How AI answer engines decide which sources to cite: the evidence

Ask ChatGPT, Perplexity, Copilot, Claude or Google’s AI Mode a question and the answer usually arrives with a few links. Most pages that could have been cited are not. For firms that depend on being found by clients, that gap matters, and it has attracted plenty of confident advice.

The honest position is narrower. The platforms say a lot about their crawlers and little about how they choose sources. Independent research fills some gaps, but much of it is early, and some comes from companies that sell visibility tools.

What a citation is worth

Pew Research Center tracked the browsing of 900 US adults in March 2025. When Google showed an AI summary, users clicked a traditional search result in 8% of visits, against 15% when no summary appeared. They clicked a link inside the summary itself in just 1% of visits[1].

Google says clicks from results pages with AI Overviews are “higher quality”, with visitors more likely to spend more time on the site, though the page gives no figures[2]. A citation is worth having, but it is not a traffic plan on its own.

What the platforms say

Google: ordinary search rules, plus query fan-out

Google’s documentation is unusually direct: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” To be shown as a supporting link, a page must be indexed and eligible to appear in Google Search with a snippet[2].

What differs is how answers are assembled. Both features may use “query fan-out”, issuing multiple related searches across subtopics and data sources, so they can show a wider, more diverse set of links than a classic search. They may also use different models, so their links vary. No new machine-readable files, AI text files or special schema.org markup are needed[2].

Microsoft: passages, not pages

Microsoft is more specific. In AI search, it wrote in October 2025, “ranking still happens, but it’s less about ordering entire pages and more about which pieces of content earn a place in the final answer.” Copilot breaks content into smaller pieces that “can be evaluated for authority and relevance”[3]. The same post concedes there is “no secret strategy for being selected by AI systems”[3].

Since February 2026, Bing Webmaster Tools has offered an AI Performance report in public preview. It shows how often a site is cited across Microsoft Copilot, AI summaries in Bing and select partner integrations, which pages are cited, and the “grounding queries” the AI used to retrieve them[4].

OpenAI, Perplexity and Anthropic: mainly about access

OpenAI, Perplexity and Anthropic document their crawlers clearly, but the pages reviewed here do not explain how they select sources. They do make the precondition clear. Sites that opt out of OpenAI’s OAI-SearchBot “will not be shown in ChatGPT search answers”, though they can still appear as navigational links[5]. PerplexityBot exists “to surface and link websites in search results”[6]. Anthropic warns that blocking Claude-SearchBot “may reduce your site’s visibility and accuracy in user search results”[7].

Anyone who claims to know the exact selection formula for these systems is guessing.

What independent research adds

Ranking well helps less than it did

Ahrefs analysed 863,000 keyword results pages and 4 million URLs cited in AI Overviews. In its March 2026 update, 38% of cited URLs also appeared in the top 10 results for the same query, with the rest split almost evenly between positions 11 to 100 and beyond 100[8].

Its July 2025 study put the figure at about 76%. Ahrefs suggests fan-out queries may matter more since AI Overviews moved to Gemini 3, but it also changed its methodology, so the comparison is indicative, not exact[8].

Third-party coverage carries weight

A September 2025 study compared AI search engines such as ChatGPT, Perplexity and Gemini with Google. It found “a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content”, and that the AI services differ from each other in domain diversity and freshness[9]. It is an arXiv preprint, so treat it as early evidence.

Pew’s figures fit the pattern: Wikipedia, YouTube and Reddit together made up 15% of sources in Google’s AI summaries, and government sites were more common there (6%) than in standard results (2%)[1]. For a business, what reputable third parties publish about you may count as much as your own pages. A well-made announcement is one route to that coverage: see how to write a press release that journalists and AI assistants pick up. For more evidence on why AI answers lean on publications, see why AI answers lean on publications, and how your brand gets into them.

Evidence-rich content does better, at least in the lab

A paper accepted at the KDD 2024 academic conference tested nine ways of rewriting web content for generative engines. The best performers were adding citations from reliable sources, quotations and statistics, which improved visibility by 30% to 40% on one of the authors’ two measures. Keyword stuffing offered “little to no improvement”[10].

The caveats are real: the main experiments used GPT-3.5 Turbo, the Perplexity test used 200 queries with source texts uploaded as files rather than found on the web, and today’s systems are different[10].

What remains unknown

Google does not expose the fan-out queries behind each answer, and AI Overviews are probabilistic, with citations changing from one search to the next[8]. Bing’s report counts citations but not where a source sits in an answer[4]. Nobody outside these companies can see the full process.

How to be a source worth citing

Make authorship and expertise obvious

Google says the same foundations apply to its AI features as to Search, including helpful, reliable, people-first content[2]. Its guidance on that asks whether content shows “clear sourcing, evidence of the expertise involved, background about the author or the site”[11]. It encourages accurate bylines that lead to information about the author, and calls fabricated author profiles a form of deception[11].

No platform says a byline on its own earns an AI citation. The case is indirect, but consistent.

Publish something only you have

Google’s self-assessment begins with: “Does the content provide original information, reporting, research, or analysis?” Its quality guidance values “unique, original information or perspectives that aren’t already available on other websites”[11]. Bing advises supporting claims with examples, data and cited sources[4].

For a broker, wealth manager or professional-services firm, that might be a client survey, anonymised market data, or a sourced explainer on a regulatory change. A page that repeats what everyone else says gives an engine little reason to choose it.

Structure pages so passages stand alone

If engines lift pieces rather than whole pages, each section should make sense by itself. Microsoft recommends descriptive headings, direct question-and-answer pairs, lists and tables, and “sentences that make sense even when pulled out of context.” It warns against long walls of text, answers hidden in tabs or expandable menus, core information kept in PDFs, and key details shown only in images[3]. Bing adds that clear headings, tables and FAQ sections help AI systems reference content accurately[4].

On structured data, the two companies differ in emphasis. Microsoft says schema turns plain text into “structured data that machines can interpret with confidence”[3]. Google says structured data should match the visible text, but there is “no special schema.org structured data that you need to add”[2]. Use schema where it describes the page truthfully, and do not expect it to do the work alone.

Be findable and crawlable

None of this matters if crawlers cannot reach the page. Google lists allowing crawling in robots.txt “and by any CDN or hosting infrastructure”, internal links, and keeping important content in text form[2]. Perplexity notes that sites behind a web application firewall may need to allow its bots explicitly[6]. Security settings can shut out answer engines without anyone in marketing knowing.

On freshness, Bing says regular updates help AI systems reference the current version of a page, and recommends IndexNow to notify participating search engines of changes[4]. Google, for its part, warns against changing dates to make pages seem fresh when the content has not substantially changed[11].

Choose your crawlers deliberately

OpenAI and Anthropic run separate bots for answers and for training, and Perplexity says its bots do not collect training data.

Provider	Answers and search	Fetches on a user’s request	Training
OpenAI	OAI-SearchBot	ChatGPT-User	GPTBot
Anthropic	Claude-SearchBot	Claude-User	ClaudeBot
Perplexity	PerplexityBot	Perplexity-User	None listed; neither bot trains models

OpenAI says each setting is independent, so a site can allow OAI-SearchBot while disallowing GPTBot[5]. User-initiated fetchers behave differently: for ChatGPT-User, robots.txt rules “may not apply”, and Perplexity-User “generally ignores robots.txt rules”[5][6]. Anthropic says its bots honour robots.txt[7].

Google needs the most care. AI Overviews and AI Mode are governed by Googlebot and preview controls such as nosnippet[2]. Google-Extended is a robots.txt token, not a separate crawler. It covers training future Gemini models and grounding in Gemini Apps and on Vertex AI, and does not affect inclusion or ranking in Google Search[12]. Blocking it will not remove you from AI Overviews, but it tells Google not to use your pages to ground answers in the Gemini app.

What to do next

1. Check robots.txt and any CDN or firewall rules so that Googlebot, OAI-SearchBot, Claude-SearchBot, Claude-User and PerplexityBot can reach your public pages, unless you have chosen otherwise.
2. Decide your training policy separately, using GPTBot, ClaudeBot and Google-Extended, remembering that Google-Extended also affects Gemini grounding.
3. List the ten questions clients ask most and make sure each has a page with a descriptive heading, a short direct answer, and supporting detail in HTML rather than a PDF or image.
4. Add real bylines linked to author pages, and source the factual claims on your key pages.
5. Plan regular original data or analysis that others in your sector would want to reference.
6. Verify the site in Google Search Console and Bing Webmaster Tools, and use Bing’s AI Performance report to see which pages are cited and for which grounding queries.