Playbook · AI search and GEO

Updated

How do AI assistants choose which sources to cite?

Nobody outside the AI companies knows exactly how AI assistants choose which sources to cite, and none of them publishes the rules. The studies that state a method point the same way: assistants mostly cite pages that already rank in the search index they draw on.

76.1% of pages cited in Google’s AI Overviews rank in Google’s top 10 for the query. From 1.9 million citations in 1 million AI Overviews.
Source: Ahrefs, 21 Jul 2025

What the providers say

Very little. OpenAI says ChatGPT ranks search results using multiple factors meant to find relevant, reliable information, and that placement is not guaranteed. It also says ChatGPT search sometimes sends rewritten queries to other search providers. Google says a page must be indexed and eligible for a snippet to appear in AI Overviews, that the site must be included in Search Console’s generative AI setting, which is the default, and that no special optimization is needed beyond that.

Studies with a stated method

  • Google AI Overviews: 76.1% of cited pages rank in the top 10, 9.5% rank 11 to 100, and 14.4% do not rank in the top 100 (Ahrefs, 1.9 million citations, July 2025).
  • ChatGPT search: over 87% of citations matched Bing’s top 20 organic results and 56% matched Google’s (Seer Interactive, about 100 questions, February 2025).
  • Brand mentions: across 75,000 brands, mentions of the brand on the web correlated with AI Overview visibility at 0.664, and backlink count at 0.218 (Ahrefs, May 2025, Spearman correlation).
  • Page content: on a 10,000-query benchmark with a GPT-3.5 answer engine, adding quotations and statistics raised a source’s visibility, while keyword stuffing gave little or no gain. On Perplexity, the paper reports keyword stuffing doing about 10% worse than doing nothing (Aggarwal et al., KDD 2024).

Why the evidence is thin

  • Three of the four studies come from companies that sell SEO services or tools, Ahrefs included.
  • Samples are small for this question. Seer calls its own sample limited.
  • Correlation is not cause. Ahrefs says so in its brand study.
  • Models and search partners change often. These numbers are from 2024 and 2025 and may not hold today.
  • The GEO paper tested a research setup, not ChatGPT itself.

What I would do with this

Put your effort where the studies agree: rank in normal search first, in Google and in Bing, since that is where the citations come from. Yes, Bing. Then make each page easy to quote and get named on sites people trust. It is dull advice, and it is what the data supports. Treat any claim about one model’s secret ranking factors as a guess until the provider documents it.

The bot rules and steps for ChatGPT and Perplexity are in the guide on getting cited by them.

Read next

Sources

Start

Tell me where the product is stuck.

In 30 minutes we look at where things stand, what you need from a CTO right now, and which option fits. Either way, you leave with a clear next step.

How the call goes

  1. 01You tell me about the product, the team and the next 6 months.
  2. 02I ask about the code, the stack and what keeps breaking.
  3. 03We pick the option that fits, or agree that none does.

09:00 to 20:00 CET

Founding rate: the first 3 clients get 20% off, in return for a named testimonial and a case study. On a monthly plan, the 20% applies to the first month.