When an AI assistant gets your business wrong, and how to find where the error came from

Sooner or later an assistant says something wrong about a business. It quotes last year's prices, the closing time from before the hours changed, a branch that shut in the spring, a service the business dropped, or the name of the previous owner. The person asking has no reason to doubt it.

The record behind an answer can usually be changed, but which record that is depends on the route the fact took into the answer. The background on how assistants build answers is in the AI search guide. This post is about finding the route, because a fix aimed at the wrong one does nothing.

Assistants decide prompt by prompt whether to search. Anthropic documents Claude searching when a request depends on information that is current or changing, including details about organisations that might have changed, and answering directly when the request draws on stable knowledge.¹ For a fact about a business, that leaves three routes.

Training data: stale or invented

Every model has a knowledge cutoff. Anthropic lists one per Claude model, from August 2023 to June 2026, and warns that a model may not know about anything after it.² A price you changed in early 2026 can still be the price a model learned.

The other failure is invention. Kalai and colleagues argue that facts which appear only once in training data, like an obscure person's birthday mentioned in a single obituary, are where base models hallucinate: if a fifth of birthday facts appear exactly once, expect a base model to hallucinate on at least a fifth of them.³ I think most facts about a small business are of that kind: the owner's name or the Sunday hours may appear on only a handful of pages.

A retrieved page: usually repeated, sometimes misread

When an answer cites a page, the wrong fact is usually on it. Seer Interactive ran 178 branded phone number prompts through seven AI search products and published the results in December 2025. Only 64% of the numbers matched the brand's own customer service pages, yet 93% appeared on the page the answer cited, and 59% of cited sources were third-party sites such as GetHuman.⁴ The assistants were mostly copying pages that were wrong.

Sometimes the cited page does not say it. In Liu, Zhang and Liang's 2023 audit of four generative search engines, only 74.5% of citations supported the sentence they were attached to.⁵ Google put some of its 2024 AI Overview errors down to misread queries and to satirical or forum content taken at face value.⁶ A wrong page can also beat a right memory. In the ClashEval benchmark of more than 1,200 questions, six leading models took incorrect retrieved content over their own correct knowledge more than 60% of the time.⁷

A structured listing: the record the assistant trusts

Several assistants draw on listings data for local questions. Google's AI Mode draws on the Knowledge Graph and information about the real world as well as web pages,⁸ and Google offers developers a Gemini tool that answers from Maps data on more than 250 million places, opening hours included.⁹ OpenAI says ChatGPT works with third-party providers that supply local information such as nearby business listings, and its help page does not name them.¹⁰

Listings also change without the owner touching them. Google may publish a user's suggested edit if the owner has not responded within four days and other public information supports it.¹¹ Apple edits a place card's name, status or hours after reviewing customer reports.¹² Bing builds each listing from Bing Places and other sources, and may prefer another source when the Bing Places data is more than six months old or conflicts with it.¹³ It is one more reason agreement between sources matters.

For one wrong fact this takes about an hour. The AI recommendations tutorial has a prompt log and the setup for testing cleanly, so I will not repeat them here.

Fix your own site and main listings first, then the pages the assistants cited, then everything else your quoted searches turned up. The master-record method is in the consistency section of the entities guide.

Your own pages

Old price lists, pages for a closed branch and pages for services you have dropped are easy to miss, because nobody thinks of them as live. Update or remove them and ask Google to recrawl, which takes a few days to a few weeks; asking again does not speed it up.²¹ Anthropic says removing content from your site is the best way to keep it out of Claude's web search answers, and owners can ask it to block a URL already appearing in Claude's outputs.²² For a branch that has closed or moved, see moving premises without losing map visibility.

Listings

Google reviews Business Profile changes before they go live, and lets anyone suggest an edit to a profile they do not manage.²³ Its help page now gives owners four days to reject a suggested edit once notified, a change Search Engine Roundtable reported in September 2026.²⁴ Bing publishes changes in cycles of 7 to 12 days in North America and up to two weeks elsewhere, and each edit to a pending listing resets the date. It only applies basic details such as hours when they agree with trusted external data, and for a business that has moved it asks for a new listing with the old one marked closed.¹³ On Apple Business, accept or contest Apple's edits from the location's alert. A name change can take up to five working days to review.¹²

Third-party pages and registers

Directories, review sites, news stories and comparison pages need a request to whoever runs them. Claim the profile where you can, and otherwise write to the editor with the correct fact and a link to the page on your own site that states it. If a page has been corrected but Google still shows the old version, anyone who does not own it can request a refresh with the Refresh Outdated Content tool, which only applies once the live page no longer carries the old text.²⁵

For a wrong owner, check the register first. A UK limited company must tell Companies House within 14 days about changes to its directors or to its people with significant control.²⁶ I would fix that before writing to anyone else, because it is the public record other people check. If the error is about the owner as a person, OpenAI's privacy policy for Europe and the UK lets them ask for inaccurate information about themselves in ChatGPT's output to be corrected or removed, a request OpenAI says it weighs against the law and what its models can technically do.²⁷

Every assistant has a way to flag a bad answer. None of the help pages I read promises to correct a specific answer or gives a timescale.

Use the buttons, and name the cited URL and the correct fact in the report. Do not wait on them, though. The follow-up the pages describe is general, and the source records are the part you control.

Training data is the slow route. A model keeps what it learned up to its cutoff until it is replaced,² and when a newer model ships, nobody outside the provider knows which pages its training data included. I think people overlook one cost of blocking training crawlers here: if you block them, your corrected pages are among the things future models cannot learn from. The robots.txt post covers that trade-off.

What you can do is make the current fact easy to retrieve. Claude is documented as searching when a question concerns details about an organisation that might have changed,¹ and retrieved content often wins over what a model remembers.⁷ That cuts both ways, which is why the stale copies matter as much as the correct one.

None of the help pages I read says how often an assistant's index refreshes, how long a corrected page takes to replace the old fact in answers, or when a product will switch to a newer model. My expectation, and it is only that, is that page and listing fixes show in search-based answers within weeks, and that an error from training data can last months or longer.

If an assistant gets your business wrong today, I would spend the first hour on the checks above and the second on the copies they turn up. Report the answer as well, then rerun the same prompts a month later and see which assistants have moved.