# How ChatGPT, Gemini and Google choose which sources and brands to cite

Rodrigo Beckmann · Published 2026-11-13 · Category: [GEO](https://rodrigobeckmann.com/en/blog/category/geo/) · HTML version: https://rodrigobeckmann.com/en/blog/how-ai-chooses-sources/

An AI with search runs several searches, retrieves pages and picks which ones to cite. Why each engine picks different sources, what query fan-out is, and why ranking on Google doesn't guarantee the AI Overview.

## In short

- An AI with search picks sources in two steps: it searches the web and retrieves pages, then decides which ones to use and cite. Google says AI Overviews and AI Mode may fire several related searches at once, a technique called query fan-out.
- Because of fan-out, ranking for the keyword doesn't guarantee the citation: according to Ahrefs, in March 2026 only 38% of pages cited in AI Overviews ranked in Google's top 10 for the same query, down from about 76% in July 2025.
- Each engine chooses differently: in a Kevin Indig study of 20,000 prompts, published in May 2026, 91% of cited URLs appeared in only one of three engines, and just 2.37% appeared in all three.

An AI with search picks the sources for an answer in two steps: first it searches the web and retrieves pages, then it decides which ones to use and which ones to cite. On Google, the first step is usually more than one search: [AI Overviews and AI Mode may use a "query fan-out" technique](https://developers.google.com/search/docs/appearance/ai-features) that fires several related searches at once. Brands get into the answer along a similar path, and, according to the available studies, what third parties say about a brand weighs more than the brand's own site.

So there is no single list of sources that "AI" prefers. Each engine searches, filters and cites in its own way, and the choice shifts over time and across languages. This guide pulls together what the platforms document and what public studies have measured, with data from Brazil where it exists. A disclosure first: I'm co-founder and CTO of First Answer, which publishes the weekly indices cited here.

## How an AI picks the sources for an answer

Google describes the process in its [guide to optimizing for generative AI features](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide), last updated on July 10, 2026. Search's AI features use retrieval-augmented generation (RAG, also called grounding): Search's ranking systems retrieve relevant pages from the index, the model reads the relevant part of each one, and writes the answer with links to the pages that support it.

Reading a page is not the same as citing it. A model can use a piece of information without crediting the source, and citations are only the subset of inputs that get attribution. That was the point of [why citations are the new backlinks](https://rodrigobeckmann.com/en/blog/why-citations-matter): a link moved a page up the ranking, a citation puts the source inside the answer.

OpenAI documents less. On its [crawlers page](https://developers.openai.com/api/docs/bots), it says OAI-SearchBot is the crawler used to surface websites in ChatGPT's search answers, and that sites that opt out of it won't be shown in those answers (they can still appear as navigational links). GPTBot is a separate crawler, which collects content to train models, and the two settings are independent. How ChatGPT decides which pages to cite, OpenAI doesn't explain. The Gemini app has no equivalent documentation either: Google's guides only cover AI Overviews and AI Mode.

## What query fan-out is, and why it changes the game

Query fan-out is the technique in which the model generates several related searches, run at the same time, to gather more results before answering. The example is Google's: for "how to fix a lawn that's full of weeds", fan-out queries might include "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". According to Google, this lets it show a wider and more diverse set of links than a classic search.

In practice, the cited page doesn't need to rank for the original question. It needs to rank for one of the derived ones. In March 2026, [Ahrefs](https://ahrefs.com/blog/ai-overview-citations-top-10) matched 4 million URLs cited in AI Overviews, across 863,000 queries, against the organic results for the same query:

| Organic position for the original query | % of pages cited in AI Overviews |
|---|---|
| Top 10 | 37.1% |
| 11 to 100 | 26.2% |
| Outside the top 100 | 36.7% |

In Ahrefs' previous study, from July 2025, about 76% of cited pages came from the top 10. The drop to 38% comes with a caveat Ahrefs itself records: between the two studies, it improved how it captures citations, and AI Overviews moved to Gemini 3. Ahrefs reads the change as fan-out playing a bigger role in source selection. It's a plausible hypothesis, not a confirmation from Google.

Google also warns what not to do with this. The guide says that creating separate content for every variation of a query, fan-out queries included, mainly to manipulate AI answers violates its scaled content abuse spam policy. The way in is covering the topic well, not multiplying pages.

## Why do I rank on Google but not show up in the AI Overview?

There are three documented reasons, and none of them is a missing special optimization.

- **AI Overviews don't always show.** According to Google, they only appear when its systems decide they add something to classic Search, and as such they "often don't trigger".
- **The citation comes from other searches.** Through fan-out, the answer can cite pages that rank for derived questions. In the Ahrefs dataset, more than a third of cited pages weren't in the top 100 for the original query, and 18.2% of those were YouTube videos.
- **AI Overviews and AI Mode are different systems.** Google says they may use different models and techniques, so the answers and links will vary.

The technical requirement is the same as for regular Search: to be a supporting link in an AI answer, a page must be indexed and eligible to appear in Search with a snippet. Google says there are no additional requirements, no special structured data, and that meeting everything doesn't guarantee the page will be shown. Using `nosnippet` or limiting the snippet also limits what can appear in an AI answer.

## Which websites does AI cite the most?

It depends on the engine and on the week. The weekly indices [Reddit in AI search](https://rodrigobeckmann.com/en/research/indices/reddit-in-ai-search/) and [YouTube in AI search](https://rodrigobeckmann.com/en/research/indices/youtube-in-ai-search/) track, by model, the share of answers that cite each domain among their sources. The numbers for week 45, November 2–8, 2026:

| Model | Answers citing Reddit | Answers citing YouTube |
|---|---|---|
| Google AI Overviews | 22.5% | 61.7% |
| Google AI Mode | 12.9% | 46.3% |
| Grok | 2.5% | 18.0% |
| Perplexity | 3.8% | 2.2% |
| Gemini | 2.4% | 1.6% |
| ChatGPT | 0.1% | 0.03% |
| Copilot | 0.03% | 0.2% |

*Source: data observed and catalogued by First Answer. The base doesn't represent every search made on the models; methodology and limits are on each index page.*

Two things stand out. First, Google AI Overviews and the Gemini app belong to the same company and pick sources very differently. Second, in week 45 (November 2–8), ChatGPT barely cited either domain.

It wasn't always like this. In week 1 of the index (December 29 to January 4), Reddit appeared in 21.1% of ChatGPT answers. In August it fell to 7.7% and then 2.1%, and it has stayed under 1% since week 35. Promptwatch, as reported by [Search Engine Journal](https://www.searchenginejournal.com/why-reddits-chatgpt-citation-drop-isnt-fully-explained/586479/), recorded the same drop in ChatGPT citations, in two steps, on August 8 and 14. According to SEJ, the cause is not yet established.

It wasn't the first time either. In a study of 230,000 prompts between July and October 2025, [Semrush](https://www.semrush.com/blog/most-cited-domains-ai/) saw ChatGPT cite Reddit in close to 60% of answers in early August 2025 and in around 10% by mid-September. Over the same period, AI Mode kept a steadier mix, with LinkedIn, YouTube, Reddit and Google's own pages at the top.

Brazil has less public data, and it doesn't look like the US. On August 17, 2026, [Profound](https://www.tryprofound.com/reports-guides/comportamento-de-prompts-e-citacoes-no-brasil) reported that over the previous three months Reddit accounted for 9.0% of ChatGPT's citations in Brazil, against 2–3% in the US, and Wikipedia for 1.8%, against 7.8% in the US. The page doesn't describe the sample. The report's three months run to mid-August; in the index, Reddit's presence in ChatGPT has been under 1% since week 35 (August 24–30). The two measures aren't the same thing, either: one is the share of citations, the other the share of answers.

## How does ChatGPT decide which brands to recommend?

OpenAI doesn't publish brand recommendation criteria. What exists are correlation studies, and they point the same way: what the web says about a brand counts more than what the brand publishes.

In December 2025, [Ahrefs](https://ahrefs.com/blog/ai-brand-visibility-correlations) matched 75,000 brands against millions of answers from ChatGPT, AI Mode and AI Overviews:

- **YouTube mentions** had the highest correlation with brand presence across the three AIs (about 0.737).
- **Branded web mentions** came right after, between 0.66 and 0.71.
- **Number of site pages** showed almost no relationship (about 0.194), and backlinks also correlated weakly.
- **In ChatGPT**, classic authority metrics mattered less: 0.266 for Domain Rating and 0.352 for branded search volume. AI Mode tracked established-brand signals most closely.

Ahrefs itself warns that this is correlation, not causation. Improving one of these metrics doesn't guarantee showing up.

A University of Toronto study by [Chen et al.](https://arxiv.org/abs/2509.08919) (September 2025) reached the same reading another way. AI search engines showed a "systematic and overwhelming bias" toward earned media (press, reviews, independent sites) over brand-owned and social content. In US car queries, for example, 81.9% of the sources of web-enabled GPT were earned media, against 45.1% on Google. The authors also describe a big-brand bias that makes life harder for niche brands. The Toronto study also appears in the [guide to what GEO is](https://rodrigobeckmann.com/en/blog/what-is-geo).

Your own content still counts, for what it adds. Google's guide asks for "non-commodity" content, with a unique point of view and first-hand experience, rather than a summary of what's already everywhere. Dates matter too: in an analysis of 17 million citations from July 2025, [Ahrefs](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/) found that pages cited by AI assistants were on average 25.7% fresher than those in organic results. The exception was AI Overviews, which cited slightly older pages than organic search.

## One brand, many engines

Doing well in one engine says almost nothing about the others. In May 2026, [Kevin Indig](https://www.growth-memo.com/p/the-consensus-gap) analyzed 20,000 prompts across ChatGPT, Perplexity and Google AI Overviews: 91.07% of cited URLs appeared in only one of the three engines, and only 2.37% appeared in all three for the same prompt. In an intent breakdown drawn from roughly 2.6 million URLs across the study's combined sample, even commercial prompts, which should converge more, reached just 2.4% overlap across the three engines. The sample is European and Spain-heavy, and Indig asks readers to treat the numbers as directional.

Language splits things further. In the Toronto study, the same questions translated into five languages made ChatGPT cite sets of sites with near-zero overlap across languages. According to the authors, it effectively switches site ecosystems by language, more than Google does. Portuguese wasn't among the languages tested, but the implication for anyone working outside English is direct: what a brand has earned in English doesn't carry over by itself to answers in another language.

## Why the studies disagree

Read two reports on AI sources and you'll often find numbers that don't match. Before deciding one is wrong, check four things:

- **The measure.** Share of citations (Profound) is not the same as share of answers citing a domain (Semrush, the weekly indices).
- **The period.** ChatGPT cut Reddit's weight twice in a little over a year. A number from August may not hold in November.
- **The sample.** Country, language and type of question change the result, as the Indig and Toronto studies show.
- **The method.** Ahrefs itself changed how it captures citations between its 2025 and 2026 studies.

It's the same problem as measuring brands, which I covered in the [guide to measuring AI visibility](https://rodrigobeckmann.com/en/blog/how-to-measure-ai-visibility): without a stated sample, environment and period, one "AI sources" number is hard to compare with another. And answers change on every run, for the reasons in [why AI gives a different answer every time](https://rodrigobeckmann.com/en/blog/why-ai-answers-change).

## What to do

1. **Make sure you're eligible.** Indexed pages with snippets allowed on Google; OAI-SearchBot allowed in `robots.txt` for ChatGPT. Blocking GPTBot is a separate decision, about training.
2. **Cover the question, not just the keyword.** Fan-out searches a question's subtopics. A page that answers the follow-ups well has a better chance of being retrieved, as long as that doesn't turn into one page per variation.
3. **Work off your site.** Press, independent reviews, videos and communities are what the studies associate with brand presence. According to [Ahrefs](https://ahrefs.com/blog/ai-overview-citations-top-10), as of March 2026 YouTube was the most cited domain in AI Overviews.
4. **Publish what only you have.** Your own data, first-hand experience, analysis with a position. Keep dates visible and updates real.
5. **Measure by engine and by language.** A blended "AI visibility" average hides that a brand can be strong in one engine and absent in another.

## What this data doesn't show

- **Causation.** The brand studies measure correlation. No one outside knows how much each signal weighs.
- **ChatGPT and the Gemini app from the inside.** OpenAI and Google don't document how these two pick sources; Google only documents AI Overviews and AI Mode.
- **Brazil in detail.** The Ahrefs, Semrush, Indig and Toronto studies don't cover Portuguese. The most cited Brazilian figure, from Profound, doesn't publish its methodology.
- **The long run.** Sources that dominated an AI in January vanished by August. Any "most cited sites" list has an expiry date.

## The choice in one sentence

An AI doesn't cite whoever ranks best for a keyword: it cites, among the pages its many searches find, the ones that support the answer, and each engine makes that choice with its own rules and sources, which change from one month to the next.

## Frequently asked questions

### How does ChatGPT decide which brands to recommend?

OpenAI doesn't publish the criteria. The available studies point to what third parties say about the brand: in an Ahrefs study of 75,000 brands, from December 2025, brand mentions on YouTube and across the web had the highest correlations with presence in ChatGPT, AI Mode and AI Overviews, while the number of pages on the brand's site showed almost no relationship. That is correlation, not causation.

### What is query fan-out?

It's the technique in which the AI turns one question into several related searches, run at the same time, to gather more results before answering. In Google's own example, a question about fixing a lawn full of weeds produces searches such as "best herbicides for lawns" and "how to prevent weeds in lawn".

### Why do I rank on Google but not show up in the AI Overview?

Because AI Overviews don't only cite pages that rank for the original query. They run related searches (fan-out) and can cite pages that rank for those. According to Ahrefs, in March 2026, 37.1% of cited pages ranked in the organic top 10 for the original query and 36.7% weren't even in the top 100. AI Overviews also only show when Google decides they add something to the search.

### Which websites does AI cite the most?

It depends on the engine and on the week. In the week of November 2–8, 2026, YouTube appeared in 61.7% of Google AI Overviews answers and in 0.03% of ChatGPT answers. In an August 2026 report, Profound put Reddit at 9.0% of ChatGPT's citations in Brazil, against 2–3% in the US.

### Does blocking GPTBot remove my site from ChatGPT?

No. According to OpenAI, GPTBot collects content to train models, and OAI-SearchBot is the crawler that surfaces sites in ChatGPT's search answers. The two settings are independent: you can block GPTBot and allow OAI-SearchBot. Sites that block OAI-SearchBot won't be shown in ChatGPT's search answers.

## Sources

1. [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features) — Google Search Central, 10 Dec 2025 · primary source
2. [Google's Guide to Optimizing for Generative AI Features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) — Google Search Central, 10 Jul 2026
3. [Overview of OpenAI Crawlers](https://developers.openai.com/api/docs/bots) — OpenAI
4. [Update: 38% of AI Overview Citations Pull From The Top 10](https://ahrefs.com/blog/ai-overview-citations-top-10) — Ahrefs, 02 Mar 2026
5. [Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)](https://ahrefs.com/blog/ai-brand-visibility-correlations) — Ahrefs, 12 Dec 2025
6. [The Consensus Gap](https://www.growth-memo.com/p/the-consensus-gap) — Growth Memo (Kevin Indig), 11 May 2026
7. [Generative Engine Optimization: How to Dominate AI Search](https://arxiv.org/abs/2509.08919) — University of Toronto (arXiv), 10 Sep 2025
8. [The Most-Cited Domains in AI: A 3-Month Study](https://www.semrush.com/blog/most-cited-domains-ai/) — Semrush, 10 Nov 2025
9. [Comportamento de prompts e citações no Brasil](https://www.tryprofound.com/reports-guides/comportamento-de-prompts-e-citacoes-no-brasil) — Profound, 17 Aug 2026
10. [Why Reddit's ChatGPT Citation Drop Isn't Fully Explained](https://www.searchenginejournal.com/why-reddits-chatgpt-citation-drop-isnt-fully-explained/586479/) — Search Engine Journal, 19 Aug 2026
11. [New Study: AI Assistants Prefer to Cite "Fresher" Content (17 Million Citations Analyzed)](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/) — Ahrefs, 28 Jul 2025

**Rodrigo Beckmann** — GEO specialist, co-founder and CTO of First Answer. [More about the author](https://rodrigobeckmann.com/en/about/)