--- title: "How do you fix wrong information about your brand in AI answers?" description: "Trace the error to the page it came from and fix that page. AI answers are built mostly from live web sources, often your own old pages." canonical: "https://underneath.agency/resources/fix-wrong-brand-information-in-ai-answers" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # How do you fix wrong information about your brand in AI answers? You fix it at the source: find the page the AI answer is repeating, then correct, retire or outweigh that page. Most wrong facts about a brand are not invented by the AI. They are old, conflicting or third-party facts that are still live on the web. ## The short version 1. In one agent study, only 7–10% of the finished answer came from the AI’s built-in memory; the rest came from pages it read ([Finder and colleagues](https://arxiv.org/abs/2609.34951), 2026, a vendor study). 2. Of 64 software prices that differed from the official pricing page, 39 were on another page of the vendor’s own site ([our pricing study](https://underneath.agency/research/ai-pricing-accuracy-study)). 3. Where a business’s Google profile number was missing from its website, 30.6% of phone numbers AI gave differed from the profile, against 1.6% where it was there ([our business facts study](https://underneath.agency/research/ai-business-facts-accuracy-study)). 4. In a simulated test, a single polluted page among the search results fooled AI assistants into recommending a fake product up to 27% of the time ([Luo and Chen](https://arxiv.org/abs/2606.13610), 2026). ## Can you actually fix what an AI assistant says about you? Often yes, because today’s AI answers are rebuilt from live web pages each time, not fixed in memory. [Finder and colleagues](https://arxiv.org/abs/2609.34951) ran 37,927 AI agent journeys about 1,056 real businesses. Only 7–10% of the finished answer came from the model’s training knowledge, whether or not the business’s site was readable. The rest came from what the agent fetched. The authors work for ora, which sells the readiness score their study uses, so treat the exact figures with care. That matters for anyone trying to correct a mistake. If an answer is assembled from pages, changing the pages changes the raw material. It also means a wrong fact usually has an address you can find. In [our pricing study](https://underneath.agency/research/ai-pricing-accuracy-study), only 4 of 64 differing prices could not be found on any source we could fetch. ## Why does source provenance matter in generative search? Because an AI answer inherits the strengths and errors of whatever pages it draws on, usually without telling the reader. Provenance means where a claim came from. [Xu, Iqbal and Montgomery](https://arxiv.org/abs/2605.14021) checked 98,020 claims in Google’s AI Overviews, the AI summaries at the top of Google’s results. They found 11.0% were not supported by the pages cited, and that the credibility of a source and the accuracy of the claim were largely unrelated. In 1.39% of claims, the cited sources disagreed with each other. When your own pages disagree, the AI can pick either version. Two live guides cover the wider problem: [how often AI answers say things their sources do not support](https://underneath.agency/resources/ai-answers-unsupported-claims) and [whether an AI can cite your page for something it does not say](https://underneath.agency/resources/ai-citing-pages-for-claims-they-do-not-make). Our pricing study shows the same inheritance. When a quoted price differed from the official page, we looked for that figure elsewhere. It appeared on the product’s own cited pages 90.0% of the time, against 20.0% on pages cited for other products. The AI was copying from its sources, not guessing. Provenance also explains the darker risk. In the simulated test by [Luo and Chen](https://arxiv.org/abs/2606.13610), one polluted page among ten real search results was enough to push a fake brand into recommendations up to 27% of the time. AI assistants resisted best in categories whose real brands they already knew well. A defense that ranked sources by credibility removed only about a sixth of the fakes. ## Which kind of error are you dealing with? Start by sorting the error, because each type has a different source and a different fix. | What the AI got wrong | Where it usually comes from | Where to fix it | |---|---|---| | An old price, plan or product name | Old pages, help articles or announcements on your own site | Update or retire every page that states it | | A phone number, address or hours | Your website and listings disagree | Publish one version everywhere | | A fact it leaves out | Your site is hard for AI agents to read | Put key facts in plain text on readable pages | | A complaint or reputation claim | Review and complaint platforms | Resolve the issue where customers report it | | A claim no source makes | Unclear; the AI may have misread a page | Document it and check again over several runs | The live guide [Do AI agents make up facts about my business, or just leave them out?](https://underneath.agency/resources/do-ai-agents-invent-or-omit-business-facts) covers the gap between missing and wrong facts. This guide is about the repair. ## How do you fix errors that start on your own pages? Make every page you control say the same current thing, and remove the old versions that compete with it. Conflicting first-party sources are the most common cause we have measured. In [our business facts study](https://underneath.agency/research/ai-business-facts-accuracy-study), we asked four AI engines for the contact details of 159 local businesses. For 24.2% of the businesses we could check, the Google profile number did not appear anywhere on their own website. Those businesses produced most of the differing numbers: 30.6% of the numbers AI gave for them differed from the profile, against 1.6% for the rest. Prices follow the same pattern. Of 64 differing software prices, 39 were on another page of the vendor’s own site, such as an old plan FAQ or a post about an earlier price change. Pages with a monthly/annual price toggle had 58.4% fully faithful prices, against 72.9% without one. A toggle hides half the price list from a single read of the page. Readability matters too. In the ora agent study, answers built from a business’s own site were 41% more accurate than answers about the same business built from the wider web. If an agent cannot load your page, it fills the gap from other people’s pages. See [where AI agents get their answer when they can’t read your site](https://underneath.agency/resources/when-ai-agents-cant-read-your-site). ## What about errors that come from other people’s sites? Fix them where they live, because AI engines lean heavily on review platforms and user-written pages. In [our reputation study](https://underneath.agency/research/is-it-legit-ai-reputation-study), four AI engines answered “Is this brand legit?” for real brands. 88.0% of answers cited a review or complaint platform. Claims attached only to review platforms were negative 56.5% of the time, against 6.4% for claims attached only to the brand’s own website. Trustpilot and the BBB accounted for 61.7% of all review-platform citations. The same study found a weakness worth knowing. 71.4% of claims that a problem was common rested on evidence that did not show how common it was. A handful of loud complaints can read as a pattern. Resolving recurring complaints, and answering them publicly, changes the evidence the engines cite. For directories, listings and news articles, the fix is ordinary and slow: ask the owner to correct the page. No study we reviewed tested whether a correction request to an AI company itself changes its answers. ## How do you know a fix has worked? Check repeatedly, on every major engine, because one answer proves little and engines differ widely. In our business facts study, 4.4% of Gemini’s answers had a fact that differed from the Google profile, against 37.1% for Perplexity. A fix that shows up on one engine may not show up on another. Run the same buyer questions on each engine you care about. Tone moves more than presence. In one vendor’s tracking data, whether the AI framed a brand positively or negatively flipped about 6.7 times more often than whether it mentioned the brand at all ([Kumar](https://arxiv.org/abs/2606.20065), 2026). Judge a correction over many runs, as explained in [why ChatGPT gives a different answer about your brand each time](https://underneath.agency/resources/why-ai-answers-about-your-brand-change). ## What should you do about it? Treat wrong AI answers as a source-cleanup job with a named owner and a regular check. 1. **Log each error with its engine, question and date.** Note any page the answer cites; it is your first lead. 2. **Search for the wrong fact itself.** Look on your own site first, including help centers, old plan pages and press posts. 3. **Retire or update old pages.** Redirect or rewrite pages that state outdated prices, plans, numbers or claims. 4. **Publish one version of each core fact.** Use the same phone number, hours, plan names and prices on your site, profiles and listings. 5. **Show prices and key facts in plain text.** State monthly and annual prices side by side instead of behind a toggle. 6. **Work the review platforms.** Fix recurring complaints at the root and reply where customers report them. 7. **Ask third parties to correct their pages.** Prioritize the pages the AI answers actually cite. 8. **Re-check on a schedule.** Ask the same questions on each engine over several runs before calling it fixed. To decide who that named owner should be, see [which teams should own AI search visibility](https://underneath.agency/resources/who-should-own-ai-search-visibility). If you want help with this kind of source cleanup, see our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization). ## What does the research not tell us yet? The research shows where wrong facts come from, but no study has yet timed how fast a fix shows up. - No study we reviewed measured how long it takes for a corrected page to change AI answers, on any engine. - No study tested correction or feedback channels offered by AI companies. - The agent study comes from a company that sells agent-readiness scoring and uses its own score. - Our studies cover local businesses and software prices, on single days, so other industries may differ. - The polluted-page test was simulated on frozen search results, mainly in Chinese, not on the live web. - Whether an AI that cites a bad source can be talked out of it by better sources elsewhere is untested. ## Frequently asked questions ### Can I ask ChatGPT or Google to correct wrong information about my company? The research we reviewed has not tested that route. What it does show is that answers draw on live pages, so correcting the pages the answer relies on is the evidence-backed path. ### Why does AI show an old price for my product? Usually because an old price is still on the web, often on your own site. In our pricing study, 39 of 64 differing prices were on another page of the vendor’s own site. ### Does fixing my Google Business Profile fix AI answers? It helps only if your other sources agree. Phone numbers differed from the profile 30.6% of the time when the website did not show the profile number, against 1.6% when it did. ### Can a competitor plant false information about my brand in AI answers? It is possible in principle. In a simulated test, a single planted page fooled AI assistants into recommending a fake product up to 27% of the time, and none of the tested defenses was adequate. ## Sources - Finder, Elovic, Shalev and Yosef (2026), [AX is the New AEO](https://arxiv.org/abs/2609.34951), arXiv:2609.34951. - Xu, Iqbal and Montgomery (2026), [Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact](https://arxiv.org/abs/2605.14021), arXiv:2605.14021. - Luo and Chen (2026), [One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders](https://arxiv.org/abs/2606.13610), arXiv:2606.13610. - Kumar (2026), [Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines](https://arxiv.org/abs/2606.20065), arXiv:2606.20065. - Underneath (2026), [How faithfully do AI assistants quote software prices?](https://underneath.agency/research/ai-pricing-accuracy-study) - Underneath (2026), [Do AI answers match a business’s Google profile?](https://underneath.agency/research/ai-business-facts-accuracy-study) - Underneath (2026), [“Is this brand legit?” How AI assistants build a reputation](https://underneath.agency/research/is-it-legit-ai-reputation-study) --- This is the Markdown twin of https://underneath.agency/resources/fix-wrong-brand-information-in-ai-answers. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.