How AI answer engines choose sources

AI answer engines choose sources through a retrieval step separate from the model’s own knowledge, and Google documents that its AI features are rooted in core Search ranking. SiteBienfait audits what can be verified: crawl access, entity clarity and quotable facts. Google Search Central, updated 2025-12-10, states that SEO best practices remain relevant for AI features.

Do answer engines publish how they choose sources?

Partially. Google publishes guidance on how AI features relate to Search, including the retrieval technique it calls query fan-out. Perplexity and Anthropic publish which crawlers fetch pages and what each is used for. None of them publishes a complete ranking specification for AI answers, and none promises that a given page will be used.

That gap is the honest starting point for GEO work. Anything beyond the published documentation is inference, and inference should be labelled as such.

How does Google choose supporting links?

Google Search Central, updated 2025-12-10, explains that AI Overviews and AI Mode may use query fan-out: issuing multiple related searches across subtopics and data sources to build a response. While the answer is generated, models identify more supporting web pages, and the response links to them.

The same page says these features are rooted in core Search ranking. The optimisation guide, updated 2026-07-10, adds that core ranking systems focus on high-quality content while other systems block spam, and that the generative features depend on both. For a site owner, that means the ordinary foundations — crawlable HTML, clear headings, unique facts — are the documented inputs.

What do Perplexity and Anthropic document?

Perplexity’s crawler documentation, updated 2026-10-29, describes PerplexityBot as designed to surface and link websites in search results, and states it is not used to crawl content for AI foundation models. It recommends allowing PerplexityBot in robots.txt to appear in search results. A separate Perplexity-User agent handles user-initiated visits.

Anthropic’s article, last updated 2026-11-04, separates three agents. ClaudeBot collects web content that could contribute to model training. Claude-SearchBot navigates the web to improve search result quality. Claude-User supports user actions when someone asks Claude a question. The distinctions matter: a site that blocks the training crawler has not blocked the search crawler, unless it says so explicitly.

What do the documented paths have in common?

Three things recur across the published material. The page has to be fetchable by the relevant crawler, which makes robots.txt and redirect behaviour part of GEO. The facts have to be in the HTML, not assembled only by client-side script. And the entity has to be described consistently, so a retrieval step can tell which business a sentence is about.

Quotable structure helps too: an answer in the first paragraph, headings phrased as questions, and numbers attributed with a source and a date. None of that is a special AI file format. It is the same discipline that makes a page useful to a person.

What cannot be controlled or promised?

Whether the retrieval step returns the page. Whether the model chooses it over a competitor. Whether the answer includes a link. How often the answer changes between runs. Which engine surface shows the answer at all.

Any claim that a change will produce citations is unfalsifiable from outside the engine. The measurable version is narrower: the page is crawlable, the facts are extractable, and a fixed prompt set records whether citations appear over time.

How should this be measured?

Keep a fixed list of prompts buyers use. On a weekly cadence, record whether the site is named, which URL is cited, which sources appear instead, and the date. Compare months, not single runs, and note engine updates separately from site changes.

SiteBienfait’s GEO route applies this as a change set: crawler access, answer blocks, entity facts and citation checks, with inclusion explicitly not promised. The redirect map checker confirms the URL layer, since every cited URL has to resolve to the page that was cited. The companion guide on what GEO is and what it cannot control sets out the limits in more detail.

Sources

• Google Search Central, AI features and your website, last updated 2025-12-10.
• Google Search Central, Optimizing for generative AI features on Google Search, last updated 2026-07-10.
• Perplexity, Perplexity crawlers, last updated 2026-10-29.
• Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler?, last updated 2026-11-04.