How Do AI Answer Engines Decide What to Cite

An AI answer engine decides what to cite through a pipeline: a question is rewritten into searches, an index returns candidate pages, a system ranks and reads the best passages, and a model writes an answer and attributes some of what it used. The bottom line: being indexed only gets you into the candidate pool, and the published data shows most candidates are never cited, so what matters is being retrievable, clear, and corroborated.

Related: our guide to answer engine optimization defines the term and lists eight steps.

2020year Lewis et al. introduced retrieval augmented generation at NeurIPS

34.5%of ChatGPT queries triggered a web search as of February 2026, per Semrush

15%of 548,000 pages ChatGPT retrieved appeared in the final answer, per AirOps

70%chance a Google AI Overview changes between observations, per Ahrefs

Where the pipeline comes from

The design is called retrieval augmented generation. Patrick Lewis and colleagues introduced it in a 2020 NeurIPS paper as a way to combine a language model's learned knowledge with an external index it can search at question time. Consumer products followed: OpenAI launched ChatGPT search on October 31, 2024, Google began running AI Overviews regularly in May 2024, and Perplexity built its product around retrieval from the start.

The stages

Stage	What happens	What it means for you
1. Decide whether to search	The system answers from memory or searches. Semrush found ChatGPT searched on 34.5% of queries in February 2026, down from 46% in late 2024	Some answers never use your pages at all
2. Rewrite the question	A long prompt becomes short searches. The average ChatGPT prompt was 23 words and its searches 4.2 words in Semrush's 2024 data	Your pages must match the sub questions
3. Retrieve candidates	An index returns pages. ChatGPT search leans on Bing, with 87% of 500+ SearchGPT citations matching Bing's top results in Seer's test	Being in the right index and crawlable is the entry ticket
4. Rank and read	Passages are scored and a few are passed to the model	Answer first, self contained passages survive
5. Generate and cite	The model writes and attributes part of what it used	Only about 15% of retrieved pages were cited in AirOps' data

What research says about the reading step

The "Lost in the Middle" study in the Transactions of the Association for Computational Linguistics found that models use information at the start and end of a long input better than the middle. AirOps, as summarized by Quattr, found 44 percent of ChatGPT citations came from the first 30 percent of a page, and that content updated within three months was cited roughly twice as often, findings Quattr treats as directional.

Why the same question gives different answers

Each step varies. Ahrefs found an AI Overview has about a 70 percent chance of changing between observations, and that AI Mode and AI Overviews cited different sources 87 percent of the time. SparkToro and Gumshoe found less than a 1 in 100 chance that two responses returned the same brand list across 2,961 runs. Judge visibility by the share of many runs, not one answer.

What to do at each stage

Allow the search crawlers so you can be retrieved: OAI-SearchBot, Claude-SearchBot, and PerplexityBot.
Put your content in the raw HTML, since the major AI crawlers do not run JavaScript.
Cover the sub questions around each topic.
Write answer first passages with numbers, sources, and dates.
Earn corroboration on the sources answers cite.
Measure with a fixed question set run many times.

Questions people ask