Why do AI citations sometimes not support the claim?

When an AI engine answers with retrieval on, several things happen in sequence. It rewrites the question into search queries, fetches a set of pages, writes an answer, and then attaches links. OpenAI's help page on ChatGPT search says it typically rewrites your query into one or more targeted queries before sending them to search providers.

The link step is not a guarantee that the linked page produced the sentence. The model can draw a fact from one retrieved page and cite another. It can also state something from its training data and attach a link that is only loosely related.

So a citation tells you a page was in the mix. It does not tell you that page says what the answer says. That is why the difference between cited and retrieved sources matters before you fix anything.

As of October 2026, the major AI APIs expose citations in different ways, and none of them shows you what the model actually relied on.

• OpenAI: the Responses API returns the full list of URLs the model consulted separately from the citations. OpenAI's documentation says the number of sources is often greater than the number of citations.
• Anthropic: Claude's web search tool cites text spans, with a quoted snippet of up to 150 characters, plus each result's page age.
• Perplexity: citations point to whole sources by an index number, not to a specific passage.

These units are not comparable. A span citation lets you check one quoted line. A whole-source citation only tells you which page to read.

There is a third set nobody exposes: the sources that actually shaped the answer. A page can be retrieved, influence the wording, and never be cited. Another can be cited while contributing nothing.

A 2026 mechanistic study, How Do LLMs Cite?, looked inside Llama-3.1-8B to see how citations get produced. It found citation comes from a distributed, multi-stage process the authors call an attributional ensemble.

The key result: by amplifying specific attention heads and layers, the researchers repaired over 90% of missed citations and eliminated 69% of spurious ones, without harming answer accuracy. The sources did not change. Only the model's citing behaviour did.

That means citation output is partly separable from source use. A model can use a source and fail to cite it, or cite a source it did not use.

Two limits apply. This was one open model in a research setup, not ChatGPT or Perplexity in production. And it measures mechanism, not how often production engines get citations wrong. The practical reading is narrow: do not treat a citation count as a measure of influence.

When you open a cited page and the claim is not there, it usually falls into one of these cases.

1. Wrong attribution. The fact came from another retrieved page, and the link points to the wrong one.
2. Memory with a decorative link. The model answered from training data, frozen at a cutoff, and attached a related page. This often shows up as an outdated answer rather than a hallucination.
3. The page changed. The engine fetched an older version. Your current page no longer says it.
4. The crawler saw a different page. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but do not execute them. If your pricing only appears after JavaScript runs, the engine read a page without it.
5. Loose paraphrase. The page says something nearby, and the model stretched it into a stronger claim.
6. Manipulated content. A 2026 study of LLM search agents showed that manipulated pages in the result set can change what the agent endorses. Hidden text and injected instructions are illegitimate, and they can misrepresent a brand on third-party pages.

Some flagged claims are simply ones nobody can check. That is an unverifiable claim, which is different from a provably false one.

Start from the evidence, not the link.

1. Open the cited page and search for the exact sentence. Note whether it is there, partly there, or absent.
2. If the sentence is there and wrong, the source is the problem. Fix it if it is your page. Contact the publisher if it is not. When engines cite other sites, being named correctly on those sites is associated with getting into the answer, though nothing guarantees it.
3. If the sentence is absent, treat it as an attribution error. Check other likely sources, using the steps in finding which source caused a wrong AI fact.
4. Check what the crawler can read. View the raw HTML of your own page with JavaScript off. Confirm the correct fact is there as a plain, specific sentence. A stated price or number on a fetchable page gives the engine a sentence it can quote.
5. Re-ask the question with search on and off where you can. A matching wrong answer both ways points to model memory.

For engines that cite whole sources, correcting Perplexity answers covers the source-side fixes. For the wider plan, see fixing AI getting your brand wrong. You control your pages and your access settings. You do not control which link a model attaches.