about your company comes from one of two places. The model either answers from what it learned in training, or it searches the web first and writes from the pages it fetched. That choice is made per question, and you usually can't see it.

It helps to split the process into four stages:

• Parametric. What the model knows with no web access. It is frozen at a training cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path behind chat-latest. Anything you changed after those dates can't appear in an answer given without search.
• Access. Whether the engine's crawler may fetch and store your pages at all.
• Retrieval. Whether your page enters the model's context for a given question. OpenAI's help page says ChatGPT search rewrites a prompt into one or more targeted queries, so the query that finds you is not the one you typed.
• Generation. Which retrieved pages the model actually uses and cites.

This changes how you test. The same question can get a stale answer from training on one run and a current, cited answer on the next. A visible list of sources is the best sign the engine searched. No sources usually means you are reading the parametric answer. Record which one you got, or you can't tell an outdated answer from a wrong one.

Before you ask any engine anything, write down the truth. You can't grade an answer without a reference.

The fact sheet

One page covering the facts buyers, journalists and investors ask about:

• Legal and trading name, plus any former names
• What you sell, in one sentence
• Pricing tiers and current prices
• Founding year, headquarters and owner
• Key product names and what was discontinued
• Whether you are operating, and where
• The competitors you are most often confused with

For each fact, note the URL on your site where it lives and the date it last changed. That date matters later, when you decide whether an error is outdated or invented.

The prompt set

Keep it to roughly 10 to 20 prompts so you can repeat it. Mix three kinds:

1. Identity prompts: "What is [Company]?", "Who owns [Company]?", "Is [Company] still in business?"
2. Fact prompts: "How much does [Company] cost?", "Where is [Company] based?"
3. Buyer prompts: questions a prospect asks without your name in them.

Say you run a dental clinic in Austin. An identity prompt is "What does [Clinic] offer?" A buyer prompt is "Which dental clinics in Austin do same-day crowns?" Identity prompts show whether the facts are right. Buyer prompts show whether you are named at all.

Freeze the wording. If you change a prompt between runs, you are measuring the prompt, not the engine.

Test each engine separately: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Claude and Microsoft Copilot. Treat AI Overviews and AI Mode as two surfaces. Google's documentation says they may use different models and techniques and can return different links.

Use the same conditions every time:

• Logged out, or a clean account with memory off and no custom instructions. OpenAI says ChatGPT may use saved memories when it rewrites a search query. Your own history skews the result, which is why logged-in and logged-out testing give different answers.
• Same location and language. AI Overviews don't appear on every search and vary by location.
• Run each prompt more than once. One answer is an anecdote.

Log every answer in a spreadsheet with these columns: date, engine, model or plan shown, prompt, full answer text, whether sources were shown, cited URLs, and your error classification.

If you want a quick first read before building the log, the free checker on this site asks up to three buyer questions and scores the answers. It is a snapshot, not continuous tracking. Your own log is what you compare month to month.

A list of "wrong answers" is not useful. Each error type points to a different cause and a different fix. Grade each flagged answer by type, then by severity.

The split between invented and stale is covered in AI hallucination vs outdated answers. For the empty-answer case, see why ChatGPT says it has no information about your company. A false closure claim has its own causes, set out in why AI might say your company closed.

Then add severity. Critical errors change a buying decision: closed, acquired, wrong price, wrong product. Moderate errors mislead but don't stop a sale. Minor errors are wrong details. Fix critical errors first, whatever engine they appear on. For a closer look at grading ChatGPT specifically, see how to tell whether ChatGPT is accurate about your company.

The evidence is clear about where your control starts and stops. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found no technique with a stable, cross-platform causal effect on whether engines discover a page. The levers that hold up remove blocks.

In your control

• Crawler access, bot by bot. OpenAI's crawler page says each setting is independent: GPTBot governs training use, while OAI-SearchBot controls whether you appear in ChatGPT search answers. Blocking all AI bots in one rule can remove you from ChatGPT search.
• Firewall and CDN reachability. OpenAI's help page says eligibility needs both: allow OAI-SearchBot, and confirm your host or CDN allows its published IP addresses. A firewall can turn away a crawler your robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
• Server-rendered facts. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. A price that appears only after scripts run may be invisible to them.
• Accidental suppression. Google documents noindex, nosnippet, max-snippet and data-nosnippet as controls that keep content out of AI features. Templates sometimes apply them by mistake.
• Consistent, specific facts. State each fact once, plainly, in HTML that loads without scripts. A specific sentence gives the engine something it can quote.
• Third-party mentions. When engines cite other sites instead of yours, being named correctly on those sites is associated with appearing in the answer. It is an association, not a guarantee.

Not in your control

• What the model learned before its cutoff
• Whether the engine searches for a given question
• Which pages it retrieves and which it cites

Schema markup, llms.txt files and "AI-friendly" rewriting are not levers for AI answers. Google says structured data isn't required for generative AI search and that Google Search ignores llms.txt. Existing markup does no harm, but adding it won't fix a wrong answer.

Once your log shows the errors, work in this order. Each step removes a cause the next step depends on.

1. Fix your own site first. Find every page that states a fact the engines got wrong. Make them agree. A price that differs between the pricing page and the footer gives an engine two answers to choose from.
2. Check crawler access. Read your robots.txt for each retrieval bot: OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot and bingbot. Then confirm your firewall actually lets them through.
3. Check rendering. View the raw page source of your pricing, about and product pages. If the key facts aren't in it, they need to be rendered on the server.
4. Check snippet controls. Search your templates for nosnippet, max-snippet:0, noindex and data-nosnippet on pages that should appear.
5. Check index status per engine. Look at Google Search Console and Bing Webmaster Tools separately. In Otterly's 2026 experiment, Google deindexed both test sites and Google's AI surfaces stopped citing them, while Bing kept them and Copilot kept citing them. One healthy index tells you nothing about the other.
6. Correct the third-party pages engines cite. Your log's cited-URL column shows which directories, reviews or articles feed the wrong answer. The playbook on responding to negative AI answers covers this step.
7. Re-run the same prompts. Search-backed answers can change after a recrawl. Answers from training data change only when a new model ships. For ongoing monitoring, see tools that alert you to false AI claims.

• Testing while logged in. Memory and custom instructions shape the answer you see. Your buyers don't have your history.
• Letting one engine stand for all. Gemini, AI Overviews and AI Mode are three Google products. ChatGPT and Copilot draw on different indexes. Log each one separately.
• Blending everything into one score. A single number averages answers given with and without search across engines that cite in different ways. It can hide one engine that gets you badly wrong.
• Treating a citation as proof. A cited page is not always the page that shaped the answer, and a cited answer can still be wrong. Read the text, not just the link list.
• Reacting to one answer. Answers vary between runs. Act on errors that repeat.
• Changing the wrong robots.txt rule. Allowing GPTBot doesn't help ChatGPT search find you, and blocking it doesn't remove you from search. OAI-SearchBot is the search control.
• Trying to game the answer. Hidden text, cloaked pages for AI crawlers and instructions planted for models are manipulation. Research (arXiv 2606.16821) shows some AI search agents can be swayed by manipulated content, which is exactly why engines and search providers treat it as abuse. It puts your index status and your reputation at risk. Fix the facts instead.

How often should I re-check what AI says about my company? Monthly is a sensible default for a fixed prompt set, plus an extra run after any change to pricing, ownership, name or products. Engines update models and recrawl pages on their own schedules, so a clean result in one month doesn't carry forward. Keep the prompt wording frozen between runs so differences reflect the engines, not your questions. Log the date on every answer.

Why do ChatGPT and Perplexity give different answers about my company? They use different models, different search providers and different rules for when to search. One may answer from training data while the other fetches your current pages. They also cite differently, so the sources behind each answer won't match. That's why each engine needs its own row in your log and its own read of what went wrong.

Will adding schema markup fix a wrong AI answer? The evidence doesn't support it. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm. To fix a wrong answer, correct the fact on your own pages, make sure crawlers can reach and read them, and correct the third-party pages engines cite.

How long does a fix take to show up in AI answers? Nobody publishes a timeline, and it depends on the stage. An answer built from a live search can change once the engine recrawls the corrected page. An answer from training data won't change until the vendor ships a model trained after your fix. Check whether the wrong answer showed sources: that tells you which kind of wait you face.