A prospect on a sales call mentions, almost in passing, that they nearly didn't book the demo. They had read that your product doesn't integrate with the accounting software they use. It does, and has for over a year. When the sales rep asks where they read it, the answer is simple: "I asked ChatGPT." Somewhere, an AI tool is still describing a version of your business that stopped existing twelve months ago, and nobody on your team knew.
This is one of the quieter costs of AI search. Chatbots state outdated facts with complete confidence, and most brands only find out by accident. The good news is that these errors are rarely random. Every wrong answer has an address: a page the tool looked up, or patterns it learned from sources during training. Find the address, fix what lives there, and the answer can change. Complaining to the chatbot, on its own, usually doesn't.
This post covers correcting errors. Building a clear, consistent brand identity in the first place is covered in brand entity optimization, which is the best long-term prevention for everything below.
Why chatbots repeat old facts
AI tools get brand facts wrong for a small number of recurring reasons. The biggest is training memory. Every model learns from data up to a cutoff, so its background knowledge of your business is a snapshot. If you changed pricing, leadership, products or locations after that point, the model may still describe the old version when it answers without searching the web.
The second is retrieval. When an AI tool does search the web, it can find and trust an outdated page: an old pricing page you never took down, a directory listing nobody updated, a news article from three years ago or a competitor comparison written before your latest release. Sources AI tools cite are often surprisingly old, so stale pages keep influencing answers long after they stop being true. Other causes include confusion with a similarly named company and inconsistent information across your own profiles, which leaves the model to guess.
Diagnose the layer before you fix it
The single most useful step is working out where a specific error lives, because the two main layers need different fixes and very different amounts of patience.
| Error type | Where it usually lives | Main fix | Typical timescale |
|---|---|---|---|
| Old fact, sources cited | A page the tool retrieved | Correct or retire that page | Days to weeks |
| Old fact, no sources | Training memory | Make the new fact consistent everywhere | Months |
| Mixed up with another company | Weak identity signals | Strengthen and connect your profiles | Weeks to months |
| Inconsistent details | Your own conflicting profiles | One source of truth, applied everywhere | Weeks to months |
A practical rule of thumb: if the wrong answer cites sources, follow the links, because the outdated page is probably right there. If it cites nothing, the error is likely coming from the model's memory. Test the same question with web search switched on and off where the tool allows it; if the answer changes, you have learned which layer you are dealing with.
Make
Powerful no-code automation platform connecting all your marketing tools together
Best for: No-Code Automation
Step one: audit what AI says
Most brands have never systematically asked AI tools about themselves. Start with a short list of factual questions, the kind prospects and customers actually ask: what does the company do, how much does it cost, who runs it, where is it based, does it integrate with a named tool, what is its refund policy, is it still offering a particular product? Add a few comparison questions naming your main competitors, since those answers often contain the most damaging errors.
Run the list across the AI tools your customers use, record each answer and every cited source, and log the errors. Because answers vary between runs, repeat anything surprising a few times before concluding it is a consistent error. The same discipline behind AI share of voice measurement applies: a fixed question set, repeated regularly, beats one-off screenshots. Real customer questions from sales and support, gathered using question research, make the audit far more representative.
The principle Every wrong AI answer has an address: a page the tool retrieved, or patterns it learned in training. Find the address and fix what lives there. Complaining to the chatbot, on its own, rarely changes anything.
Step two: decide what to fix first
An audit often turns up more errors than a team can tackle at once. Prioritise by how much harm each one does.
Score each error from 1 to 3 on two factors, then multiply:
→ Severity: how much damage if a buyer believes it (1 = cosmetic, 3 = loses a sale or creates a legal or safety risk)
→ Reach: how often the question is asked and how many tools repeat it (1 = rare and isolated, 3 = common and widespread)
Examples:
→ "Doesn't integrate with [major tool]" in comparison answers: severity 3 × reach 3 = 9. Fix now.
→ An old pricing figure quoted in pricing questions: severity 3 × reach 2 = 6. Fix this month.
→ A founder's outdated job title: severity 1 × reach 1 = 1. Fix when convenient.
Anything scoring 6 or more goes to the top of the list. Errors that cost sales or create risk outrank errors that are merely embarrassing.
Step three: fix the sources
Start with your own site. Make key pages, such as about, pricing, product, integrations and policies, state current facts explicitly and clearly. Where something changed, say so directly, for example "as of 2026, we integrate with…", because explicit statements are easier for AI tools to pick up than implied ones. Retire or redirect outdated pages instead of leaving them live, and check that crawlers can reach the corrected ones, using the checks in technical SEO fundamentals. Structure facts clearly so they are easy to extract, as covered in answer engine optimization.
Then fix the wider web. Update your structured data, Wikidata entry, business profiles and the major directories in your category so they all agree. For local and multi-site businesses, consistent names, addresses and opening hours matter enormously, the same discipline behind the local map pack and multi-location SEO. Where a third-party page carries the wrong fact, contact the publisher with a short, polite correction and evidence; many will update it.
Make the new fact outnumber the old one. Training memory shifts slowly because an old fact may be repeated across many sources. Publishing the correct information clearly, in more places, over time, is what eventually tips the balance. Genuine independent coverage of the change helps most, since AI tools weigh what others say about you heavily, as covered in earning third-party mentions.
Step four: use feedback, and set expectations
Most AI tools let users flag inaccurate answers, such as the thumbs-down option in ChatGPT, the feedback link under Google's AI Overviews and Perplexity's report option. Use them, and include the correct fact and a source link where the form allows. But treat feedback as a supplement, not a fix. It rarely produces a quick structural change, particularly for errors rooted in training memory.
Set expectations honestly with colleagues. Errors from live retrieval often improve within days to a few weeks once the source page is corrected and re-crawled. Errors from training memory can persist for months, until the model is updated. Recheck the same questions at around 30, 60 and 90 days, and work on both layers at once rather than waiting on the slow one. Meanwhile, make sure visitors who arrive with a wrong impression find the right answer quickly, as covered in building landing pages for visitors sent by AI.
And your own chatbot
The same accuracy problem applies closer to home. In February 2024, British Columbia's Civil Resolution Tribunal held Air Canada liable after its website chatbot gave a customer incorrect information about bereavement fares, rejecting the idea that the chatbot was responsible for its own statements. Outcomes vary by jurisdiction, but the lesson is clear: treat what your own AI tools say as your own statements.
Ground your chatbot in an approved, regularly updated knowledge base rather than letting it improvise. Limit it to topics it can answer reliably, make it hand over to a person for prices, policies, refunds and anything uncertain, and review conversation logs for wrong answers. The design principles in designing a chatbot flow that doesn't frustrate users apply here too. This is general information rather than legal advice.
Prevent the next one
Most outdated AI answers start with an internal change that never reached every source. Build a simple change checklist: whenever pricing, leadership, products, locations or policies change, update your own pages, retire the old ones, refresh structured data and profiles, and notify key partners and directories. Then re-audit quarterly with the same question set, so the next stale fact is caught by your team rather than by a prospect on a sales call. For the wider picture of how these disciplines fit together, see our plain-English guide to AI search optimization.
The short version
AI tools often repeat outdated facts about brands with complete confidence, and most businesses only find out by accident. The errors are rarely random: every wrong answer has an address, either a page the tool retrieved or patterns it learned during training. Diagnose the layer first. If the answer cites sources, the outdated page is probably among them; if it cites nothing, the error likely comes from training memory. Audit what the main AI tools say about your pricing, products, leadership, integrations and policies, and prioritise errors by severity and reach. Fix your own pages first with explicit, current statements, retire or redirect old pages, then align structured data, Wikidata, business profiles and directories, and ask third-party publishers to correct errors. Make the new fact outnumber the old one through consistent, independent mentions. Use platform feedback as a supplement, expect retrieval fixes in days to weeks and memory fixes in months, and recheck at 30, 60 and 90 days. Hold your own chatbot to the same standard, and prevent future errors with a change checklist and a quarterly audit.
Is AI still describing an old version of your business?
We help brands audit and correct what AI tools say about them.
Explore SEO →Frequently asked questions
Why does ChatGPT say outdated things about my company?
Usually for one of two reasons. The first is training memory: every AI model learns from data up to a cutoff date, so if you changed your pricing, leadership, products or locations after that point, the model's background knowledge of you is a snapshot of the old business, and it may repeat that snapshot when it answers without searching the web. The second is retrieval: when an AI tool does search the web, it may find and trust an outdated page, such as an old pricing page still live on your site, a directory listing nobody updated, a years-old news article or a comparison page written before your latest changes. Other common causes include confusion with a similarly named business and inconsistent information across your own profiles. Working out which of these is behind a specific wrong answer is the first step, because each needs a different fix and a different amount of patience.
How do you correct wrong information about your brand in AI answers?
Fix the sources, not just the answer. Start by auditing what the main AI tools say in response to factual questions about your business, and note any sources they cite. Where an answer cites a page, update or correct that page, whether it is on your own site or a third party's. Make your own key pages state current facts explicitly and clearly, with dates where helpful, and retire or redirect outdated pages rather than leaving them live. Update your structured data, Wikidata entry, business listings and major directory profiles so they all agree. Publish clear, crawlable content confirming the change where it matters, so the correct fact appears across more sources than the old one. Use each platform's feedback option on inaccurate answers as a supplement, but do not rely on it alone, since feedback rarely produces a quick structural change. Then re-check the same questions after 30, 60 and 90 days to see whether the fix has taken hold.
How long does it take for AI tools to update information about a brand?
It depends on where the error lives. Answers built from live web retrieval can change within days to a few weeks once the source page is corrected and re-crawled, which is why tools that search the web by default tend to reflect fixes fastest. Answers drawn from a model's training memory usually persist until that model is retrained or replaced, which commonly takes several months and can take longer for facts repeated across many sources. That is why it is worth working on both at once: fix the pages being retrieved for a quick improvement, and strengthen consistent, current information across the web for the longer-term change. A practical rhythm is to recheck the same set of questions at 30, 60 and 90 days. If nothing has moved after several months of genuine correction work, look again for an outdated source you may have missed.
Is a business responsible for what its own chatbot tells customers?
It can be. In the widely reported case of Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal in February 2024, the airline was held liable after its website chatbot gave a customer incorrect information about bereavement fares, and the tribunal rejected the argument that the chatbot was responsible for its own statements. Outcomes will vary by country and circumstance, but the case is a clear signal that businesses should treat what their own AI tools say as their own statements. In practice that means grounding a chatbot in an approved, regularly updated knowledge base rather than letting it improvise, limiting it to topics it can answer reliably, making it hand over to a person for policies, prices, refunds and anything uncertain, and reviewing conversation logs for wrong answers. This is general information rather than legal advice, so check your obligations with a qualified adviser.
