How AI search chooses its sources

AI assistants do not rank ten links for one query. They decide whether to search at all, split the question into several searches, merge what comes back, and quote from the passages that fit. Each step changes what off-page work is worth.

One warning before the numbers. Almost all the evidence on this topic is correlational and comes from companies that sell tools in the field, such as Ahrefs, Semrush, Peec AI, Profound and Muck Rack. Treat it as a direction, not a law. The patterns also change quickly.

The pipeline, step by step

1. Decide whether to search. Many answers come from the model’s memory alone, with no citations.
2. Fan out. The prompt is rewritten into several sub-queries.
3. Retrieve. Results come from one or more search indexes.
4. Merge and re-rank. Ahrefs describes a method in which a page ranking around sixth for several variations can beat a page ranking first for one.
5. Select passages, write the answer, attach citations.

This gives you two separate targets. A mention is your brand named in the answer. It depends on the model’s training data and on what the retrieved pages say about you. A citation is your URL linked as a source. It depends on retrieval. Semrush’s 2026 AI Visibility Index found that on Gemini the overlap between mentioned brands and cited domains was as low as 30%, so track them separately.

Query fan-out

Google’s own documentation says AI Overviews and AI Mode “may use a query fan-out technique”, running multiple related searches across subtopics and data sources.

The measured detail comes from vendors:

• Seer Interactive and Nectiv, cited by Ahrefs in March 2026: 59% of prompts trigger 5 to 11 fan-out queries and 24% trigger 12 to 19. The Ahrefs article adds that over 95% of fan-out queries have no recurring human search volume, so keyword tools do not show them.
• Peec AI, 2026, more than 20 million fan-outs: the average ChatGPT fan-out query doubled in length, from about 6 words in October 2025 to about 12 in January 2026. ChatGPT added the word “best” to 24.3% of advice-style questions even when the user had not used it. On non-English prompts, 43% of ChatGPT’s research steps ran on the English-language web.

The added “best” helps explain why third-party comparison lists and review sites are retrieved so often. Analyst readings of Google’s patents go further and suggest AI Mode retrieves passages and personalises results, but Google has not confirmed that interpretation.

Which index each assistant uses

Assistant	Where it retrieves from	Confidence
Google AI Overviews and AI Mode	Google’s own index, with no extra requirements beyond being indexed and eligible for a snippet	High, Google documentation
ChatGPT search	Third-party search providers plus OpenAI’s own crawler, which must be allowed in robots.txt. Bing was the launch partner. The Information reported in August 2025 that Google results obtained through SerpApi were also used; Google sued SerpApi on 19 December 2025	High for the official part, medium for the report
Microsoft Copilot	Bing’s index	Medium
Perplexity	Its own index, offered through a search API since September 2025	High, Perplexity documentation
Claude	Brave Search, according to one analysis of Anthropic’s supplier list	Medium, secondary source
Gemini app	Believed to ground answers in Google Search	Low, not verified

The practical point: you can only be cited by an assistant whose index holds your pages, so being crawlable by each assistant’s search bot, and indexed by Bing as well as Google, is a precondition. Bots that collect training data are separate from bots that power search, so blocking the first kind does not remove you from answers. See links in Bing, Yandex, Baidu and Naver.

How citations differ from rankings

AI Overviews are moving away from the top ten. Ahrefs analysed 863,000 results pages and 4 million cited URLs in March 2026. Of the cited URLs, 37.9% ranked in the top 10 for the same query, 31.2% ranked 11 to 100 and 31.0% ranked beyond 100. In July 2025 the same analysis showed about 76% from the top 10. Ahrefs attributes the shift to heavier fan-out.

Assistants overlap even less. In an Ahrefs study of 15,000 long-tail prompts from August 2025, the share of cited URLs found in Google’s top 10 was 8% for ChatGPT, 8% for Gemini, 8.6% for Copilot and 28.6% for Perplexity.

The same sites, different pages. Ahrefs’ September 2025 study of 3,311 short keywords found only 10% of ChatGPT’s citations in Google’s top 10, but 31.8% overlap at domain level.

The counterpoint. Seer Interactive tested 10,000 questions in January 2025 and found that ranking on Google’s first page correlated at about 0.65 with being mentioned by the model. Seer removed forums, aggregators and social sites from its sample, which inflates the figure.

Put together: ranking still helps, and strong domains recur, but the cited URL is often not the page that ranks for the main keyword. The off-page goal moves from one page and one keyword to being present across many third-party documents that can be retrieved.

Mentions versus links

Ahrefs’ study of 75,000 brands, updated in December 2025, gives these correlations with brand visibility:

Factor	ChatGPT	AI Mode	AI Overviews
YouTube mentions	0.737	0.740	0.712
Branded web mentions	0.664	0.709	0.656
Branded anchors	0.511	0.628	0.527
Branded search volume	0.352	0.466	0.392
Domain Rating	0.266	0.285	0.326

In the earlier May 2025 edition, which covered AI Overviews only, raw backlink count scored 0.218.

Read the table with care:

• It does not show causation. Big brands have more mentions, more searches, more links and more AI visibility for the same reason: they are big. The sample was also limited to sites above Domain Rating 40.
• “Mentions beat links three to one” compares correlation figures, not effects.
• Links are not irrelevant. Links with the brand as anchor are among the strongest factors, and a linked mention is still a mention. Seer found backlinks weak but Google rank strong, and rank depends on links.

The fair conclusion is links plus context, not an end to link building. See brand mentions and citations.

Which sites get cited

Peec AI’s analysis of 30 million citations in the US in March 2026 states that Reddit and YouTube were the only domains in every engine’s top five, although its own published ChatGPT list is Wikipedia, Reddit, Forbes, TechRadar and LinkedIn, with no YouTube. ChatGPT leaned towards editorial sources. Google’s surfaces leaned towards social platforms.

Muck Rack’s May 2026 report claims earned media makes up 84% of citations and paid or advertorial content 0.3%. We took those figures from a press release, Muck Rack sells PR software, and its “earned” category includes Reddit, Wikipedia and reviews.

Two corrections to common claims:

• The famous percentages are misquoted. “Wikipedia is 47.9% of ChatGPT citations” is a share of the top ten sources. Profound’s data on 680 million citations puts Wikipedia at 7.8% of all ChatGPT citations and Reddit at 6.6% on Perplexity.
• The sources rotate. Semrush recorded Reddit falling from about 60% of ChatGPT responses to about 10% within weeks in September 2025, and its citation share dropping from 3.8% to 0.5% in August 2026. The cause of the first fall is unconfirmed.

The limits of the evidence

• The studies are correlational, and most come from vendors.
• No public study isolates review volume as a cause of AI recommendations. That idea is speculative.
• The idea that models learn which brands appear near which category words is well reasoned, but no controlled experiment at scale supports it.
• Traffic is small. SE Ranking’s study of 101,574 sites put AI referrals at 0.32% of all traffic in 2026. Pew found in July 2025 that only 1% of visits with an AI summary led to a click on a link inside it.

Where to go next

For the tactic most often tied to AI citations, read digital PR. For the wider case on unlinked coverage, see brand mentions and citations. Our methodology explains how we treat vendor studies.