How Do AI Assistants Choose Which Sources to Cite? Three Stages

By Katleen, Founder of PARCHMNT · Updated October 29, 2026

AI assistants choose which sources to cite in three stages. A web search narrows the candidates to pages a search engine ranks. The assistant then lifts the passages that answer the question directly, and it prefers passages whose claims carry evidence, a named source and a date. Most pages fall out at the first stage, before the writing is read.

This post is built on what Google, OpenAI, Anthropic and Perplexity document about their search features and crawlers, and on Ahrefs' 2025 and 2026 counts of which pages each assistant cites. It also draws on Pew Research Center's 2025 click study and on the 2023 research that first measured what lifts a page into a generated answer. When we read the first page for this search, every result was a vendor or agency post and none came from an assistant maker.

What happens between the question and the answer?

An assistant with web search runs a search, reads the results, writes an answer and attaches citations to the passages it used. Anthropic's documentation for Claude describes the sequence plainly: Claude determines when to search based on the prompt, the API runs the searches, and the final response comes with cited sources.

Google describes the first step for AI Mode as a query fan-out technique, issuing multiple related searches concurrently across subtopics. Anthropic says a simple factual question typically takes one to three searches, and comparative research across several candidates can take ten or more. Perplexity's help center says it searches the internet in real time and gives each answer numbered citations linking to the original sources.

Three stages fall out of those descriptions:

1. Search: which pages are in the running at all.
2. Passages: which parts of those pages get lifted into the answer.
3. Trust: which of the lifted passages the assistant cites.

Stage one: which pages are even in the running?

A page is in the running when the assistant's search crawler is allowed to read it and a search engine ranks it for the question. Google's documentation says a page must be indexed and eligible to be shown in Google Search with a snippet to appear in AI Overviews or AI Mode.

The crawler part is a robots.txt line. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Anthropic says disabling Claude-SearchBot prevents its system from indexing your content, and Perplexity says PerplexityBot is designed to surface and link websites in its results. Our two-minute robots.txt check lists the lines.

The ranking part is ordinary search, and the best measurement of it has moved. Ahrefs found in July 2025 that 76.10 percent of pages cited in AI Overviews ranked in Google's top ten. Its March 2026 update found 37.9 percent in the top ten, 31.2 percent at positions 11 to 100 and 31.0 percent beyond the top 100. Ahrefs says its parsing improved in between, so the two are not directly comparable.

A top-ten page is still the most likely kind to be cited, and a page can be cited without ranking. No equivalent figure is published for ChatGPT or Claude, and neither OpenAI nor Anthropic names the search index behind its feature, so this post will not guess.

Stage two: which passages get lifted?

Assistants quote passages, not pages, and a passage is short. Anthropic's API returns citations with up to 150 characters of the cited text, about two sentences, so the sentence that answers the question has to carry its own evidence.

A paragraph that opens with the answer and names its subject survives being cut out. So does one that puts the figure, the source and the date in one sentence.

When one question has been fanned out into several searches, each search is looking for a section that answers it. A heading phrased as the real question, followed by two or three sentences that answer it, is the unit an assistant finds. Our guide to writing a blog post that gets quoted by AI covers the shape section by section.

Stage three: what makes one passage more trusted than another?

Among the passages that answer, assistants favor the ones that show their evidence. That means a figure with a linked source, a quotation from a named person, and a date that says when the fact applied. It is the clearest measured finding in generative engine optimization, and nobody publishes the weights behind it.

Aggarwal and colleagues at Princeton University and Georgia Tech built a benchmark of 10,000 queries in 2023. Adding citations, quotations and statistics to a page raised its visibility in generated answers by 30 to 40 percent. Keyword stuffing, tested on Perplexity, performed about 10 percent worse than the unchanged page. Princeton's record of the paper, presented at KDD 2024, puts the headline gain at up to 40 percent.

One detail matters for small sites: adding sources lifted the visibility of the page ranked fifth by 115.1 percent while the top-ranked page lost ground.

Google's description of its own AI features points the same way, without the numbers. "Helping people discover content from the web remains central to our approach," writes Robby Stein, VP of Product at Google Search, in the post that introduced AI Mode. As of October 2026, no company publishes how it weighs a source's authority, freshness or reputation.

Where does each assistant look?

Each assistant documents a different set of inputs, and the independent counts show them citing different webs. Ahrefs compared the fifty most-mentioned sites across ChatGPT, Perplexity and AI Overviews in June 2025 and found only 7 appeared in the top 50 for all three.

Community and video sources weigh more on Google and Perplexity than on ChatGPT. A thread answered under your own name, or a video that answers a question, does more there. Wikipedia and news lead on ChatGPT, which rewards being written about by established third parties. Our posts on getting cited by Perplexity and getting recommended by Claude take each one in turn.

How do you write one page for all of them?

Write one page that passes all three stages. Let every search crawler read it, win a specific search with it, open each section with a self-contained answer, and give each claim its evidence in the same sentence.

PARCHMNT is a writing tool that researches what ranks for a search, writes the page, checks every claim and tracks where it lands in Google and AI answers. Its 41 content checks score the passage stage: does each section open with a standalone answer, does each figure carry a source. Its crawler check covers the first stage, and its fact check gate holds back any figure whose source cannot be opened, which is the trust stage made literal.

In our experience building those checks, the pages that fail are rarely thin. We find the answer in paragraph four.

Pew Research Center found in 2025 that people clicked a traditional result in 8 percent of visits with an AI summary against 15 percent without. They clicked a link inside the summary in 1 percent of visits. A page that ranks but is not cited by AI assistants loses about half its clicks to the summary. A page that is cited is one of the few links left.

Key takeaways

• Three stages: search (indexed, crawler allowed, ranking), passages (short, self-contained), trust (sourced figures, named quotes, dates).
• Ranking is the most reliable route into AI citations, but Ahrefs' 2026 update found 31 percent of AI Overview citations beyond the top 100, so it is not the only one.
• Anthropic's citations carry up to 150 characters, so the sentence that answers must hold its evidence.
• Citations, quotations and statistics raised visibility 30 to 40 percent; keyword stuffing lowered it.
• The assistants cite different webs: only 7 of the 50 most-mentioned sites were shared across ChatGPT, Perplexity and AI Overviews.

Frequently asked questions

Do AI assistants cite pages they were trained on?

Not in the sense that matters. Citations in a search-enabled answer come from the pages the search returned at the time of the question. That is why the companies run separate search crawlers and let you block the training crawler without leaving search. An answer given without search draws on training data and usually carries no citations you can act on.

Does the date on a page affect whether it is cited?

The date is at least visible to the assistant. Anthropic's search results include a page_age field recording when the page was last updated, and Perplexity describes its search as real time. No company publishes how it weighs freshness, so keep the date the text last changed visible and honest, and refresh when the facts age rather than on a calendar.

Does structured data help an assistant choose a source?

Not by the companies' own accounts. Google says there is no special structured data you need to add to appear in its AI features. The OpenAI, Anthropic and Perplexity crawler pages do not mention schema markup.

How many sources does one answer use?

Nobody publishes a fixed number. Anthropic's documentation says a simple question typically takes one to three searches and comparative research ten or more. Each search returns several results, and only the passages used are cited.

How do AI assistants choose which sources to cite, in one paragraph

Assistants search, lift passages, and trust the passages that show their evidence, and each company does it on a different web. If you would rather have the research, the writing, the checks and the monthly asking done for you, try one article free.