Explanations of how AI answer engines pick their sources tend to blur three kinds of claim: what the platforms themselves document, what independent studies have measured, and what practitioners infer. This piece keeps them apart. Where a mechanism is documented by a platform, we say so. Where it has been measured, we give the study, its sample and its period. Where it is a hypothesis, we label it as one.

Step one: whether the engine searches at all

A citation is only possible when an engine retrieves web content instead of answering from its training data. None of the major assistants publishes how often that happens for real users, and the available proxies vary widely.

  • Measured, Google. Pew Research Center found AI summaries on 18% of the Google searches made by its 900-person panel in March 2025, rising with query length from 8% of one- or two-word searches to 53% of searches of ten words or more. Seer Interactive's 2026 data showed AI Overviews on 36% of informational queries but 5% of transactional ones, and on 95.4% of the comparison queries it tracked.
  • Measured, chat assistants. Similarweb found web citations in 6.8% of ChatGPT conversations in May 2026 (US, desktop), rising to 22.6% for travel. Muck Rack found ChatGPT cited sources in 96% of the responses it analyzed, Gemini in 82% and Claude in 55%.

The gap between those two ChatGPT figures is instructive. Prompts written by researchers to elicit recommendations are not a sample of how people use the product, and Muck Rack's post does not describe how its prompts were built.

Step two: retrieval, and whose index

Documented. Google says AI Overviews and AI Mode use retrieval-augmented generation grounded in its core ranking systems, and that a page must be indexed and eligible to show a snippet. Since August 31, 2026, a site must also remain included under Search Console's generative AI control, which is the default. OpenAI's crawler documentation says OAI-SearchBot surfaces websites in ChatGPT's search features and that sites blocking it will not be shown in ChatGPT search answers, apart from navigational links. Perplexity says PerplexityBot surfaces and links sites in its results, while its user-triggered fetcher, Perplexity-User, generally ignores robots.txt. Anthropic lists Claude-SearchBot and Claude-User and says blocking either may reduce a site's visibility in Claude's search results. Microsoft's Bing team describes Copilot breaking pages into smaller pieces that are evaluated separately.

Measured. The engines do not retrieve the same things. Ahrefs ran 15,000 long-tail prompts and found that on average only 12% of the URLs cited by ChatGPT, Gemini, Copilot and Perplexity also ranked in Google's top 10 for the same query. Perplexity was highest at 28.6%, and Copilot overlapped most with Bing. Crawler rules are also an imperfect gate: the Tow Center at Columbia found in early 2025 that Perplexity identified excerpts from publishers that had blocked its crawler, and that several tools cited syndicated copies of articles on Yahoo News or AOL instead of the originals.

Step three: query fan-out

Documented. Google describes query fan-out as a set of concurrent, related searches on subtopics of the user's question; its own example turns a question about a weedy lawn into searches about herbicides, chemical-free removal and prevention. At I/O in May 2025, Google said AI Mode's Deep Search can issue hundreds of searches. In November 2025 it said Gemini 3 lets fan-out run more searches and find content it may previously have missed.

Measured, and consistent with fan-out. In Ahrefs' data, the share of AI Overview citations that also rank in Google's top 10 for the original query fell from 76% in July 2025 to about 38% in March 2026, although Ahrefs improved its citation parsing in between and says the studies are not directly comparable. BrightEdge puts the figure at about 17%. Ahrefs also found that when AI Mode and AI Overviews answered the same query in September 2025, they cited the same URLs only 13.7% of the time, even though their answers had 86% semantic similarity.

Hypothesis. The common inference is that pages which rank well for the sub-queries win the citations. Google's description makes it plausible, but it has not been demonstrated directly, because Google does not expose its fan-out queries. When Ahrefs tested one version of the idea in 2025, expecting cited pages from outside the top 10 to rank for many long-tail queries, it found the opposite. Google also warns that creating separate pages for fan-out variations, primarily to manipulate its AI features, violates its scaled content abuse policy.

Selection signals: what the evidence supports

Authority and third-party coverage. The most consistent finding is a tilt toward earned media. A University of Toronto preprint (Chen et al., September 2025) queried the API versions of GPT-4o search, Perplexity, Claude and Gemini with product-ranking prompts and found they drew far more heavily on third-party editorial and review sites than Google, which returned a more balanced mix including brand-owned and social pages. Ahrefs' study of 75,000 brands found brand web mentions correlated more strongly with AI Overview visibility (0.664) than backlinks (0.218). Both findings are correlational, and Google cautions that seeking inauthentic mentions does not help.

Freshness. Ahrefs' analysis of 17 million citations found AI assistants cited younger pages than organic search on average; ChatGPT favored the newest content, while AI Overviews cited pages slightly older than organic results. The Toronto study found recency varied by engine and vertical, with Claude's automotive citations averaging about 331 days old. Freshness appears to matter more to some engines than others; it is not a universal rule.

Content features. The GEO paper (Aggarwal et al., KDD 2024) found that adding statistics, quotations and citations to a source raised its visibility by up to 40% in a simulated engine built on GPT-3.5, while keyword stuffing did not help and a more authoritative tone produced no significant gain. The limits are in the setup: a controlled engine working from the top five search results per query, with 2023-era models.

Formatting and markup. Here the platforms disagree. Microsoft's October 2025 guidance recommends clear headings, question-and-answer formats, short self-contained answers and schema markup. Google's July 2026 guide says content does not need to be chunked, that structured data is not required for its generative AI features, and that Google Search ignores llms.txt files. We found no independent controlled study that settles the question for either engine; our comparison of AEO and SEO looks at the guidance in more detail.

Community sources. Reddit's prominence is measured, not assumed: Profound found it accounted for 46.7% of citations among Perplexity's ten most-cited domains between August 2024 and June 2025. It is also volatile; Semrush saw ChatGPT's Reddit citations fall from close to 60% of responses to around 10% within weeks in 2025. The Search Arena study (Miroyan et al., ICLR 2026), built on more than 24,000 conversations, found users tended to prefer answers citing community-driven platforms. We look at the Reddit case in why Reddit shows up in AI answers.

Where the evidence stops

Two limits apply to nearly every finding above. First, answers vary between runs. SparkToro found fewer than one in 100 repeated runs of the same prompt produced the same list of recommended brands, and Ahrefs has reported that about 45% of AI Overview citations change between generations. Single-snapshot studies understate that noise.

Second, a citation is not proof that the source supports the claim. Liu, Zhang and Liang's 2023 audit of four generative search engines found only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported the sentence they were attached to. The Search Arena study found users favored answers with more citations even when the cited content did not support the claims.

The evidence, sorted by strength:

  • Documented by the platforms: retrieval from a search index; crawler-based eligibility and opt-outs; Google's query fan-out.
  • Measured repeatedly: low overlap between engines and with classic rankings; a tilt toward third-party editorial sources; high run-to-run variability.
  • Measured, but contested or engine-specific: freshness effects; the weight of Reddit and other community sites.
  • Still hypothesis: that ranking for fan-out sub-queries drives citation; that formatting or markup changes help outside controlled experiments.

For how these differences play out engine by engine, see our review of published cross-engine comparisons.