How long to fix wrong AI answers about you
How long to fix wrong AI answers about you depends on where the answer comes from. Stated response times, index refresh, a re-test schedule and a log.
How long it takes to fix a wrong AI answer about you depends on where the answer comes from, not on which assistant gave it. If the assistant read a web page, the answer usually changes after that page changes and the search index picks up the change, which Google says can take anywhere from a few days to a few weeks. If the answer comes from the model's own training, no provider page we read gives a time at all.
That second sentence is the uncomfortable one. None of the provider pages read for this guide states how quickly a report or privacy request about a wrong answer will be decided, so a number of days quoted for "fixing ChatGPT" is not in the providers' own documentation. What you can do is learn which clock you are waiting on, act on the parts with a known mechanism, and measure on a schedule.
This guide covers time only: what drives the wait, what each provider's page does and does not say about response times, how index refresh works according to Google's own pages, and a re-test schedule with a log. It does not explain why assistants invent facts, how to diagnose the error or what the law says. Provider pages change, and nothing here is legal advice.
- There are three clocks: the web page that must change, the search index that must catch up, and the model's own memory. Only the first two have a mechanism you can act on.
- For answers that cite a web page, the wait is the owner's response time plus the recrawl. Google's page says crawling can take a few days to a few weeks, and asking again does not speed it up.
- Provider pages for OpenAI, Anthropic, Perplexity and Gemini state no response time for reports or privacy requests. Where no time is given, plan on none.
- Re-test on a fixed schedule, with the same questions and several runs each, and log every round. One good answer after a retry proves nothing.
Sources: Google Search Central pages on recrawling and on removing information, and the help pages of OpenAI, Anthropic, Perplexity and Google for Gemini, all read for this guide.
Three clocks decide how long a wrong AI answer takes to fix
The wait is the sum of three separate clocks, and most frustration comes from treating them as one.
The first clock is the source page. If an assistant repeated something from a web page, a directory entry, an old article or a profile, that page is the origin of the error. It changes only when its owner acts, which is outside your control and any provider's.
Related readAI Answers About Job Candidates: Audit and FixThe second clock is the index. Assistants that search the web, and Google's AI features, read pages through a search index. A corrected page helps only after the index has seen the change. Google publishes statements about this clock, covered below.
The third clock is the model's own memory: what the system absorbed during training and now produces without looking anything up. Providers describe this clock in their policy pages only in outline. OpenAI's page on how its models are developed, for example, does not state a training data cutoff or describe how newer data is added.
| Where the wrong answer comes from | What must change | Clock you can read | What you control |
|---|---|---|---|
| A cited web page | The page, then the index | Owner's reply, then Google's stated recrawl range | Evidence, the request to the owner, the recrawl request |
| Model memory | The provider's system | None stated in the pages read | A report or privacy request; strengthening accurate sources |
Do the parts with a readable clock first. Treat the memory clock as something you monitor, not something you schedule.
Answers built from web pages move when the page and the index move
An answer that cites or paraphrases a web page is the most fixable kind, and its timeline is the page's timeline. ChatGPT's help page on its search feature says it will search the web automatically when a question might benefit from it, shows inline citations, and may share disassociated search queries with third-party search providers such as Bing. The page does not say how fresh results are or how often they are refreshed, so that part has no stated clock.
Google's AI features follow a more documented path. Google's page on AI features and your website says a page must be indexed and eligible to appear in Search with a snippet to be used in AI Overviews or AI Mode, with no additional technical requirements. It also says indexing and serving are not assured even when a page meets all requirements, and that these features may use a query fan-out technique, issuing multiple related searches across subtopics. An answer can therefore draw on several pages, so fixing one source does not always change it.
Related readAI answers about doctors and lawyers: what to fixThe order of operations that follows from this:
- Find every page that states the claimSearch for the exact wording and close paraphrases. Do not stop at the cited sources, because a fan-out answer may use pages it never shows.
- Ask each owner to change itSend a short request with the correct fact and one piece of evidence. The owner's reply time is the largest unknown.
- Confirm the change is liveLoad the page logged out and in a private window. A change saved in an editor is not a change a crawler can see.
- Request a recrawl for pages you ownFor a few URLs use Google's URL Inspection tool, and for many use a sitemap. Request once per URL.
- Strengthen an accurate page you controlGive assistants a clear, current page that states the correct fact beside your name.
- Start the re-test scheduleBegin rounds only after steps 2 to 4 are done. Testing earlier measures nothing.
A fast provider cannot rescue a slow owner: if the page stays unchanged, the original claim is likely to return the next time the assistant searches.
Answers from the model's own memory have no clock you can read
When an assistant answers without searching, it is producing text from what training left in its parameters. OpenAI's page on how its models are developed states no training data cutoff and does not describe how newer data is added. It says the model does not retain copies of what it processes, that individuals in certain jurisdictions can object to the processing of their personal information, and that requests can be declined where there is a lawful reason.
Anthropic's page on privacy rights requests relating to its training data says people can ask for erasure or correction under laws such as the GDPR and UK GDPR, including where their data appears in a Claude response. It notes that models do not store text like a database, and says requests may be declined for a lawful reason such as public interest. Neither page states how long a request takes; Anthropic's says only that requests are considered according to local legal requirements.
That is the finding. If the error lives in memory, you cannot schedule the fix. You can file a request, keep accurate sources in place, and keep testing. Why assistants produce such errors is a separate subject.
Seeing a corrected answer once does not mean the memory changed: answers vary between runs, and a search-enabled assistant may simply have fetched a better page that day.
Related readWrong company information in ChatGPT: how to fixWhat each provider's own pages say about response times
This is the table most people want and cannot find, because there is little to put in it. The OpenAI row rests on its help page on personal data removal. "None stated" means exactly that, not "fast" or "slow".
| Provider and route | What the page says | Stated response time | What to plan on |
|---|---|---|---|
| OpenAI, remove personal data from ChatGPT responses | Requests are assessed case by case; ID or alternative proof is required; reviewers rely on what you provide and cannot independently investigate disputed facts | None stated | No date to plan around; may be asked for more information |
| Anthropic, training data rights requests | Contact route for erasure, correction or objection; may decline for a lawful reason | None stated | No date; you may complain to a supervisory authority if unhappy |
| Perplexity, Report on a response | Three-dot menu, then Report | None stated, and the page does not say what happens after you report | Treat it as a signal, not a ticket |
| Gemini, feedback and Report legal issue | Feedback adds the conversation; a legal issue route exists under More below the response | None stated | No date; keep your own evidence |
The OpenAI page lists what reviewers weigh, including whether the person has a public or professional role, accuracy as judged from what you provide, the age of the person (information about children and young people generally attracts stronger privacy interests) and public interest, and says it may contact you for clarification. A complete first request, with identity proof and exact examples, is the only time saving the pages themselves support.
There is one number in this area, and it comes from a regulator. The UK Information Commissioner's Office says an organization must respond to a rectification request at the latest within one month of receipt, extendable by a further two months if the request is complex. None of the provider pages read says they apply it to a chatbot answer, so use it as a benchmark for chasing, not as a promise.
Search index refresh as Google's own pages describe it
Google's pages are the most explicit about timing. Read them as ranges Google states, not as promises.
On recrawling, Google's page on asking Google to recrawl your URLs says crawling can take anywhere from a few days to a few weeks. It says requesting a recrawl several times for the same URL will not get it crawled any faster, that there is a quota for submitting individual URLs, and that large numbers of URLs should be handled through a sitemap. A second request adds nothing. A sitemap with accurate dates is the right tool when you have changed many pages.
Related readGoogle AI Mode reputation: audit what it saysOn removal, the picture splits by who owns the page. Google's removal page says the Removals tool can take a page hosted on your own site out of results within a day, and that such requests last about six months. It is a temporary hide: the page itself must still change.
For a page you do not own, Google's Refresh Outdated Content tool applies when a page or image no longer exists or has deleted important content. Its page says it only updates Google's search result and does not remove a page from the web, that pending requests can take a few days to process, that for a changed page the snippet is removed and the result refreshed the next time the crawler visits, and that approvals expire after 180 days or sooner if the URL no longer exists. It is not for information still on a live page.
Notice what is missing: the time between Google's index updating and an AI answer changing. Google's page on AI features gives no timeline for inclusion, so that interval is unknown.
A realistic timeline for a source-driven correction, with an invented scenario
The scenario is invented, with generic parties. A consultant finds that an AI assistant describes her as head of a firm she left, citing a third-party directory entry. It is out of date, not defamatory. The table shows the order of events and what is known at each point; where no source gives a time, it says unknown.
| Stage | What happens | How long it takes | Who controls it |
|---|---|---|---|
| 1. Evidence | Save the answers, dates, questions and cited links | Same day | You |
| 2. Owner request | Ask the directory to update the entry | Unknown; depends on the owner | The directory |
| 3. Live change | Entry shows the correct role when loaded logged out | Unknown | The directory |
| 4. Index refresh | Search index sees the new version | Google states a few days to a few weeks for crawling | |
| 5. Assistant layer | Answer starts citing the corrected page | No source states this interval | The provider |
| 6. Stabilization | Corrected answer appears in most runs, in every assistant tested | Measured by your own log | You, by testing |
In our experience the largest delay in a case like this is usually stage 2, not the technology, because nobody follows up with the owner. That is a pattern from practice, not a statistic. If the owner does not reply, the options are a formal request to the platform behind the page or a stronger accurate page, and neither has a promised time. If the claim is also in the assistant's memory, steps 1 to 4 fix the search-enabled answers and leave the memory-based ones unchanged, so test for each separately.
Related readWhat Siri, Alexa and Google Assistant say about youA re-test schedule that shows whether the fix worked
The rounds below are a testing plan, not a prediction that anything will change by a given date. Start only after the page-side steps are done.
- Baseline, before any changeRun each fixed question several times per assistant and log the results. Without a baseline, later rounds cannot be compared.
- Day of the live changeConfirm the page is changed and record the time.
- One week laterA light round. Expect little; it catches a fast refresh and any regression.
- Two to four weeks laterFull-size round, inside the window Google's recrawl range points to, so the first that can show a real difference.
- Monthly after thatRepeat until the corrected answer appears in most runs across two rounds in a row, then slow down.
- After any eventA new article or an assistant update can reverse a fix. Run an extra round and log the trigger.
The rule for calling it fixed is simple: the corrected answer must appear in most runs, in every assistant you track, in two consecutive rounds. One good answer is a lucky run. A fix that appears in one assistant and not another is a partial result, and the log should say so.
For the ongoing part, a monitoring service such as our reputation monitoring exists, but doing it yourself works if you keep identical wording and a written log. Rewording a question starts a new baseline.
A log template you can copy
Keep one row per question, per assistant, per round. These columns are enough.
| Column | What to write | Why it matters |
|---|---|---|
| Date and round | Baseline, week 1, week 3 and so on | Shows the timeline against your actions |
| Assistant and mode | Name, whether web search was on, logged in or out | Search-enabled and memory-only answers behave differently |
| Exact question | Word for word | Rewording breaks comparison |
| Runs and result | For example, wrong in 4 of 6 runs | Answers vary, so one run proves nothing |
| Action since last round | Request sent, page changed, report filed, with the date and any reference | Links cause to effect, and shows who has not replied |
Record failures as carefully as successes, because a fix that reverses is the case you most need to see, and keep dated screenshots of any harmful answer for a later provider request or a lawyer.
What stalls a correction, and what to do when the clock stops
Corrections stall in four predictable ways.
The owner does not answer. Move to the platform or host behind the page using its own reporting route, or build accurate pages that outrank the wrong one. In our experience, pressure on a page owner can create a second story.
The page is fixed but the answer is not. Check that the page is live logged out, that the index has had the time Google describes, and that the claim does not appear on another page, since fan-out answers can draw on several.
Related readWhat Microsoft Copilot says about you: check and correctThe error is in memory only. There is no page to fix and no stated clock: file the provider's request, keep an accurate source in place and keep testing.
The information is accurate. If it is true, lawful and of public interest, no route removes it and we would not help hide it. The work is adding accurate, current context.
Your next step is the baseline, not the form
Before filing anything, run the baseline round and write down every page the answers cite. It is the only thing that makes the later weeks measurable. Reputation Partners can run the first audit and set up the routine if you prefer, and we will not give you a fix date that no source supports.
Save the wrong answers with dates and cited links today, then identify the page behind each one before you file a report. If you want a second pair of eyes on which clock applies to your case, a private review is the place to start. See what AI assistants say about you.
Questions people ask
Why does a corrected page not change the AI answer straight away?
Because the page is only one step. The owner must change it, the search index must pick up the change, and the assistant must then read it. Google says crawling alone can take a few days to a few weeks, and no page we read states how long the assistant layer adds. Memory-based answers may not change at all.
Does asking Google to recrawl more than once help?
No. Google's page says requesting a recrawl multiple times for the same URL will not get it crawled any faster, and individual URL submissions have a quota. Use URL Inspection for a few pages and a sitemap for many, request once after the page is live, then wait through the stated range.
Does ChatGPT say how long a removal request takes?
The OpenAI help page on personal data removal from ChatGPT says requests are assessed case by case, asks for examples, links and identity proof, and says OpenAI may ask for clarification. It does not state a response time. Plan without a date, and send a complete first request so you avoid a round of follow-up questions.
Is there a legal deadline for an AI company to answer?
Regulators set deadlines for the organizations they cover. The UK Information Commissioner's Office says a rectification request must be answered at the latest within one month, extendable by two months if complex. Whether that applies to a given provider depends on law and facts, so ask a lawyer.
How many times should I test before I call it fixed?
Run each fixed question several times per assistant, since answers vary. Call it fixed when the corrected answer appears in most runs, in every assistant you track, across two rounds in a row. One good answer is not evidence. Log failures too.
What if the wrong answer comes from a page I do not own?
Ask the owner to change it, with the correct fact and evidence. If the page is gone or substantially changed, Google's Refresh Outdated Content tool can update the search result, though approvals expire after 180 days. If the owner refuses and the information is accurate, the route is adding better context.
