A prospect asks ChatGPT whether your product supports SSO. It says no. It does.
Nobody told you, there is no notification, and the prospect has already moved on to the tool the answer recommended instead. This is the failure mode that upsets teams most when they first measure AI visibility, because unlike being omitted, it is actively working against you.
First: is it invented, or is it repeated?
This is the whole diagnosis, and skipping it wastes months.
Repeated claims come from a source. An article from two pricing changes ago, a review listing nobody updated, a competitor’s comparison table written to flatter them, your own documentation describing a deprecated limit. The engine is behaving correctly; the web is wrong.
Invented claims have no source. The model had a gap and produced a plausible-shaped detail to fill it. This is the true hallucination, and it is rarer than the reputation of the word suggests.
Tell them apart by asking the engine where it got that. On a retrieval-based engine you will usually get a citation, and half the time the answer is immediately obvious from the URL. If no source is offered and repeated asking produces none, treat it as invented.
The distinction decides everything after it. Repeated claims are fixable in days. Invented ones are a longer game against a model that will be retrained on whatever the web says by then — which makes the fix identical anyway, just slower.
Capture the case properly
Before doing anything about it, make it reproducible. A vague report that “ChatGPT says our pricing is wrong” cannot be acted on or verified as fixed later. Record:
- The exact prompt, verbatim.
- The engine, and the date.
- The claim, quoted rather than paraphrased.
- Whether a source was cited, and which.
- Whether it reproduces on a second, fresh session.
That last one matters more than it looks. Answers vary between runs, and a claim that appears once in five attempts is a different problem from one that appears every time. Both are worth fixing; only one is urgent.
The sequence that works
1. Correct the source
If a citation points at a stale review listing, update the listing. If it points at an old article, ask the publisher for a correction — most will, for a factual error, and it costs an email. If it points at your own site, you have found the cheapest fix available to you.
This is unglamorous and it is where nearly all the results are. Engines read the web; the web was wrong; fix the web.
2. State the fact unambiguously on your own site
Somewhere on your domain there should be a sentence that answers the question directly, in a form that can be lifted whole. Not “enterprise-grade security features” but “Noma supports SAML single sign-on on the Enterprise plan.”
Vague marketing language is a major cause of hallucination about product capabilities. A model that cannot find a clear statement either omits the fact or infers one, and its inference will not reliably favour you.
3. Get the correction into the sources engines read
Your own site is one input among many, and rarely the most trusted. If the review platforms and trade publications in your category still carry the old fact, the answer will keep reflecting them. Work out which sources actually get cited for your category — and they are often not the ones you would guess — and correct the record there.
4. Re-test on a schedule
Corrections propagate unevenly. A source updated today may be re-crawled this week or next month, and different engines will pick it up at different times. Put the prompt on a recurring check rather than declaring victory after one good answer.
5. Escalate the serious ones
For claims that are genuinely damaging — a safety allegation, a regulatory claim, an accusation — providers do offer channels for reporting harmful inaccuracies. Use them for those, and do not waste them on a wrong price tier. Bring your reproducible case; a vague complaint goes nowhere.
Which errors are worth the effort
You will find more wrong claims than you can chase. Rank them:
- Claims that lose deals. A missing capability, a wrong price, a compliance claim. These are the ones a buyer acts on immediately.
- Claims that reproduce every time. Consistency means a source is driving it, which also means it is findable and fixable.
- Claims on high-intent prompts. A wrong detail in an answer to “is X worth it” costs more than the same error in a definitional question.
Everything else can wait. Some of it will fix itself when the underlying source is corrected for a different reason.
The uncomfortable part
There is no support line. No form that guarantees a correction, no SLA, no confirmation that a fix landed. You are influencing a system indirectly, by changing what it reads, and then waiting to see whether the output moved.
That is genuinely worse than the tools we are used to. It is also why measuring matters more here than in channels where you can simply check the setting: without a record of what each engine said and when, you cannot tell a fix that worked from one that did nothing.