Retrieved but not cited: what AI engines read before they answer

By Abiot Y. Derbie - Founder, RankEcho - Updated 2026-11-03

Key findings

These figures describe a small, recent sample: the questions of seven audits for three websites, asked through each engine's API over three days. They show what the engines did, not why.

• Most pages an engine reads do not make it into the answer. ChatGPT cited 42% of the pages its search returned and Claude 54%.
• Place in the results matters on Claude. Pages in its top three results were cited 70% of the time; pages 11th or later, 35%.
• Archive, tag and sitemap pages get read and passed over. ChatGPT cited 1 of the 18 it was given. On lifebylogic.com, 21 of the 74 times ChatGPT read one of its pages were listing pages, and none was cited.
• ChatGPT often answers without searching at all: 57 of 143 runs (40%) had no web search, so nothing could be retrieved. Claude searched in every run.
• The same page is not treated the same way twice. Of the pages returned in three or more runs, most were cited in some runs and not others.

Why retrieved but not cited is a useful signal

An engine that searches before it answers does two things: it fetches a set of pages, then it writes an answer and links some of them. If your page was never fetched, the problem is being found for that question: a page that matches it, or mentions on the pages the engine does fetch. If it was fetched and not cited, the engine had your page in front of it and preferred another. That is a problem with the page: whether it answers the question directly, early and specifically enough to quote.

Most visibility tools only see the answer, so they cannot tell these two cases apart. The OpenAI and Anthropic APIs list the pages their web search returned next to the pages the answer cites, which is what this study counts.

The sample

The study counts every run between 2026-11-01 and 2026-11-03 in which the OpenAI API (the models behind ChatGPT) or the Anthropic API (Claude) answered an audit question with web search allowed, and the response listed what the search returned. That is 246 runs for 143 questions from 7 audits of 3 websites: lifebylogic.com, a free psychology-test site; RankEcho's own site; and a small law firm, which is not named.

A returned page counts as cited when the answer links the same URL, ignoring the scheme, a trailing slash and OpenAI's tracking tag. Up to 20 returned pages are stored per run; 18 runs reached that cap, so a few returned pages are missing. ChatGPT's answers cited 27 links that were not in its stored list, mostly in capped runs; they are left out of the rates.

Finding 1: higher in the results, more often cited (on Claude)

Claude's citation rate fell steadily with place in its search results, from 70% for the top three to 35% for the 11th and later. Claude sometimes ran more than one search, so later places include the results of a second search.

ChatGPT's list showed no clear order: the top three were cited most, but pages 11th and later were cited more often than pages 7 to 10. Its list of sources may not be in ranking order, so read its row with care.

Finding 2: which kinds of page get passed over

Each returned page was sorted by its URL alone, so the groups are rough. Archive, tag and sitemap pages were returned but almost never cited. Comparison and list pages were cited less often than average by ChatGPT. ChatGPT's search returned many homepages, and it cited half of them.

Finding 3: the same page is cited one day and not the next

88 pages came back in three or more ChatGPT runs and 65 in three or more Claude runs. Most were cited in some of those runs and not in others: 54 of 88 for ChatGPT and 40 of 65 for Claude. Only 13 and 17 were cited every time, and 21 and 8 never.

So one run that does not cite your page proves little. Ask the same question several times, and judge a page by its rate across runs.

A real case: lifebylogic.com

lifebylogic.com publishes free psychology and cognitive tests with guides that explain them. Across its audits, ChatGPT's search returned one of its pages 74 times and the answer cited it 22 times (30%). Claude's returned one 58 times and cited it 40 times (69%).

The gap on ChatGPT comes largely from what it found. 21 of its 74 reads were the sitemap, numbered archive pages of the article library and a lab's article list. None of those was cited. Without them, ChatGPT cited 42% of the site's pages it read. The pages that were cited were the ones that hold an answer: the homepage, the list of tests, and individual tests and guides.

The lesson is not to hide listing pages. It is that a listing page is a weak landing spot for a question: when search returns it, the engine still has to find the answer elsewhere. The fix is a page that answers the question directly, linked clearly from the listing, so search has something better to return.

What to do when your page is retrieved but not cited

• Put the answer to the question in the first lines of the page, in plain words an engine can quote: a number, a definition, a yes or no with the reason.
• Make the page about one question. A page that covers ten topics loses to a page that answers the one that was asked.
• Add the specifics the competing cited pages have and yours lacks: prices, steps, dates, named sources. Look at what was cited instead of you for the same question.
• Make sure the page that answers is the one search returns, not a list, tag or archive page that only links to it.
• Measure across several runs before and after a change, on each engine separately.

Limits

This is a small, recent and observational sample: three websites, seven audits and three days, with most runs from two of the sites. Each site's questions were written for that site. Engines change their models and search behaviour often. The figures do not show why an engine chose one page over another, and they should not be read as a prediction for another site, engine or date.

Only ChatGPT and Claude are counted, because their APIs list the search results separately from the citations. Gemini's and Perplexity's responses list the same pages as both, so they cannot show this. The kinds of page come from URL patterns, not from reading the pages. RankEcho will update the study as more audits record what each search returned.

Frequently asked questions