Articles

How ChatGPT Chooses Which Sources to Cite

ChatGPT chooses its sources in three stages: it rewrites your question into search queries, retrieves a small shortlist of pages from a search index, and then a language model decides — sentence by sentence — which of those pages to build the answer from. Each stage is a separate competition, and pages lose at different stages for different reasons.

The answer pipeline
1
Reformulate
The question is rewritten into one or more search queries — you compete on the paraphrases, not the buyer’s words.
2
Retrieve
A search index returns a shortlist of roughly five pages. Miss it and nothing downstream matters.
3
Synthesize
A language model builds the answer from the shortlist, sentence by sentence, citing the pages it used.

The study that founded the field formalized this pipeline, and the citations and query fan-out visible in today's products still match it. Here's each stage, and where you win or lose.

Stage 1: Reformulation — your question isn't the query

The engine rewrites the user's question into one or more search queries: "best PM tool for a small construction crew?" becomes best construction project management software 2026, sometimes several queries fanned out and merged. Google's AI Overviews do this aggressively — Ahrefs attributes the growing share of citations from obscure pages to exactly this "query fan-out".

What it means: you're competing on the reformulations, not the buyer's phrasing. This is one structural reason keyword targeting decays in AI search: the model paraphrases before it ever searches.

Stage 2: Retrieval — the brutal shortlist

The queries hit a search index, and synthesis works from a shortlist of roughly five pages (count the citations in any answer). Two consequences:

  • AI search is more winner-take-all than Google. Ranking 8th used to earn a trickle of clicks; missing a five-page shortlist earns nothing.
  • SEO's table stakes still apply — indexing, crawlability, AI crawlers not blocked — and 2025 replication research found retrieval rank moves AI citations more than page rewrites do.

But the shortlist is decoupling from classic rankings: by March 2026, only 38% of Google AI Overview citations came from the organic top 10, down from ~76% in mid-2025. You don't need to rank #1 to be retrieved. You need to exist, be crawlable, and match a reformulation.

AI Overview citations that rank in Google’s top 10
July 2025
~76%
March 2026
38%

Share of cited URLs appearing in the organic top 10. Source: Ahrefs, 863K keyword SERPs, March 2026. The other 62% now comes from pages classic rankings would never show you.

Stage 3: Synthesis — where content quality takes over

The model reads the shortlist and writes the answer. Nothing here counts keywords or links; what wins is what a language model finds usable — closer to an editor on deadline choosing what to quote than to a ranking algorithm. Specific, verifiable, well-written material gets lifted; padded pages get skipped. The measured tactics for winning this stage — and their effect sizes — are in How to get cited by ChatGPT.

The winning move is prefabricated sentences: "setup takes 4 days on average, per a 2025 survey of 300 customers" can enter an answer nearly verbatim, with your name on the citation.

When no retrieval happens at all

Semrush's clickstream analysis found ChatGPT invoking live search on only about 34.5% of queries as of early 2026. The rest of the time the answer comes from training data — the statistical residue of the historical web, where long-established brands dominate and there are no citations to win.

You can spot the difference instantly: retrieval-backed answers carry citations; memory answers don't. Win the cited answers you can influence now; every third-party mention you earn is also a candidate for whatever future models train on.

Position inside the answer matters too

Visibility isn't binary. The founding study's metric weights a source's share of the answer by position, with early sentences counting far more — much like the click-through curve down a results page. The prize isn't being citation [4] at the bottom; it's being the source the first two sentences are built on. Track where you appear, not just whether.

Same pipeline, different judges

ChatGPT, Perplexity, and AI Overviews share the shape but differ in indexes, shortlist sizes, and taste — and the differences move fast. Semrush's analysis of 100M+ citations found each engine has its own most-cited domains, and ChatGPT's Reddit citations collapsed from roughly 60% to about 10% of its citation mix between early August and mid-September 2025. Measure each engine separately; one doesn't generalize to the other.

Where are you losing?

Symptom Losing stage Fix
Your site never appears in search for the question Retrieval Build the page; SEO fundamentals; unblock AI crawlers
You rank, but answers never cite you Synthesis Make the page citable
Answers have no citations at all No retrieval fired Play the long game; win cited answers elsewhere
Cited sources are roundups that omit you Retrieval (theirs) Get into the third-party sources

Read next: Why ChatGPT recommends your competitors · How to check what ChatGPT says about you

Want this done for you? Citational tracks your buyers’ questions daily on ChatGPT & Perplexity, shows who’s cited, and drafts the pages that win the answer. See how it works or claim your slots.