Monitoring What ChatGPT Says About Your Brand

Several things change what a person sees when they ask ChatGPT about your company. None of them is visible in the answer itself. For the full picture of how answers get built, see how ChatGPT describes your brand, and why.

The plan and the model

ChatGPT routes people to different models depending on their plan. OpenAI's model docs give the GPT-5.6 family a knowledge cutoff of February 16, 2026. The Instant model behind the chat-latest alias has a cutoff of August 31, 2025. One path can know about roughly five and a half months of company news that the other cannot. Secondary reports, as of September 2026, put Free and Go users on GPT-5.6 Luna and Plus users on GPT-5.6 Sol. OpenAI has not confirmed that mapping on its own pages.

Whether search ran

ChatGPT decides, question by question, whether to search the web. When it does not, the answer comes from what the model learned in training, frozen at its cutoff. When it does, OpenAI says ChatGPT "rewrites your query into one or more targeted queries" and sends them to search providers. Published estimates of how often that happens disagree sharply. Profound puts it at about 18% of conversations, Semrush's US clickstream data put it at 34.5% of queries in February 2026, and cloro reports 80.5% on its own prompt set. They measure different things, so none of them is the rate for your category.

Memory and custom instructions

OpenAI's help page states that "if memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." Product results also take memory and custom instructions into account. A check from your own account therefore partly reflects you: your past chats and your saved context.

Ordinary variation

Even with everything else held fixed, a generated answer is not guaranteed to repeat word for word. One run is an anecdote, not a measurement.

The most useful split for monitoring is whether search ran. The two kinds of answer draw on different material and respond to different fixes.

This is why a fix can appear to work for one prompt and not another. A corrected pricing page can show up in searched answers while unsearched answers keep repeating the old price. For realistic timelines, read how long it takes AI to update what it says about your brand.

One more caveat applies to searched answers. In OpenAI's API, the list of sources the model consulted is returned separately from its citations, and OpenAI notes the number of sources is often greater than the number of citations. A page can shape an answer without being cited, so citation counts undercount influence.

Be plain with yourself about the split. It decides where effort pays off.

In your control

• Search crawler access. OpenAI's crawler docs separate the bots. OAI-SearchBot controls whether your site can appear in ChatGPT search answers. GPTBot governs training use only, and OpenAI says each setting is independent of the others.
• Network reachability. OpenAI's help page says eligibility needs both: allow OAI-SearchBot in robots.txt, and make sure your host or CDN allows traffic from OpenAI's published search bot IP addresses. A firewall can return a 403 to a crawler that robots.txt allows, and nothing in your analytics will flag it.
• Content in the HTML. Vercel's crawler study found ChatGPT's crawlers fetch JavaScript files but do not execute them. A price that only appears after JavaScript runs may be invisible to them.
• The facts on your own pages. A specific fact stated plainly on a fetchable page (your price, your founding year, a one-line description) gives the engine a sentence it can quote.
• Where you are named. When ChatGPT cites other sites instead of yours, being named accurately on those sites is associated with appearing in the answer. That is an association, not a promised effect.

A quick robots.txt rule that allows the search crawler looks like this:

User-agent: OAI-SearchBot
Allow: /

To see whether your firewall actually lets each AI crawler through, the AI Crawler Checker requests a page as each crawler and shows which ones get turned away.

Out of your control

• Which model a given user is routed to.
• Whether ChatGPT decides to search for a given question.
• How it rewrites the question into search queries.
• How results are ranked. OpenAI's full disclosure is that ChatGPT ranks results "using multiple factors" and that "placement is not guaranteed."
• What the current model already learned in training.

The goal is a baseline you can compare month to month. A logged-out session is one configuration, not what every customer sees, but it removes your own memory and history from the result.

1. Write a fact sheet. List the facts a buyer would ask about: plans and prices, founding year, headquarters, product names, ownership, key integrations, and things you do not offer. This is your answer key. The method for judging answers against it is covered in how to tell whether ChatGPT is accurate about your company.
2. Write a fixed prompt set. Mix three types: questions that name you ("What does [brand] cost?"), category questions that do not (say you run a dental clinic software company: "What scheduling software do small dental clinics use?"), and comparisons against alternatives. Keep the wording identical every run.
3. Use a clean session. Open a private browser window, stay logged out, and run from the same country and network each time. No memory, no custom instructions, no earlier chat in the same thread.
4. Run each prompt more than once. Three runs per prompt shows you which claims are stable and which flicker.
5. Record everything. For each answer, save the date, the prompt, the full text, whether citations appeared, the cited URLs, any model label shown, whether your brand was named, and each factual claim made about you.
6. Keep a schedule. Run monthly, and again after any change ChatGPT should pick up: a price change, a rebrand, a new product.
7. Optionally, add a controlled run. OpenAI's API lets you run the same prompt with web search on and off and compare the two. The API is not the consumer app, so label these results separately.

If you want a quick first snapshot before building this, the free checker on this site runs up to three buyer questions and scores the result. It is a snapshot, not continuous tracking.

A pile of screenshots is not a result. Turn each run into numbers you can compare.

• Classify each claim. Mark every factual claim about you as correct, wrong, outdated, or unverifiable. Unverifiable is different from wrong: ChatGPT stating a headcount you never published is a gap in your fact sheet, not proof of an error.
• Weight by damage. A wrong price or a false acquisition matters more than a slightly off headcount. Decide your weights once and keep them fixed.
• Report a rate, not a count. Divide errors by answers returned. A raw error count rises when you add prompts, which makes a bigger check look worse for no real reason.
• Split by search state. Report searched and unsearched answers separately. Averaging them hides which fix applies.
• Track presence, not position. Record whether you were named and cited. No ground truth exists for a "rank" inside a ChatGPT answer, so any rank number is a guess.

If manual scoring becomes the bottleneck, see what tools can alert you when AI says something false.

Work through the causes in this order. Each step rules out a cheaper explanation before you spend effort on a harder one.

1. Check whether search ran. If the wrong answer has no citations, it probably came from training data. You cannot patch that directly. Make sure the correct fact is on a fetchable page so searched answers can get it right, and expect unsearched answers to lag until a newer model.
2. Open the cited pages. If search ran, read what ChatGPT cited. If it is your own page, the old fact may still be there. If it is a third-party page, the error is theirs, and asking them to correct it is the fix.
3. Look for contradictions on your own site. The same price stated differently on the pricing page, an old blog post and the footer gives the engine two answers to choose from. Pick one and update the rest.
4. Check for old coverage. Archived press and outdated reviews can keep a retired fact alive. See whether old news stories shape what AI says today.
5. Check access. Confirm robots.txt allows OAI-SearchBot, your firewall returns a 200 to it, and the fact is in the HTML before JavaScript runs.

Then rerun the affected prompts on your normal schedule and compare against the baseline. A fix makes a correct answer possible; it does not guarantee one.

• Checking from your own account. Memory and custom instructions can shape the result. Your logged-in view is not your customer's view.
• Trusting one run. A single screenshot shows one sample of a variable process. Run each prompt several times.
• Blending searched and unsearched answers. They respond to different fixes on different timelines. Keep them apart.
• Confusing the crawlers. Allowing GPTBot does not get you into ChatGPT search, and blocking it does not remove you. OAI-SearchBot governs search.
• Chasing tactics without evidence. Schema markup, an llms.txt file, and rewriting pages in an "AI-friendly" style have no credible evidence behind them for getting cited. A July 2026 survey of 45 GEO studies (arXiv:2607.14035) found no technique with a stable, cross-platform causal effect on discoverability.
• Monitoring only ChatGPT. Other engines draw on different indexes and can say different things about you. For a wider view, read what to monitor across AI brand monitoring tools or start with free AI brand monitoring across several engines.

Can I see exactly what ChatGPT tells a specific customer? No. What a customer sees depends on their plan, the model they are routed to, whether search ran, and their own memory and custom instructions. None of that is visible to you. A logged-out check gives you a stable, comparable baseline. It is one configuration, so treat it as a consistent sample of how ChatGPT describes you, not a copy of any individual's answer.

How can I tell whether ChatGPT searched the web for an answer? Look for citations or source links in the answer. If they appear, search ran and you can open the cited pages to see where each claim came from. If there are none, the answer most likely came from the model's training data. Record this for every answer, because the two kinds of answer need different fixes.

Does blocking GPTBot remove my brand from ChatGPT? No. According to OpenAI's crawler documentation, GPTBot governs whether your content may be used for training, and OAI-SearchBot governs whether your site can appear in ChatGPT search answers. OpenAI states each setting is independent. Blocking GPTBot does not remove you from search, and allowing it does not put you there.

How often should I run a ChatGPT brand check? Monthly is a sensible baseline for most brands, with an extra run after any change you expect ChatGPT to reflect, such as new pricing or a rebrand. Keep the prompts, session setup and scoring identical each time. Consistency matters more than frequency, because the value of monitoring is the comparison between runs.

Is checking through the OpenAI API the same as checking the ChatGPT app? Not quite. The API lets you control whether web search runs, which makes clean comparisons possible, and it returns the sources consulted separately from citations. But the app routes users across models and decides on its own when to search. Use API runs for controlled tests and label them separately from app results.