How AI Engines Choose Which Sources to Cite (2026)

AI engines choose citations in two stages, and almost all the attention goes to the wrong one. Retrieval decides which pages become candidates. Selection decides which candidates get quoted. Most brands optimize hard for the first and never measure the second.

The gap is large. In an AirOps analysis of 548,534 pages retrieved across 15,000 ChatGPT prompts, reported by Search Engine Land in March 2026, only 15% of retrieved pages were cited. The other 85% were fetched, read, and discarded before the answer was written.

Stage one: retrieval still runs on classic signals

Getting into the candidate pool is largely a conventional SEO problem. In the same dataset, 55.8% of cited pages ranked in Google’s top 20, and pages at position 1 were cited roughly 3.5 times more often than pages outside it.

Two things complicate that.

Which index matters. ChatGPT’s search leans heavily on Bing. A Search Engine Land case study in April 2026 followed a well-reviewed hotel appearing in only 1.5% of 68 runs of the same prompt, and found Bing rankings aligned with ChatGPT mentions far more closely than Google rankings did.

The query is rarely the one typed. Engines fan a prompt out into many sub-queries. In the AirOps data, 15,000 prompts became 43,233 queries, 32.9% of cited pages appeared only in those expanded results, and 95% of the expanded queries had no traditional search volume. You cannot target those directly, only cover the territory they come from.

Stage two: selection runs on different signals

Once a page is retrieved, a different test applies: can this passage be lifted and trusted as the answer?

Search Engine Land’s November 2025 analysis of the content traits models quote most found short, self-contained answer capsules were the strongest citation driver, with 34.3% of cited posts combining an answer capsule with original or owned data. The same analysis flagged heavy link density inside those capsules as a drag.

That matches what the decoupling data shows. AI Overview citation overlap with the top 10 organic results fell from roughly 76% in late 2024 to about 17 to 38% by February 2026, depending on methodology. Ranking gets you considered. Something else gets you quoted.

The signals that correlate with citation

Two readings are worth drawing out.

Brand beats links. Branded mentions track AI visibility roughly three times more closely than backlinks do. That is a different acquisition strategy: accurate description in places engines already trust, rather than link volume. See digital PR for AI search.

Scale correlates, but is not the only route. The referring-domain finding describes very large sites. Smaller brands compete on the other two: review-platform presence and consistent branded mentions, both of which are achievable without enterprise scale.

This is also why the underlying work is entity SEO: making your brand an unambiguous, consistently-described thing the model can identify and trust.

What the research actually established

The first formal framework here is the GEO: Generative Engine Optimization paper (Aggarwal, Murahari et al.), peer-reviewed at ACM SIGKDD KDD 2024. It introduced the first paradigm for optimizing content to be cited inside generative answers rather than to rank in a list of links.

Its headline result: methods improved a source’s visibility by up to about 40% on the authors’ GEO-bench. Read that carefully. It is a benchmark maximum, measured on a synthetic GPT-4-based engine using proxy metrics, not a live commercial engine. It tells you the shape of what works, not a number to expect in production. Follow-up work in 2025 and 2026 suggests current methods may be insufficient in fully realistic settings, so treat GEO as an active research area rather than a solved playbook.

Legacy tactics do not transfer

The same paper is blunt about the old toolkit. Keyword stuffing reduced citation visibility by about 9% on the Position-Adjusted Word Count metric, and the authors concluded traditional SEO tactics “offer little to no improvement” for generative engines.

So the reflex to cram in target phrases actively works against you. What the research and the 2026 correlation data point toward instead:

1. Be citable, not just rankable. Write clear, self-contained statements a model can lift without ambiguity. See content chunking for AI.
2. Put a verifiable fact in the passage. Original or owned data paired with a clean answer is the highest-performing combination in the citation research.
3. Build entity corroboration. Consistent facts across your site, structured data, and third-party profiles. See structured data for AI search.
4. Earn branded mentions, not just links.
5. Cover the question space, because a third of citations come from queries nobody typed. See topic clusters that AI engines cite.
6. Stop optimizing for density. It measurably hurts.

It also explains why so much of the work happens off your own site, and why being retrieved is not the same as being cited, which we unpack in retrieval vs citation.

How to tell which stage is failing you

A quick diagnostic, using your own prompt set:

Where citation selection leaves you

Citation and ranking are drifting apart, and the signals that predict citation lean toward entity strength: third-party corroboration, branded mentions, and passages that can be quoted without editing. The one clear “don’t” from peer-reviewed work is keyword stuffing. Everything else is a matter of making your brand an entity engines can recognize, verify, and quote with confidence.

If you want to know how AI engines currently see and cite your brand, and which of the two stages is failing you, request an AI visibility audit and we will map it.