Short answer, as of August 2026: you cannot edit what an assistant says about you, and there is no complaints queue that will do it for you. The only reliable correction is to change what the model can read — fix the fact at its source, state it unambiguously and prominently on your own site, and get the third-party records that disagree updated so the corroboration flips. Expect the visible answer to change over weeks, not hours, and expect the slowest engine to be the one still repeating the old fact a month after everyone else has moved on.
This is a supply problem dressed as a reputation problem, and treating it as the latter wastes the first fortnight.
Why assistants repeat facts that are years out of date
Three mechanisms produce a confidently wrong answer, and they need different responses.
Training-time memory. Some of what an assistant "knows" about your business is baked into the model from a snapshot of the web taken at some point in the past. Nothing you publish today changes that layer directly. It changes only when the model is retrained or when a live retrieval step overrides it.
Retrieved-but-stale sources. More often the assistant fetched something current at read time — but the thing it fetched is itself out of date. A directory listing nobody has touched since 2023, an aggregator that scraped your old pricing page, a syndicated press release, a review site profile with a former address. The retrieval worked perfectly; the source was wrong.
Conflict resolution gone the wrong way. When sources disagree, the engine has to choose. It tends to favour whichever version appears in more places and in more independent-looking places. If your correct fact exists on one page — yours — and the incorrect fact exists across nine directories, weight of evidence is against you.
The third mechanism is the one people find hardest to accept, because it means being right on your own website is not automatically decisive. The underlying model of how disagreement gets resolved is set out in how ChatGPT, Perplexity and Gemini pick sources.
First, establish what is actually being said
Before fixing anything, capture the error precisely. Vague reports — "AI says the wrong thing about us" — cannot be worked on.
Ask the question the way a customer would, on each engine you care about, and record the verbatim answer with the date. Do this for the specific fact in dispute: your price, your hours, your service area, whether you still offer a discontinued line. Then ask the follow-up that matters most: where did you get that? Most assistants will name or link sources when asked, and that list is your work order.
Repeat the check across at least a few sessions. A wrong answer that appears once is noise; a wrong answer that appears consistently across sessions is a fact the system holds. This is also the point at which having a standing prompt set pays off, because you notice the drift on a chart rather than when a customer mentions it — see choosing which prompts to track.
The correction order of operations
| Step | What you change | Typical time to visible effect | Notes |
|---|---|---|---|
| 1. Your own page | The canonical statement of the fact, dated | Days to 2 weeks | Must be in server-rendered HTML, not behind a tab |
| 2. Structured data | Schema.org fields carrying the same value | 1–3 weeks | Machine-readable agreement with the visible text |
| 3. Google Business Profile | Hours, address, phone, categories, services | Days to 2 weeks | Highest-trust local record for most engines |
| 4. Major directories | The same value everywhere it appears | 2–6 weeks | Refresh cycles are the vendor's, not yours |
| 5. Aggregators and data brokers | Upstream feeds that repopulate directories | 4–12 weeks | Skipping this makes step 4 undo itself |
| 6. Press, profiles, syndicated copy | Old releases, partner pages, author bios | Variable | Often the source a model quotes verbatim |
| 7. Re-verify | Same prompts, same engines, dated | Ongoing | Change is gradual and uneven across engines |
Step 5 is the one most correction efforts skip, and skipping it is why the same wrong phone number reappears three months after being fixed. Directories repopulate from upstream feeds; correcting the visible listing without correcting the feed is deleting an email without unsubscribing.
The mechanics of getting these records to agree — and why partial agreement is worse than none — are covered in entity consistency and the knowledge graph, and the specifics of keeping the Google record current are in Google Business Profile automation.
Make the correct fact easy to prefer
Fixing sources is necessary. Making your version the easiest one to use is what accelerates it.
Be unambiguous and current. "Our pricing starts at $X per month, as of August 2026" is a usable sentence. "Contact us for pricing" is not a fact at all, and when your page declines to state something, the engine will happily use whatever a third party said instead — usually an old number from a comparison article. This is the practical case for publishing your prices rather than gating them.
Say it in prose, not only in a widget. A price rendered by a client-side pricing component may not exist for a crawler that does not run JavaScript. If your correction lives only in an interactive element, it may never be read at all — the failure mode described in JavaScript rendering and AI search visibility.
Remove the old fact rather than burying it. Superseded pricing pages, old service pages, archived location pages and previous-brand press releases are all still readable. If a page must stay for legal or historical reasons, date it explicitly and state what replaced it. An undated old page and an undated new page look equally authoritative to a machine.
Do not contradict yourself across your own estate. The most common self-inflicted version of this problem is a footer, a PDF and a landing page that each carry a different phone number. Fix your own house before complaining about the neighbours.
What about asking the platform to fix it?
Some engines offer feedback mechanisms, and where they exist they are worth using — a thumbs-down with a correction, a "report an issue" link, or the publisher controls that a few platforms provide. Use them, but do not build the plan on them. They are one-off, unacknowledged, and there is no visibility into whether the report was actioned.
Two cases deserve escalation beyond content work: a genuinely defamatory statement, and a factual error that causes real harm — a wrong address sending customers to a stranger's home, an incorrect safety claim, a phone number belonging to someone else. Those warrant a direct report through the platform's formal channel and, if the wrong data originates from an identifiable third-party source, a correction request to that source directly. Everything short of that is best treated as the supply problem it is.
Set the expectation before you start
The honest timeline runs roughly: your own site within days once re-crawled; your Google record within days to a fortnight; directories over several weeks depending on their cycles; aggregators over one to three months; and training-time memory only whenever the next model refresh happens, which nobody outside the labs can schedule.
That means partial improvement is the normal shape of success. One engine will update quickly, another will keep repeating the old fact for weeks, and the temptation will be to conclude nothing is working. Judge it on the trend across engines, not on a single check — the same discipline that makes any AI citation timeline legible.
DigiRank's AI Visibility Tracker records the verbatim answers and cited sources for your prompt set across six engines on a schedule, so a stale fact surfaces as a dated record you can act on, and the citations module audits your listings across twenty-one directories so you can see which records still carry the old value. Tracking and the citations audit start on the $99/mo Starter plan; the Citations Builder that pushes corrections into listings is included from Agency at $249/mo with a 14-day trial. If the wrong fact is being repeated because a competitor's page answers the question better than yours does, what to do when a competitor is cited and you are not is the companion piece.
Frequently asked questions
AI is saying wrong information about my business — how do I get it corrected? You cannot edit the answer directly. Correct the fact on your own site in server-rendered text with a date, mirror it in structured data, update your Google Business Profile, then push the same value through the directories and the upstream aggregators that feed them. The answer changes when the weight of readable evidence changes, typically over two to six weeks.
Why does ChatGPT still have my old pricing? Either it is recalling a training-time snapshot, or it retrieved a current page that is itself out of date — an old comparison article, a stale directory profile, or your own superseded pricing page. Ask it where the figure came from; most assistants will tell you, and that source list is your work order.
Can I report a false AI answer to the platform? Some engines offer feedback or publisher controls, and it is worth using them, but they are unacknowledged and unreliable as a plan. Reserve formal escalation for genuinely harmful errors, such as a wrong address or a phone number belonging to someone else, and treat everything else as a content and data-supply fix.
How long does it take for a correction to show up? Your own site within days of being re-crawled, your Google record within days to two weeks, directories over several weeks, and upstream aggregators over one to three months. Model-memory errors persist until the next training refresh, which is outside anyone's control.
Why did the old information come back after I fixed it? Almost always because the directory was corrected but the upstream data feed that populates it was not. Directories repopulate from aggregators, so a fix applied only at the visible layer gets overwritten on the next sync.
Does deleting the old page help? Usually yes, but redirect it rather than returning a 404 so any accumulated authority carries over. If the page must remain for legal or historical reasons, date it clearly and state what superseded it, because an undated old page looks exactly as current as an undated new one.
Should I publish a correction page? Only if the error is widespread and consequential. A dated, factual statement of the current position on the relevant service or pricing page is more effective than a standalone rebuttal, which mostly adds another page for the engine to weigh.
How do I know the correction worked? Re-run the same prompts on the same engines on a schedule and record the answers with dates. Improvement is uneven across engines, so judge it on the trend over several weeks rather than on any single check.