If an AI assistant quotes your old prices, names a service you no longer offer or sends buyers to the wrong location, start by saving the answer and checking its sources. Correct information you control, request factual corrections elsewhere and retest under comparable conditions. You cannot force an assistant to use your preferred wording or guarantee that the next answer will change.
A mention is not automatically a good result. Being recommended for a service you cannot deliver creates work for your team and confusion for the buyer. Equally, a correct answer linking to an older guide is not necessarily a problem if that guide still helps the reader.
OpenAI explicitly warns that search results and citations can be incomplete, outdated or incorrect. A footnote is a starting point for verification, not proof that every sentence beside it is supported. Google also says its AI features may use multiple related searches and that responses and links can vary. That makes a saved example more useful than a recollection of what the assistant said.
Build a correction record you can actually use
Save the exact prompt, answer passage, assistant, date, market, search setting and cited URLs. Note whether this was a consumer app or an API. Keep the original answer: a new run is another observation, not a replacement for the evidence.
Use one row per claim. The columns can be simple: claim, cited URL, final destination after redirects, evidence on the page, verified current fact, severity, owner and next check. Mark the support as direct, partial, absent or inaccessible. An inaccessible page is unknown, not automatically false.
Illustrative example: an assistant says a fictional workshop offers Saturday collections at £25. Its cited service page supports the £25 charge but says nothing about Saturdays. That is two checks: verify the fee against the current approved price list, and investigate the opening-hours claim separately. Do not mark the whole answer accurate because one number matches.
Prioritise mistakes that could cause a failed visit, a wrong purchase or an unrealistic expectation. Slightly awkward descriptions belong behind incorrect service areas, availability, eligibility and prices. Give each correction a named owner so the spreadsheet does not become a list nobody acts on.
Fix the source before asking for a different answer
For your own site, put the current fact on the page where a customer would expect to find it. Include conditions: whether a price includes VAT, where the service is available and what a stated turnaround depends on. Check related pages for contradictions. Update structured data only to match the visible information; it is not a place to hide a stronger promise.
For a business profile you manage, use the provider’s editing process. For an independent publisher, prepare a short factual correction with the incorrect sentence, current evidence and requested amendment. Do not demand that a genuine critical opinion be removed. If nobody cites a source for a claim, record that limitation rather than inventing the assistant’s reasoning.
Ask your web team to investigate duplicates, retired pages and redirects when those are genuinely part of the problem. Do not redirect a useful guide to a sales page simply because you prefer the sales URL to receive a citation. Customers still need the guide.
For Google AI features, the official guidance is ordinary SEO: indexable pages, crawl access, useful internal links and important information in text. Google says there is no special AI schema or mandatory AI text file. For ChatGPT search, OpenAI identifies OAI-Searchbot access and host access as eligibility checks, while explicitly saying placement is not guaranteed. Neither technical access nor a factual correction guarantees inclusion.
Retest, record limits and decide what to do next
Run the original question again in a fresh conversation with comparable market and search settings. Save the new answer separately. Try a couple of natural variations without instructing the assistant to favour your brand or URL. Check again later for a commercially important error: the next answer alone cannot prove that your edit caused the change.
Your useful outcome is a cleaner source of truth and fewer observed factual errors, not a promise that every model now understands the business. Track corrections completed, unresolved high-risk claims and repeat errors. Keep this alongside, rather than inside, your broader mention count.
Foundry’s AI SEO service covers fact and content structuring, crawler access and ongoing visibility tracking. If your team has a specific wrong answer to investigate, bring that answer and the current approved fact to the conversation. It is a clearer starting brief than “make us rank in AI”.
