It is a real observed error and it is not yet evidence of a persistent pattern, and the distinction decides whether you should rewrite a page or simply watch.
Somebody checks ChatGPT, finds it describing their product incorrectly, and the room concludes there is a crisis. Sometimes there is. Often what has been found is one instance of something that does not reproduce, and treating it as a pattern means spending real effort on a problem that was not there.
What a single wrong answer could be
Several things produce one wrong answer and nothing else.
The phrasing of the question. A slightly different wording can retrieve a different set of sources, and a question written the way an insider asks it is not the question a buyer asks.
A retrieval miss. The engine did not reach the page that holds the correct information on that attempt, so it composed an answer from what it did reach.
Session conditions. Whether memory is on, whether web access was used, which model version served the response, and what else was in the conversation all change the output.
None of those is a fact about your brand. All of them look identical to a fact about your brand when you have one screenshot.
What a pattern looks like instead
A pattern is the same question returning the same wrong thing across more than one engine, or across repeated attempts under comparable conditions.
If ChatGPT, Claude, Gemini and Perplexity all describe your product as single account only when it supports twelve, that is not a retrieval accident. Something they share is wrong, which usually means a source they all reach is wrong, or the correct information is nowhere they can reach.
That is worth rewriting a page over. The single instance is worth noting and watching.
Decide the threshold before you look
The failure mode here is deciding what counts as a pattern after seeing the data, which is how a single bad screenshot becomes a quarter of work.
Set it in advance. A reasonable threshold for most questions is the same wrong answer appearing in at least two of four assistants on the same question, asked the same way, recorded with the date. One assistant, once, does not clear it.
Writing the threshold down first is what separates a measurement from a reaction.
Fixing the source does not prove you fixed the answer
This is the part that gets skipped most often, and it is the part most likely to produce a false sense of resolution.
Suppose the threshold is cleared, you rewrite the documentation, and thirty days later the answers are correct. It is tempting to conclude the rewrite worked.
It might have. Answers also move on their own, for reasons that have nothing to do with you. Model versions change, retrieval changes, and the set of sources an engine reaches on a given day changes.
To get closer to an answer, recheck comparable observations under the same conditions, and look at whether the revised page is actually among the sources the new answer draws on. If the revised page is not visible among the sources, you cannot establish whether it contributed. Even a citation would not, by itself, prove the change caused the improvement.
That is a slower conclusion than most reporting in this category offers. It is also the only honest one available from outside the system.
What to record so the comparison is possible later
A screenshot is not a record. To compare a later reading to an earlier one you need the question exactly as asked, the engine, the date, the conditions, and what the answer cited.
Without those, a recheck is a new observation rather than a comparison, and you cannot tell whether anything changed or whether you simply asked differently the second time.
This is the practical reason to record answer observations in a structured way rather than collecting screenshots in a folder. Not because the structure is impressive, but because the comparison is impossible without it.
What to hold constant between the first check and the second
Five things, and holding them steady costs nothing except writing them down the first time.
The question, word for word. Not the same question in spirit. The same characters. A rephrase retrieves differently and produces a difference you will misread as movement.
The engine and, where visible, the model version. ChatGPT, Claude, Gemini and Perplexity are separate observations and should never be averaged into one. Within a single engine, a version change between readings is a confound you want recorded rather than hidden.
Whether web access was used. An answer composed with live retrieval and an answer composed without it are different measurements of different things.
Whether memory or personalisation was on. If the account has history, the answer is partly about the account. A fresh session removes that.
The date, in every case. Not the month. The date, because the thing you are measuring changes faster than monthly.
A recheck that holds all five is a comparison. A recheck that holds none of them is a fresh screenshot with an older screenshot sitting next to it.
Frequently asked questions
ChatGPT said something wrong about my company. Should I be worried?
Treat it as a real observed error worth recording, not yet as a pattern. Ask the same question of Claude, Gemini and Perplexity before deciding whether to change anything.
How many wrong answers make a pattern?
Set the threshold before you look. The same wrong answer in at least two of four assistants, on the same question asked the same way, is a defensible starting point.
Why do AI answers about my product change between attempts?
Question phrasing, retrieval variation, session conditions and model version all affect the output. That variation exists independently of anything about your brand.
If I fix my page, will the AI answer change?
It may, and you will not be able to prove your change caused it. Recheck under the same conditions and look at whether the revised page appears among the sources.
Can I make ChatGPT or Gemini correct a specific claim?
Not directly. You can correct the sources they can reach and verify whether the corrected source appears in later answers, which is a slower and more honest process.
What should I record when I find a wrong answer?
The exact question, the engine, the date, the conditions of the session, and the sources the answer cited. Without those a later check cannot function as a comparison.
ABOUT AXIS SUITE
Axis Suite by TrendAxis is the independent intelligence layer that explains what AI believes about your brand, why it believes it, and what decision that belief ultimately drives. See the recorded observations and the method behind them at axissuite.ai and in the Proof Center.