Wrong descriptions come from stale sources, and the fastest way to locate them is to ask the engine to cite what it used, then check publication dates rather than arguing with the output.
Correcting the engine in the chat feels productive and changes nothing. The page it read is still there, still saying the same thing, and it will be read again tomorrow.
The instinct when an AI describes your product wrongly is to tell it so. That works for exactly one conversation, in one session, for one person. The engine has not learned anything and the next user gets the same answer, because the problem was never in the model. It was in a document.
This is the procedure for finding that document.
Why is arguing with the output pointless?
Because the answer was assembled from retrieved material, and the material has not changed.
Most answer engines fetch documents at query time and generate from them, the pattern set out in Lewis et al. A correction you type is a fact in the conversation, not a fact in the corpus. Close the tab and it is gone.
The engine will agree, apologise, and use the correct figure for the rest of that session. Nothing outside the session changes.
Slower, duller, and it changes what every future reader is told, because it changes what the retriever finds.
How do you find the source?
Ask the engine directly. It is more forthcoming about this than people expect, and it is faster than any amount of searching for the offending page yourself.
Temporary or incognito, memory off. Ask the question that produced the wrong description and confirm it happens without your own chat history involved.
If it does not reproduce, you were looking at your own conversation rather than at what a stranger sees.“Which source did you use for the pricing figure?” rather than “what are your sources?”. Naming the wrong fact gets you the document behind that fact instead of a generic list.
A general request returns everything it touched. A specific one returns the culprit.Published date, last-updated date, or the copyright line if it has neither. You are looking for when the page last said something true.
Almost every wrong description is a correct description that expired.If your pricing page states the figure in prose the engine can extract, the stale source is outranking you. If your figure only exists inside a table image or behind a toggle, it may not be extractable at all.
Sometimes the fix is on your side after all, and it is worth ruling out before you email anybody.What are the usual culprits?
Wrong descriptions cluster into a small number of source types, and they are worth knowing because they have different remedies.
Directory listings
Written once, never revisited, and frequently the oldest thing about you on the open web. High retrieval weight because they are structured and on-topic.
Usually editable
Review platform profiles
Often carry pricing and feature fields you filled in at launch. Engines trust them because they are third-party, which is exactly why a stale one is expensive.
Usually editable
Roundup articles
Somebody wrote “the ten best X tools” two years ago and never updated it. Dense, comparative, and precisely the shape a recommendation question retrieves.
Requires an ask
Your own old pages
An unretired announcement, a changelog entry, a legacy pricing page still indexed. The one category entirely within your control.
Fix today
Which engines will show you their sources?
Not all of them, and knowing which is which saves a lot of wasted asking.
The live-search family cites reliably because it just fetched the documents. Anthropic states that for its web search tool “citations are always enabled”, with each one carrying the URL, the title and up to 150 characters of the text actually used. That last field is what makes this trace possible at all: you get the passage, not just the page.
Perplexity is built around visible inline sources, which makes it the fastest place to run this check even if it is not where your buyers are. Google’s AI surfaces cite links drawn from the Search index. Engines answering from stored knowledge often cite nothing, because nothing was fetched.
Run the trace on whichever engine shows the most sources, even if it is not your primary. You are looking for the bad document, not for a visibility score.
You will spend twenty minutes asking a system that has no document to point at, and conclude wrongly that the source is untraceable.
What if it will not cite anything?
Then you have learned something separate and useful: the answer probably did not come from retrieval at all.
Engines answering from stored knowledge often have no document to point at, because none was fetched. That is a different problem with a different timescale. A stale page can be fixed this month; a stale impression in a model’s weights persists until a future model generation, and the only input you have is building consistent public evidence that reaches the next training cycle.
Test which one you are dealing with by asking the same question in a way that invites a lookup: naming a recent year, asking for current pricing, or asking for sources explicitly. If citations appear, you are in retrieval territory and the trace above applies.
Does structured data fix this?
It helps with identity and not with contradiction, and the distinction matters here.
Google describes structured data as helping search understand the content of a page, and makes no claim that it improves ranking position. It makes you easier to identify. It does not overrule a third-party page stating a different price.
What does help is stating the fact plainly, in text, on a page that is easy to retrieve. A price written in prose in a sentence is extractable. A price rendered inside an image, or only visible after clicking a monthly-annual toggle, may not be reachable at all, which leaves the stale third-party figure as the best thing available.
How do you know when it is fixed?
Not by asking once. Answers vary between identical runs, so a single clean answer proves nothing more than a single wrong one did.
Three times, in clean sessions, on the engine where you found the error. One correct answer is not evidence.
The page has to be recrawled before the change reaches an answer. OpenAI documents around 24 hours just for a robots.txt change to register.
Paste both full answers somewhere dated. It is the only record that the fix did anything, and screenshots of verdicts will not do.
If the wrong description turns out to be one symptom of a broader problem rather than a single stale page, the four-cause diagnostic covers the rest, and the five-minute test is the quickest way to see whether accuracy is your only issue or just the most visible one.
One last thing worth saying plainly. A confident, specific, wrong statement about your pricing is more damaging than being absent, because a reader has no reason to doubt it. Absence loses you a deal you never knew about. A wrong price loses you one you would have won.