How AI assistants choose which sources to cite: what we know and what is guesswork

When an AI assistant answers a question and lists sources, how did it pick them? The honest answer is a mix: a few things are documented, some patterns show up in studies, and a lot is simply not public. This post sorts what we found into three piles, known, likely and unknown, and gives the evidence for each statement so you can judge it yourself.

What it means for an assistant to pick a source

An assistant can answer in two ways. It can answer from what the model learned in training, with no sources. Or it can search the web at the moment you ask and build the answer from what it finds. Citations come from the second way.

The company documents we read describe that second way in similar terms. Anthropic's documentation for Claude's web search tool says Claude determines when to search based on the prompt, that the searches can repeat several times in one request, and that the final response includes cited sources. It says Claude searches for current or changing information, such as recent events, prices and details about specific organizations, and answers directly for stable knowledge. Google describes AI Overviews and AI Mode as using the core Search ranking systems to retrieve pages, plus "query fan-out," which means issuing several related searches. OpenAI and Perplexity each run separate crawlers for search and for training, which we cover below.

1. 1You askA question about the world
2. 2DecideThe assistant decides whether to search
3. 3SearchIt runs one or more searches
4. 4ReadIt reads the pages that come back
5. 5AnswerIt writes a reply and cites some of them

That distinction matters for everything below. Being part of a model's training data and being cited in a live answer are different things. We only deal with the second.

Known: the gate is access, not a secret ranking

These statements come straight from the companies' own documentation.

• Google: a page must be indexed and eligible to be shown in Google Search with a snippet to be a supporting link in AI Overviews or AI Mode, and crawling must be allowed in robots.txt and by any CDN or hosting layer (Google Search Central, updated December 10, 2025).
• OpenAI: OAI-SearchBot is used to surface websites in ChatGPT's search features. A site that disallows it won't be shown in ChatGPT search answers, though it can still appear as a navigational link, and OpenAI says changes to robots.txt can take about 24 hours to take effect. GPTBot is a separate crawler for training, and ChatGPT-User visits a page when a user asks, so robots.txt rules may not apply to it.
• Perplexity: PerplexityBot is designed to surface and link websites in search results, and Perplexity says it isn't used to crawl content for AI foundation models. To be visible, a site should allow it in robots.txt.
• Anthropic: developers who use the web search tool can set an allow list or a block list of domains, and an organization's administrator can restrict which domains it searches. So the source pool can be filtered by whoever deploys the assistant.

What this tells you: blocking the right crawler can keep you out. It does not tell you that allowing one gets you in. Access is a necessary condition, not a ranking factor.

Known: citations and Google rankings overlap less than you'd expect

If an assistant searched the web and simply took the top results, its sources would match Google's top 10. Measured data says they often don't.

Ahrefs studied about 4 million URLs cited in AI Overviews across 863,000 keywords (data published March 2026). Only 37.9% of the cited URLs also ranked in the top 10 for the same search. Another 31.2% ranked 11 to 100, and 31.0% ranked below 100. In an earlier study from July 2025, the figure was about 76%.

For other assistants, Ahrefs ran 15,000 long-tail queries in Google and Bing and then asked the same questions of AI assistants (data collected in early July 2025, published August 11, 2025). The share of cited links that were also in Google's top 10 was 28.6% for Perplexity, 8.6% for Gemini, 8.2% for Copilot and 8.0% for ChatGPT's in-text citations. (We covered the averaged figure in the future of link building.)

Google's two AI features also differ from each other. In a September 2025 sample of 540,000 query pairs, Ahrefs found that only 13.7% of citations overlapped between AI Overviews and AI Mode. Yet the answers agreed in meaning about nine times out of ten. They said much the same thing and cited different pages.

Why? Google confirms that query fan-out exists. Ahrefs's explanation of the low overlap is that the pages appearing most often across the sub-searches get cited. That explanation is a reasonable inference, and Ahrefs hedges it, noting that fan-out "may" be playing a bigger role since a model update. It isn't something Google has measured for us. Also keep in mind that Ahrefs sells AI-visibility tracking, and all these studies use its own dataset.

Known: the same question gives different sources each time

Three measurements point the same way.

Ahrefs followed more than 43,000 keywords, each with at least 16 recorded AI Overviews, over one month. The content changed between consecutive checks about 70% of the time, and on average an answer stayed stable for about 2.15 days. Between consecutive responses to the same query, 45.5% of the cited sources were new. The meaning barely moved (a similarity score of 0.95 out of 1), so the answer stayed the same while the words and sources shifted.

SparkToro and Gumshoe tested ChatGPT, Claude and Google AI Overviews with 12 prompts, 600 volunteers and 2,961 runs (published January 28, 2026). When asked for brand or product recommendations, there was less than a 1 in 100 chance of getting the same list of brands twice. The same list in the same order was rarer still, about 1 in 1,000 by their estimate. But some brands came up very often: a top cancer hospital appeared in 69 of 71 answers (97%), while other brands appeared in roughly 30% to 40% of answers in their categories.

Likely: signs that point one way but don't prove it

Here we have data, but only correlations, small samples or lab settings. Treat each as a lead.

Brands that are mentioned more on the web are cited more. In December 2025, Ahrefs studied 75,000 domains with a Domain Rating above 40, using each domain's highest-volume keyword (at least 800 monthly searches). Using Spearman correlations, branded web mentions scored 0.664 for ChatGPT, 0.709 for AI Mode and 0.656 for AI Overviews. Domain Rating scored 0.266, 0.285 and 0.326, and Ahrefs called the correlation with the raw number of backlinks very weak. Ahrefs says correlation isn't causation. Well-known brands are mentioned a lot and also cited a lot, so a hidden third factor, simply being big, could explain both. We go deeper in brand mentions vs backlinks.

Assistants cite somewhat newer pages than Google's results do. Ahrefs analyzed 16.975 million cited URLs (published July 28, 2025). Pages cited by AI assistants averaged 1,064 days since publication against 1,432 days for organic results, about 25.7% fresher. ChatGPT's citations averaged 958 days, and AI Overviews matched organic results at 1,432. The authors point out that the average cited page was still 2.9 years old, so this doesn't show that updating a page earns citations.

A few large sites take a big share of citations. Pew found that Wikipedia, YouTube and Reddit made up 15% of the sources listed in the AI summaries it examined. In an Ahrefs analysis of 78.6 million searches in June 2025, Wikipedia accounted for 16.3% of ChatGPT's citations, 12.5% of Perplexity's and 8.4% of AI Overviews. The shares differ a lot by platform: YouTube was 16.1% for Perplexity and 9.5% for AI Overviews and not in ChatGPT's top 10. Ahrefs weighted citations by keyword search volume, because no volume data exists for assistants.

Pages with quotes, statistics and cited sources may be used more. A KDD 2024 research paper tested nine writing changes on a benchmark of 10,000 queries and reported that its best methods could raise visibility in generated answers by up to 40%. Adding quotations did best in their tests, and stuffing in keywords, the old SEO habit, did poorly. This was a lab test run in 2023 and 2024 on the authors’ own test setup and on a 200-query sample from Perplexity. The authors also note that engines change.

Microsoft's advice. In its February 10, 2026 announcement of AI Performance in Bing Webmaster Tools, Microsoft recommends strengthening depth and expertise, improving structure and clarity, supporting claims with evidence, keeping content fresh and accurate, and keeping text, images and video consistent. That's the company's guidance, not a measured result.

Unknown: what nobody outside the labs can tell you

These are the questions we couldn't answer from any source we could open.

• Whether a backlink or a mention causes a citation. We found no controlled test where one thing was changed and the citations were watched. A June 2025 Ahrefs test compared a link-based domain score with the sites mentioned most in AI answers and found -0.12 for AI Overviews, 0.01 for ChatGPT and -0.34 for Perplexity. The author said the sample, the top 50 sites per system, was too small to be reliable. So "links don't matter for AI" is also unproven.
• How each assistant weighs the pages it finds. Google says its AI features are rooted in its core ranking and quality systems. OpenAI and Perplexity documentation describes which crawler surfaces sites, not how results are ordered. Anthropic's documentation describes when Claude searches and how citations are returned, not how results are ranked. None of them publishes weights.
• Whether being cited sends visitors. Pew found that people clicked a link inside an AI summary in 1% of visits to pages that had one. Seer found brands cited in an AI Overview had a higher organic click-through rate, 0.70% against 0.52%, but cited brands may simply be stronger. Google's new report shows impressions only, with no clicks or position. Microsoft says its report doesn't show placement, ranking, authority or the role of a page within an answer.
• What will still hold in a year. Ahrefs's own figure for AI Overviews fell from about 76% to 37.9% in under a year. Rules that look solid today can change quickly.
• Whether special files or markup help non-Google assistants. Google says llms.txt and similar files don't help or harm its Search. We found no official statement from other assistants. We look at this in what llms.txt does.

Our view: Anyone who tells you exactly how an assistant ranks sources is guessing or selling. The honest position is that access is documented, instability is measured, correlations exist, and cause is unknown.

How strong is each kind of evidence?

Not every source in this post deserves the same weight. This is the order we use.

From strongest to weakest evidence

1. Company documentationWhat the company says its system does
2. Large measured studiesShow what happens, not why
3. CorrelationsTwo things move together
4. Small or lab testsLimited samples or controlled setups
5. Unsupported claimsNo method, no numbers

Company documentation is the strongest evidence of how a system works, but it covers only what the company chooses to publish. Large measured studies, like the Ahrefs and SparkToro samples, tell you what happens in their data. Correlations show association, and the Ahrefs December 2025 numbers sit here. Small samples and lab setups, like the 2023 research paper, show what's possible. And a claim with no method behind it is a guess.

How to test a claim about AI citations

Take a claim that's easy to find: "More web mentions means more visibility across every AI platform we studied." Ask five questions before you act on it.

Five questions for any AI citation claim

• Who measured it?And what do they sell
• What was counted?Citations, mentions or clicks
• Who was in the sample?Big brands or everyone
• What was compared?Over time, or across sites
• Did anyone change one thing?If not, it's correlation

Here's how that claim holds up. Who measured it? Ahrefs, which sells a tool that tracks AI visibility. That doesn't make the study wrong, but it's good to know. What was counted? Mentions in blog posts, anchors, video transcripts, descriptions and titles, and also YouTube mentions where the brand appeared in the video title. Who was in the sample? Domains with a Domain Rating above 40, using one high-volume keyword each. What was compared? Brands against each other, at one point in time. Did anyone change one thing and watch? No.

Our view: the data fit the idea that bigger, better-known brands get mentioned more and cited more. They don't show that adding mentions to a small brand would lift its citations. The claim, as worded, goes further than the evidence.

What to take away

• Make your pages reachable. If you want to be eligible, make sure the relevant crawlers aren't blocked and Google can index the page with a snippet. That's the one step that's documented.
• Treat any single check as a sample. Run the same prompts several times and look at how often you appear, not whether you appeared once. Our guide to checking AI mentions yourself shows how.
• Don't pay for a ranking-factor list. Nobody outside the labs has one. See Google's own words on fake mentions in AI search vs SEO: what is changing and what is not.
• Invest in what holds up under every outcome. Pages worth citing, and real coverage from sites your readers use. For how buyers actually use AI when choosing software, see where software buyers look, and for expert-quote placements, expert quotes and AI answers.
• Be careful with promises. We don't sell AI citations, and no one can honestly guarantee them. What we do is arrange real placements on real sites that you approve first, through guest posting, niche edits or blogger outreach. For sensitive topics, our guide to YMYL and E-E-A-T explains what we check.

Sources

• Google Search Central: AI features and your website (updated December 10, 2025)
• Google Search Central: Optimizing your website for generative AI features on Google Search (updated July 10, 2026)
• Google Search Console Help: Generative AI performance report (Search) (impressions only; no clicks or position)
• OpenAI: Overview of OpenAI crawlers (OAI-SearchBot, GPTBot, ChatGPT-User)
• Perplexity: Crawlers (PerplexityBot, Perplexity-User)
• Anthropic: Web search tool (documentation)
• Bing Webmaster Blog: Introducing AI Performance in Bing Webmaster Tools (February 10, 2026)
• Ahrefs: Update: AI Overview citations and the top 10 (March 2026; 863,000 keywords)
• Ahrefs: Only 12% of AI cited URLs rank in Google's top 10 (August 11, 2025; 15,000 queries)
• Ahrefs: Are AI Mode and AI Overviews just different versions of the same answer? (September 2025 data; 540,000 query pairs)
• Ahrefs: AI Overviews change every 2 days (43,000+ keywords, one month)
• SparkToro and Gumshoe: AIs are highly inconsistent when recommending brands (January 28, 2026; 2,961 runs)
• Ahrefs: Brand visibility and AI correlations (December 12, 2025; 75,000 domains)
• Ahrefs: Domain-level link metrics may not be good predictors of AI search mentions (June 25, 2025)
• Ahrefs: Do AI assistants prefer to cite fresh content? (July 28, 2025; 16.975 million URLs)
• Ahrefs: The 10 most mentioned domains for ChatGPT, Perplexity and AI Overviews (June 2025 data; 78.6 million searches)
• Pew Research Center: Google users are less likely to click on links when an AI summary appears (July 22, 2025)
• Seer Interactive: AIO impact on Google CTR, September 2025 update (cited vs not cited click-through rates)
• Aggarwal et al.: GEO: Generative Engine Optimization (KDD 2024)

Checked on 6 October 2026.