How AI answer engines choose the sources they cite
They retrieve candidate pages from a search index, often across several sub-queries, then cite a few. In controlled tests topical match and position matter most; formatting alone barely counts.
The short answer
AI answer engines work in stages. They retrieve candidate pages from a search index, often by splitting one question into several sub-queries. They select a few of those pages to cite. Then each cited page shapes the written answer to a different degree.
Controlled studies published in 2026 find that topical relevance and a page’s position among the retrieved results dominate which source gets cited. Concrete details such as prices and recent dates help. Formatting changes on their own barely register. The table near the end separates what has been measured from what is only claimed.
What happens between a question and a cited answer?
Four steps are worth separating, because each one has a different lever and a different body of evidence.
- Query fan-out: for many questions the engine runs several related searches across subtopics, which widens the pool of candidates.
- Retrieval: each search pulls pages from an index that could support an answer. If a page is not in that index, nothing later in the chain can help it.
- Selection: from the candidates, the model decides which few sources to cite.
- Absorption: a cited page may contribute facts, structure or wording to the answer, or it may appear as a link that shaped almost nothing.
What does Google say about how its AI features pick sources?
Google’s documentation describes the system without publishing the selection rules. Its May 2025 explainer, “AI Overviews and AI Mode in Search”, says AI Overviews use a customised Gemini model working with Google’s existing quality and ranking systems and the Knowledge Graph. It says Google integrated its core web ranking systems into the feature and built it to show information backed up by top web results. For “Your Money or Your Life” topics such as health and finance, it describes a higher bar for supporting sources.
The Search Central page on AI features sets the entry condition. A page must be indexed and eligible to be shown with a snippet. There are no additional requirements, and no special markup, structured data or AI text files are needed. The usual preview controls (nosnippet, data-nosnippet, max-snippet, noindex) also apply to AI features.
The same explainer notes that AI Mode uses more advanced models than AI Overviews, so a different set of responses and links will likely appear for the same query. Being cited in one Google surface does not imply being cited in the other.
What is query fan-out and why does it matter?
Query fan-out is Google’s name for splitting one question into subtopics and searching each. Google’s announcement of AI Mode in May 2025 described it as breaking a question into subtopics and issuing many queries at once on the user’s behalf. The same post said Deep Search can issue hundreds of searches before writing a cited report.
The AI features page says both AI Overviews and AI Mode may use the technique, and that while a response is generated, Google identifies more supporting pages, so it can show a wider and more diverse set of links than a classic search.
The practical consequence is our inference, which Google has not quantified: a page can be cited in the answer to a broad question because it ranks for one of the narrower sub-questions. Pages that answer one specific sub-question well have more ways into an answer than pages that cover a broad topic thinly.
Other engines run their own retrieval. OpenAI’s crawler documentation says OAI-SearchBot is the crawler that surfaces sites in ChatGPT’s search features, and that sites which opt out will not appear in ChatGPT search answers except as navigational links. Retrieval is therefore a separate gate for each engine.
What do controlled studies show about selection?
The clearest evidence comes from “What Gets Cited: Competitive GEO in AI Answer Engines” (arXiv 2605.25517, May 2026). The authors placed exactly two candidate sources in a model’s context, varied one content factor at a time and recorded which source the model cited first. They anonymised brands and counterbalanced the order of the two sources to separate content effects from position bias.
The scale was 252,000 trials across six models: Gemini 2.5 Flash, GPT-5 Nano, GPT-5 Mini, GPT-5.2, Claude 3.5 Sonnet and Kimi K2 Thinking. The corpus was 100 product review articles across 50 consumer categories, and the study tested 18 factors.
The results, reported as odds ratios:
- Topical relevance was decisive. An on-topic source beat an off-topic one by very large margins in every model.
- Position in the retrieved list had a very large effect on its own. The first-listed source was heavily favoured.
- Price information and a recent timestamp (against an old one) helped consistently across models.
- Completeness and trust signals, such as specifications, comparisons, confident wording and claims backed by evidence, had smaller but real effects that varied by model.
- Formatting-only changes, such as structured against dense text, had odds ratios close to 1, meaning almost no effect.
Models also differed in how sensitive they were. Kimi K2 responded to the most factors; Gemini 2.5 Flash and Claude 3.5 Sonnet were more selective.
The limits are stated in the paper. Real systems retrieve 5 to 10 or more pages, not two. Brand trust was removed by design. The content was generated with an LLM, covered only product reviews and used a synthetic search step rather than a live engine. The findings describe pairwise preference under lab conditions.
What is absorption and how is it different from being cited?
A citation is a link. Absorption is how much of the answer actually came from that page. “From Citation Selection to Citation Absorption” (arXiv 2604.25707, April 2026) measures both, using a public dataset of 602 controlled prompts run on ChatGPT, Perplexity and Google (AI Overviews and Gemini).
Search was triggered on almost every prompt on all three platforms. The number of cited sources per answer differed widely:
| Platform | Average sources cited per answer | Average influence score of a cited page |
|---|---|---|
| ChatGPT | 6.88 | 0.271 |
| 12.06 | 0.058 | |
| Perplexity | 16.35 | 0.065 |
The influence score is the authors’ own composite, from 0 to 1, built from repeated references, how early a page is used, how much of the answer it covers and text overlap with the answer. On that measure ChatGPT cites fewer pages but each one shapes much more of the answer.
Pages in the top quarter for influence averaged 1,943 words and 10.6 headings, against 170 words and 0.85 headings in the bottom quarter. Pages containing statistics, definitions, comparisons or how-to steps scored higher than pages without them. Question-and-answer formatting scored slightly lower.
The authors are explicit that these are descriptive results with no causal claim, that the influence score is a proxy and that platform behaviour may have changed since the snapshot.
Do GEO tactics transfer between engines?
Poorly, on current evidence. The original GEO paper (Aggarwal and colleagues, first posted November 2023, accepted at KDD 2024) reported visibility gains of up to 40% from rewriting content, measured on its own GEO-bench benchmark.
A critical survey of 45 GEO studies, posted in July 2026, reads that result narrowly. It finds the gains hold only once a source is already in the model’s context, so they say something about how retrieved content is used and little about how content gets retrieved. It also reports that generic heuristics transfer poorly between engines and that rewrites aimed at citation can hurt retrieval. Its conclusion is that no reviewed technique shows a stable, long-term causal effect across platforms.
What is measured and what is claimed?
| Statement | Source | Type of evidence |
|---|---|---|
| AI Overviews use Google’s core ranking systems and show information backed by top results | Vendor description of its own system | |
| Pages need only be indexed and snippet-eligible; no special markup or AI files | Vendor guidance, authoritative for Google only | |
| Clicks from results pages with AI Overviews are higher quality | Vendor claim; no figures published | |
| Topical relevance and list position dominate which source is cited | arXiv 2605.25517 | Controlled experiment, 252,000 trials, six models |
| Prices and recent dates help; formatting alone barely matters | arXiv 2605.25517 | Same experiment |
| ChatGPT cites fewer sources but each shapes more of the answer | arXiv 2604.25707 | Descriptive measurement, 602 prompts |
| Rewriting content raises visibility by up to 40% | Aggarwal et al. | Benchmark experiment; survey finds it holds only for already-retrieved sources |
| Wikipedia, YouTube and Reddit were 15% of links in Google’s AI summaries | Pew Research Center | Observed browsing and scraped results, March to April 2025 |
The Pew figure is a useful check on intuition. In its 2025 analysis of 68,879 searches, the three most-linked sites made up 15% of sources in AI summaries and 17% in standard results. The AI summaries leaned on the same big sites as ordinary results, slightly less heavily.
What to do
- Treat retrieval as the first problem. Confirm your key pages are indexed in Google and that each engine’s search crawler, including OAI-SearchBot, is allowed in robots.txt and at your CDN.
- Write pages that each answer one specific question completely. Fan-out means narrow sub-questions are entry points.
- Put concrete, checkable facts on the page: prices, dates, specifications, figures with their sources.
- Show a real, recent updated date where the content has genuinely changed.
- Do not spend effort on formatting alone. Use structure because it helps readers, and expect little citation benefit from it by itself.
- Measure citations per engine on a fixed set of prompts, and record which pages are cited and how much of the answer they shaped.