Blog · Third-party sources

How to get a wrong fact about your brand corrected in AI answers

Published 3 April 2026 · Updated 8 October 2026 · 9 min read

How to get a wrong fact about your brand corrected in AI answers
Photograph: RoonZ nl / unsplash

The short answer: You cannot edit an AI answer. You can only change the documents it reads. When a wrong fact in an answer can be traced at all, what we find at the other end of the trace is usually a correct quotation of a stale page. So the job is three steps: find the exact URL the claim came from, get that URL changed or outranked by fresher evidence, then re-run the prompt over several days to see whether the correction travelled. Expect weeks, not an afternoon.

Marketers reach for a support ticket here, and there is no ticket to file. The useful instinct is closer to a librarian's than a publicist's.

First, find out whether anything is being retrieved at all

Two very different failures produce the same wrong sentence, and the fix for one does nothing for the other.

If the engine searched before it answered, the wrong fact sits in a document it fetched. On the surfaces that show their citations, such as Perplexity, Google AI Overviews and ChatGPT with search on, you can open what was used. Where an answer shows no sources, you cannot see what was fetched, and anything you say about the origin is inference. Open every cited URL and read it with the claim in hand. The offending line is often a line you wrote, on a page you forgot, or a line a directory copied from your site in 2023 and never refreshed.

If the engine did not search, the claim came out of the weights. No page is responsible, nothing you edit this week will move it, and the only honest lever is to make the correct version so common across the open web that the next training run sees it more often than the old one. Telling the two cases apart is the whole first hour of this work, and retrieval vs training data sets out the test we use: ask the same question twice, once with web access on and once with it off, and compare the two answers.

There is published evidence for why the retrieved document matters so much. ClashEval (Wu, Wu and Zou, arXiv preprint, 2024) benchmarked six models, including GPT-4o, on more than 1,200 questions across six domains, with deliberate perturbations applied to the supporting content. The models overrode their own correct prior knowledge and followed the incorrect retrieved content over 60 percent of the time. That is a benchmark with planted errors, not a measurement of any production answer engine, so read it as a reason to take a single bad page seriously rather than as a rate you should expect to see in the wild.

Rank the sources before you write a single email

List the URLs carrying the wrong claim and sort them by how much work each correction costs against how often it appears in answers. Four tiers, in the order we work them:

Pages you own. Fix today. Wrong prices and headcounts that we can trace frequently lead back to the brand's own estate: a landing page from a previous positioning, a press release with the old funding number, a help article nobody owns. Correcting these costs you nothing but the edit. When an engine next crawls the page is outside your control, and recrawl intervals are not published.

Structured records you control. Your schema markup, your Google Business Profile, your knowledge panel. Google's knowledge panel help describes claiming a panel by signing in to an official profile you already run, such as Search Console or YouTube, and says that feedback submitted while logged in with the verified address "receives priority for review". The same page notes that not all knowledge panels are claimable. Google does not publish how long a review takes, so treat this as a documented route rather than a dated one.

Open databases. Wikidata and Wikipedia carry a good deal of entity detail. Wikidata's Help:Sources says most statements should indicate where the data comes from, that references should point to reliable sources such as textbooks, academic journals and newspapers, and that Wikipedia pages themselves are not appropriate as sources. In practice that means a filing, a regulator page or a dated news article does more work there than a page you publish about yourself. If your own page on Wikipedia is wrong, you do not edit it. You post an edit request on the talk page using the conflict of interest template, and you wait: Wikipedia's edit request guidance says it can take anywhere from a day to several months, and that for conflict of interest requests you should treat six months with no response as a declined request.

Third parties. Directories, review sites, listicles, course notes, a consultancy's category roundup. These are the slowest, and often the ones that matter most, because they are exactly the documents engines like to cite. The third-party sources AI engines keep citing covers which ones show up most. Review platforms deserve their own treatment, since product facts there are user submitted and decay quietly: see review sites and what AI repeats about your product.

What a correction request should contain

Publishers do correct things. The working rule we have formed from sending these requests is that an email which reads like a correction gets answered and one that reads like a takedown gets ignored. We have not counted outcomes across publishers, so take that as a habit worth copying rather than a measured response rate.

Send the claim as published, quoted exactly. Send the corrected version, written as a replacement sentence they can paste. Send one verifiable source for the correction that is not your marketing site: a filing, a dated announcement, a product documentation page with a visible last-updated date. Name the date the original was published, because an editor's first question is whether the piece was right at the time. Ask for an update, not a deletion. A deleted page helps nobody, and an updated page keeps the authority it has already accumulated.

One thing worth saying plainly: do not threaten. Legal escalation on a factual error can produce a correction notice, or a news item about the threat, that is itself more quotable than the error was.

The index lag nobody plans for

A corrected page is not a corrected answer. Caches, embeddings and search indexes can keep serving the old version after you publish the fix. None of the engines publish how long their copies persist, so we cannot tell you how long the lag runs on a given surface, only that we see corrected pages and stale answers coexist.

Google's removals tool exists for the urgent case, and its documentation states that requests made in the tool last for about six months, after which the content returns unless the page itself changed. The separate refresh outdated content form is for a page whose description or cached snippet is stale after content was removed from the site. Google's help page is explicit that if you have already updated a live page, you do not have to fill out that form.

Three practical consequences. Change the page content rather than the URL, so the citation you already earn keeps working. If you must move it, make sure the redirect is a real 301 and the old URL stops returning the old body, which is the sort of request that goes nowhere without a developer, so getting engineering to ship your AEO fixes is part of this project, not a separate one. And update any dated lines on the page, because a visible revision date is one of the few cheap hints that a document has moved on.

Verify, and verify more than once

Re-run the original prompt. Then run it again the next day, and again a week later, on each engine where it matters. A single clean answer is weak evidence. In our audits, which run a prompt set of about 45 prompts across five engines with retrieval on and retrieval off, repeating the same prompt on the same engine often returns a different set of cited sources. That is a sample of a client's prompts on the days we ran them, not a census of how those engines behave.

Record three things each time: whether the claim is now correct, which URLs are cited, and whether a new source has taken the old one's place. The pattern you want is the corrected page appearing in the citation list. The pattern worth a second look is a clean answer with no citations at all, which is consistent with the engine not searching and answering from memory, where the error may still live. An engine can also retrieve and simply not show the source, so treat the missing citation list as a signal to re-test rather than as proof. The self-check has a version of this loop you can run without any tooling.

Honest limits

Some corrections do not land. A claim repeated across many pages keeps surfacing from the ones you never found, and we have no way to enumerate the pages we did not find. We can tell you which document a cited claim came from, on the surfaces that show citations, and whether a correction appeared in later runs. We cannot inspect what a model holds in its weights, and we cannot make it forget.

Two numbers in this post come with conditions. The 60 percent figure is from a benchmark with planted errors, not from production answer engines. The six month figure is Google's stated life for a removals request, not a measurement of how long any index holds a stale page. We have no reliable estimate of how many corrections stick permanently: our re-runs are monthly and cover the prompts in a client's audit set, and we have seen facts revert when a stale aggregator was re-crawled.

FAQ

Can I contact OpenAI or Google to fix a wrong fact about my company? The complaint channels these companies publish are built around personal data and unlawful content rather than commercial inaccuracies. For an out-of-date price or feature, changing the underlying documents is the route we have seen work.

How long does a correction take to show up? We cannot give you a schedule, because recrawl and index refresh intervals are not published. In the corrections we have tracked, pages the client owns move first, third-party publishers move at the speed of their editorial queue, and open-database edits with a conflict of interest declaration can run to months, which matches Wikipedia's own stated range of a day to several months.

The engine cites my own page but states the opposite of what it says. Check for a contradicting page elsewhere on your own site, then check whether the cited page renders its content client side. A page that arrives empty gets summarised from whatever else the engine holds about you.

Is it worth correcting a low-traffic directory listing? If it appears in the citation list for a question your buyers ask, yes. A page can be cited whatever its own traffic, because what put it in the answer was that it was retrieved for that query.

Sources

All six were opened and checked on 8 October 2026.

Related reading

The third-party sources AI engines keep citingReview sites, G2 and what AI repeats about your productGetting engineering to ship your AEO fixesRetrieval vs training data: which one is answering about your brandLinkedIn as a cited source in AI answersThe listicles you are missing from