Classify the error before deciding what to do
The same visible symptom — a wrong claim about your product — has several distinct causes, and the response differs in every case. A claim that came from a stale version of your own pricing page is a content freshness problem. A claim that matches a competitor's marketing language may have been assembled from a competitor's material because nothing on your site answered the question. A claim that appears nowhere in any indexed source and is simply wrong is a model or sampling artefact, and no amount of editing will reach it.
- Stale first-party content: the fact was true once and your own page still says it. Fix the page and note the change date. The site scan is where you find the page that still carries the old wording.
- Third-party origin: the wording traces to another site. The work is to publish the correct fact on your own domain in a form that can be quoted. The citation report shows which pages were cited for the answer, which is where a third-party origin becomes visible.
- Genuine model error: the claim appears in no source you can find, and it does not reproduce consistently. Record it and do not spend editorial time on it.
- Prompt-sensitivity: the claim appears only when the question is phrased in a particular way. That tells you what the question is testing, not that your content is wrong.
Choose the correction path that matches the cause
Publishing a correction to an AI platform is not a lever any vendor holds. What is available is the ability to change the material a model is most likely to retrieve. That makes correction a content and evidence problem rather than a request to a third party.
- For stale content: update the page, and keep a dated change record so the history of the fact is visible
- For missing coverage: publish a page that answers the specific question directly, with a conclusion in the first paragraph and checkable facts under it
- For third-party origins: state the fact on your own domain in language a model can lift, and make sure the same wording appears on every page that touches the attribute
- For model errors: log the case, keep the response text, and re-test on a fixed schedule rather than reacting to the single instance
Why waiting for the platform to update usually does not work
Waiting is tempting because it costs nothing and the error is visibly wrong, so it feels like the platform should eventually get it right. In practice the platform has no queue for individual corrections. Models are retrained or re-prompted on their own schedule, and the inputs they draw on change for reasons you do not control. The most common outcome of waiting is that the error persists, because the cause was a gap in the material available rather than a mistake in the model's knowledge.
There is one case where waiting is right: when the cause is genuinely the model's own error and no source states the claim. In that case editing your own site has no effect, and the useful action is to record the case and to watch whether it recurs.
- Waiting works when the claim has no traceable source, and you have recorded it for later review
- Waiting fails when a page of yours is the source, because that page is still there and will keep being retrieved
- Waiting fails when the question is simply unanswered on your site, because nothing will change until something answers it
Verify the correction rather than assuming it
A published correction changes what is available, not what any particular model returns on any particular day. Treating the publication date as the finish line means the case is closed without evidence, and a still-wrong answer reappears later with no record that anyone checked.
- Re-run the exact question set on a fixed schedule, and record the answer text each time, not just a presence flag. The query tool stores each response verbatim, which is what makes the re-run comparable rather than anecdotal.
- Check the cited URL as well as the claim; a correct statement attributed to an unrelated page is still a problem
- Keep the exception open until the same question no longer produces the claim, or until the cause is confirmed to be model-side
- Note which pages you changed and when, so a later reappearance can be distinguished from a fresh problem
An incorrect answer is worth less as a complaint and more as a finding: it points at a specific gap in the material available. Treating it that way turns a recurring annoyance into a repeatable diagnostic. Sources reviewed on 2026-10-20. This article records a method with verifiable boundaries; it does not claim that any search or AI platform will index, cite or recommend anything.