How can I tell whether what ChatGPT says about my company is accurate?

You need a reference, a fixed test, two conditions and a grading rule. Here is the whole process.

1. Write a brand fact sheet. List 10 to 20 facts you can prove: what you sell, who it's for, pricing tiers, founding year, headquarters, ownership, current product names and features you've retired. Say you run a hypothetical payroll startup in Denver. Your sheet would include your plan prices, the states you serve and the integration you dropped last spring.
2. Write a fixed question set. Pick five to ten questions a buyer would ask: "What does [company] do?", "How much does [company] cost?", "Who are [company]'s competitors?". Save them word for word. Changing the wording between runs changes the answer.
3. Run each question twice. Once where ChatGPT searches the web and shows cited sources, and once where it answers from what it already knows. Use a logged-out session or turn memory off.
4. Grade every claim. Break each answer into single statements and mark each one supported, wrong or unverifiable against your fact sheet.

The result is a list of specific claims, each tied to a condition and a date. That tells you far more than "ChatGPT gets us wrong". For background on where these descriptions come from, see how ChatGPT describes your brand, and why.

ChatGPT answers about your company through two different channels.

Without search: training data

With no web access, the model answers from what it learned before its knowledge cutoff. OpenAI's model documentation lists a Feb 16, 2026 cutoff for the GPT-5.6 family and an Aug 31, 2025 cutoff for the Instant path (chat-latest). So "what ChatGPT knows" is not one fixed thing. Anything you changed after the cutoff can't appear in these answers.

With search: the live web

With search, ChatGPT fetches pages and cites them. OpenAI's help center says ChatGPT search typically rewrites your question into one or more targeted queries. It also says saved memories may shape that rewriting when memory is on. That's why a logged-in test can differ from a buyer's view. The logged-in vs logged-out comparison covers this in detail.

As of October 2026, ChatGPT decides for itself whether to search. If an answer shows no cited sources, you can't be sure it used the web. Record which condition each answer came from, or the grade means little.

Use three labels and keep them separate.

The split between wrong and unverifiable matters. FActScore, a factuality benchmark from Min et al., groups "contradicted" and "lacks evidence" into one not-supported class. For a brand check, that hides the useful signal. A wrong price is a fix target. An unverifiable headcount is a gap in your fact sheet or your public pages.

FActScore also gives every fact equal weight. You don't have to. A wrong claim about ownership or pricing hurts more than a wrong founding year, so flag severity yourself. If wrong claims involve old features, the guide on hallucination vs outdated answers helps you tell an invented fact from a stale one.

A single manual run tells you what ChatGPT said once, under one set of conditions. Several things move underneath it:

• Retrieval is discretionary. The same question can trigger a search one day and not the next.
• Routing changes the model. Different paths carry cutoffs months apart, per OpenAI's model documentation.
• Query rewriting is hidden. You see your question, not the searches ChatGPT actually ran.
• Models update. OpenAI warns its model snapshots are regularly updated.
• Small samples mislead. One wrong claim in five answers looks like a crisis. It may be noise.

Treat each run as one data point. Keep the same questions, date every run and compare runs over time. If you want a repeatable process, read monitoring what ChatGPT says about your brand. The free checker on this site gives a quick snapshot across up to three buyer questions, which is a fast first look but still a snapshot.

Sort each wrong claim by condition, because the fix depends on where it came from.

Wrong with search off

This is training data. Nothing you publish this quarter changes it. A later model trained on newer data may. Note it and move on to what you control.

Wrong with search on

This is fixable territory. Check in this order:

1. Look at the cited sources. If a third-party page carries the wrong fact, that page is the problem. Being described correctly on the sites ChatGPT cites is the route into the answer.
2. Check your own pages for contradictions. The same price stated differently on your pricing page and your footer gives ChatGPT two versions. Your About page should state key facts once, plainly.
3. Check that the fact is in the raw HTML. Vercel found ChatGPT's and Claude's crawlers fetch JavaScript files but don't execute them. A price that only appears after JavaScript runs may be invisible to them.
4. Check crawler access. OpenAI says OAI-SearchBot controls ChatGPT search, and your host or CDN must allow its IP addresses too. The AI Crawler Checker shows which crawlers your firewall turns away.

For retired features that keep appearing, see how to get AI to stop describing dropped features. Stating a specific fact on a fetchable page gives ChatGPT a sentence it can quote. It doesn't guarantee it will.