How do I find which source caused AI to get a fact wrong?

Every AI answer draws on two inputs. One is the set of pages the engine retrieved for that query. The other is what the model learned in training, frozen at a cutoff date. Your job is to work out which input carried the wrong fact, then fix the page that states it.

1. Record the answer. Save the exact wording, the platform, the date and whether it showed citations. Use a logged-out or memory-off session, because OpenAI says ChatGPT may use saved memories when it rewrites a search query.
2. Open every cited link. Search each page for the wrong fact. If one states it, you have your source.
3. Pull the retrieved set. If no cited page contains the error, rerun the question through an API with search on and capture every URL the model consulted.
4. Rerun with search off. If the wrong fact still appears with no retrieval, it lives in the model's training knowledge.
5. Fix the source. Correct your own page, or contact the site that published the error.

Say you run a hypothetical project tool priced at $49, and an AI answer says $39. The steps above tell you whether that $39 sits on an old review page, on your own outdated footer, or only in the model's memory. Each answer needs a different fix.

Citations are the cheapest clue, so check them first. They are also incomplete. Engines often use pages they never link, and they sometimes link pages that don't support the sentence next to them. Our page on why AI citations sometimes don't support the claim covers that failure in detail.

The evidence is clear that a citation is a noisy signal of use. A 2026 mechanistic study of citation in retrieval-augmented generation (arXiv 2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources. Citation emission is partly separate from which pages shaped the answer.

Platforms also cite differently. Anthropic's API cites short text spans. Perplexity cites whole sources by number. Google AI Overviews show supporting links and offer no official API, so the visible links are all you get there. Google also says AI Overviews and AI Mode may use different models and techniques, so check each surface on its own.

Three APIs separate what the model read from what it cited, as of October 2026. That split is where hidden sources show up.

Run the same question several times and log both lists side by side. A wrong fact on a retrieved but uncited page may still have shaped the answer. Our playbook on logging retrieved vs cited sources sets this up step by step.

One caveat. The API is not the consumer app. Models, retrieval settings and query rewriting differ, and OpenAI says ChatGPT rewrites your prompt into one or more targeted queries. Treat an API trace as a strong lead, not a perfect replay of what a user saw.

Retrieval is per query and discretionary. ChatGPT searches at the model's discretion. Claude.ai search is a user toggle. So some answers never touch the live web.

OpenAI documents a knowledge cutoff of February 16, 2026 for its GPT-5.6 models and August 31, 2025 for the Instant path (chat-latest). A fact that changed after those dates can come back stale in any answer given without search. If the model states something that was never true anywhere, that is a hallucination instead. Telling a hallucination from an outdated answer matters, because the fixes differ.

Be plain about the limit. Nothing you publish this quarter reaches a search-off answer. What you control is whether search-on answers find the correct fact, which is where the next section comes in.

If the wrong fact is on your site

Often the culprit is you. The same fact stated differently on the pricing page, the about page and the footer gives an engine two versions to choose from. Run a check for contradicting facts on your own site and make every page agree.

Then make sure crawlers can read the correct version. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so a price rendered only by JavaScript may be invisible to them. Put the fact in the HTML and state it plainly, in one sentence an engine can quote.

If the wrong fact is on a third-party site

Contact the site owner. Give them the correct fact, a link to the page that states it, and the date it changed. Review sites, directories and partner pages often update when asked. If a competitor's page is the source, see what to do about a misleading comparison page. For Perplexity specifically, correcting the sources it cites walks through each step.

Watch for one rare case. Some third-party pages carry hidden or injected instructions aimed at AI systems. That is a manipulation risk to report, never a tactic to copy.

Then recheck

Fixing the source removes the wrong input. It does not guarantee the next answer, because no publisher controls what a model says. Log each error and recheck it on a date, using an AI error correction tracker. The free checker on this site gives you a quick snapshot of how engines answer up to three buyer questions after a fix.

Can I see which sources Google AI Overviews used? Only the supporting links it shows. Google offers no official API for AI Overviews, so you can't pull a hidden retrieved list the way you can with OpenAI, Anthropic or Perplexity. Open each visible link and search it for the wrong fact. Check AI Mode separately, since Google says the two may use different models and techniques and return different links.

Why does the API give a different answer than the consumer app? The app and the API often run different models and retrieval settings. The app may search at its own discretion, use memory, or rewrite your prompt into other queries. The API runs exactly what you configure. Use the API to trace sources, and use a logged-out, memory-off app session to confirm what real users see.

What if no cited or retrieved page contains the wrong fact? Rerun the question with search off. If the error persists, it comes from training data or is a hallucination. If it appears only with search on, the model may have merged facts from several pages, or misread a page you already checked. Look for near matches such as an old price, a former product name, or a similarly named company.

How long does a correction take to show up in AI answers? Nobody can give a reliable timeline. Search-on answers can change once engines recrawl the corrected page and retrieve it for that query. Search-off answers change only when a vendor ships a model with a later knowledge cutoff. Recheck on a fixed schedule and record the date of each result, so you can tell a real change from normal variation.