AI Is Getting Your Business Wrong: How to Fix It

A homeowner asks ChatGPT when your shop closes. It says 5pm. You close at 7pm, and you have for two years. Nobody calls to tell you the answer was wrong. You just quietly stop getting calls after five.

Wrong AI answers are not one problem with one fix. They have three separate causes, and the fix for one does nothing at all for the other two. So diagnose before you touch anything.

Why does AI get your business hours, services, or location wrong?

Because AI engines assemble your business facts from pages you do not control. Ahrefs analyzed 863,000 SERPs and 4 million AI Overview URLs and found that only 38% of pages cited in an AI Overview also rank in Google's top 10 for that query. In the same study, 31.2% of cited URLs ranked somewhere between positions 11 and 100, and 31.0% did not rank in the top 100 at all (Ahrefs, March 2026). Those are pages you have probably never looked at.

Google confirms that AI Overviews and AI Mode use a query fan-out technique, issuing multiple related searches across subtopics and data sources to build one answer (Google Search Central, 2025). One of those sub-searches can surface a 2019 directory listing with your old hours, and that listing gets a vote.

So the wrong answer is usually not the model inventing things. It is the model faithfully repeating something wrong that is still sitting on the open web with your name on it.

Is it stale training data, a bad source, or no source at all?

There are exactly three causes, and you can separate them in about ten minutes by asking the same question two ways: once with web search or browsing turned off, once with it on.

If the wrong answer survives with browsing off, it is baked into the weights and no support ticket will remove it. Your only move is to flood the retrieval layer with the correct fact so that every time the engine does search, it finds the truth and overrides its own memory.

If the wrong answer only appears with browsing on, you have a source problem, and source problems are fixable this month.

How do you find the source an AI engine is actually quoting?

Ask the engine to cite, then open every URL it names. Pew Research Center found that 88% of Google AI summaries cited three or more sources, and only 1% cited a single source (Pew, March 2025 data), so there is nearly always a paper trail to follow.

Run the wrong-answer query in a fresh, logged-out session in ChatGPT, Perplexity, and Google AI Mode. Log four columns: the engine, the exact wrong claim, every cited URL, and the date. Then open each URL and find the sentence the engine lifted.

Also check inside Google itself. In the Business Profile editor, Google shows a Collected info panel — available for select regions, languages and business categories — where, per Google's Business Profile documentation, you can review information Google gathered about you, see the source of that info and the date it was collected, and delete it. Most owners have never opened that panel. It is often where the bad hours came from.

How do you build a canonical facts page an AI engine will trust?

One URL, plain text, one fact per line, dated. Google's own AI-features documentation says to make sure that important content is available in textual form and that your structured data matches the visible text on the page (Google Search Central, 2025). A facts page is the cheapest way to satisfy both at once.

Put it at something like /about/company-facts and state, in short declarative sentences: legal name and any DBA, street address, service area, phone, hours by day, holiday policy, year founded, license numbers, payment and financing options, and whether you take emergency calls.

Then state the negatives too. If an engine keeps saying you do not do emergency work, the page needs a line that says you do, and a line that says what you genuinely do not do. Models guess when nothing on the open web says the thing plainly. Do not make them guess.

Mark the page up with Organization or LocalBusiness schema and use sameAs to tie your identity together. Schema.org defines sameAs as the URL of a reference Web page that unambiguously indicates the item's identity, such as a Wikipedia page, Wikidata entry, or official website. Point it at your Business Profile, your LinkedIn, your license record. Our guide to what schema markup actually matters for GEO covers the property-level detail.

Which platforms do you have to correct at the source?

Google Business Profile first, because Google states that it gathers info from different sources, like user reports and licensed content, and that if those sources report your info is incorrect or outdated, Google may update your profile for you. In the profile editor, blue text means Google changed that field, not you.

That is the whole game in one sentence: a third-party data source can overwrite your hours, and then an AI engine reads the overwritten version. Fix the profile, then fix what fed it.

For the knowledge panel, claim it first. Google's documentation says you claim it by searching for your entity, clicking Claim this knowledge panel, and signing in to an official property such as Search Console, YouTube, Facebook, or X. Then use the Feedback link, click the flag next to the wrong fact, describe what it should say, and provide links to web pages that verify it. That last step is the one people skip, and it is the one that decides the outcome.

Wikidata is worth it only if you actually qualify. Its notability policy says an item qualifies if it refers to an instance of a clearly identifiable conceptual or material entity that can be described using serious and publicly available references — one of three routes to the bar, and the only one most businesses could plausibly take. A five-truck plumbing company with no press coverage will get deleted, and a deleted entry helps nobody. Earn the references first. Our post on entity SEO and knowledge graph optimization walks through how a brand becomes a recognized entity in the first place.

Can you report a wrong answer directly to OpenAI, Google, or Perplexity?

Yes, three channels exist, and none of them promises a fix or a timeline. Use them anyway, because one of them lets you report the bad source, which is far more valuable than reporting one bad answer.

OpenAI's help documentation says you can report conversations and responses through its content reporting webform, or in-product by tapping thumbs down under a message and selecting an issue. The important line: reported domains and other content may be reviewed by OpenAI's Model Quality team, which may apply filters or other mitigations to help prevent ChatGPT from relying on unreliable sources in future responses. That is a direct channel for naming the zombie directory that keeps poisoning your answers.

Perplexity's help center is blunter: for incorrect responses, click the three-dot menu below the response and select Report. For Google, the documented correction paths for business facts are the Business Profile editor and the knowledge panel Feedback link, not the AI Overview itself.

One honest limit. OpenAI's privacy portal handles removal of personal data from ChatGPT responses, reviewed case by case, with a government-issued ID or alternative proof required. It is built for people, not for a company's wrong closing time, and OpenAI states that removing data from ChatGPT does not remove it from external websites or search engines. Do not plan around it.

How often should you re-test what AI says about your business?

Monthly, on a fixed list of 10 to 15 prompts, with one named owner and a dated log. Quarterly is too slow when one wrong closing time silently eats every after-5pm call for 90 days.

• What are {business name}'s hours?
• Does {business name} do emergency {service}?
• Is {business name} still in business?
• Who is the best {service} in {city}?
• Does {business name} offer {the service they keep denying}?
• How much does {business name} charge for {service}?

Run each one logged out, in a new conversation, with memory off, on ChatGPT, Perplexity, Gemini, and Google AI Mode. Record the answer verbatim and every cited URL. A one-tab spreadsheet beats any dashboard here, because the thing you need is the citation trail, not a score.

Expect retrieval fixes to show up within days to a few weeks, and expect training-data errors to persist until the next model. If the same wrong fact survives two clean re-tests after you fixed the source, you did not find the real source. Go back to the citation log.

If AI is quoting a source you do not control, the durable fix is a source you do: a canonical facts page, correct schema, and clean platform records that every engine keeps re-crawling. That is exactly what we build in our GEO program, month to month, no lock-in, no ranking guarantees. Want to see what the engines are saying about you today? Get my free audit.