How ChatGPT, Perplexity, Gemini & Google AI each choose their sources

The SearchCounsel five-stage source-selection diagnostic

Most explanations of AI citations collapse several separate decisions into one vague idea: “the engine trusted this source.” That is too simple to diagnose why one page appears and another does not. A more useful working model is to separate the process into five stages. This is a Search Counsel diagnostic framework, not a claim about any platform’s private ranking algorithm.

This model explains why “we rank well, so why are we not cited?” is often the wrong question. A page can pass eligibility and retrieval but fail passage fit because the answer is buried. It can support the generated answer but lose the visible citation to a clearer primary source. And a brand can be recommended while a review site, forum thread, or publisher receives the link.

Use an evidence confidence framework, not “AI ranking factor” claims

AI-search advice becomes misleading when platform documentation, third-party correlation studies, one-off observations, and pure speculation are presented as equally certain. Throughout this guide, we use four evidence levels so you can tell what is actually known.

How ChatGPT, Perplexity, Google AI, Gemini and Claude differ

How ChatGPT chooses sources when it searches the web

Documented OpenAI says ChatGPT can search the web automatically when a question would benefit from current information. When search is used, ChatGPT may rewrite the user's prompt into one or more targeted queries, review the results, and issue additional searches. OpenAI also says ChatGPT Search sometimes partners with other search providers; its current help documentation links to Microsoft and Shopify privacy information for those provider relationships.

For publishers, the clearest direct technical control is OAI-SearchBot. OpenAI says public websites can appear in ChatGPT search and recommends allowing OAI-SearchBot, together with OpenAI's published IP ranges, if you want content to be eligible for search-result inclusion. OpenAI separates OAI-SearchBot from GPTBot, so allowing search discovery does not require making the same choice about model-training crawling.

What seems to matter in practice

Observed Third-party studies show that the sources ChatGPT cites vary dramatically by prompt type and research design. A September 2026 CiteGraph sample of 45,309 citations across buyer-oriented questions found heavy use of comparison and roundup pages, while Conductor's seven-month cross-engine study found Wikipedia and Reddit recurring prominently for different ChatGPT intent groups. Those findings are useful directional evidence, not a disclosed OpenAI ranking formula.

For site owners, the practical lesson is to match the information need. If the prompt asks for a definition, give a crisp sourced definition. If it asks for alternatives, a transparent comparison is more useful than a broad brand essay. If it asks for a current fact, show when the fact was last verified and link to the primary source.

How Perplexity chooses sources

Documented Perplexity describes Pro Search as a multi-step web research process. It conducts multiple searches and can draw from articles, academic papers, forums, videos, and other source types depending on the query and selected focus. The answer is then synthesized with direct links so readers can inspect the underlying sources.

This makes Perplexity especially sensitive to the quality of the retrievable evidence around a question. A page does not win simply because it is recent, because it is on Reddit, or because it has a specific format. Those patterns can show up in individual studies, but the stronger durable target is to publish the source that best resolves a sub-question with current, checkable information.

Observed Conductor's September 2025–March 2026 dataset found YouTube leading several Perplexity intent categories during its observation window. Other studies have surfaced different dominant domains or source mixes. That disagreement is a useful warning: platform-level citation percentages should be treated as time-boxed measurements, not permanent rules.

How Google AI Overviews and AI Mode choose sources

Documented Google says AI Overviews and AI Mode are built on its Search systems. Both may use query fan-out, issuing multiple related searches across subtopics and data sources while the response is being developed. Google also says the two surfaces can use different models and techniques, so the links they show can differ even for similar questions.

The most important SEO point is surprisingly ordinary: Google says there are no additional technical requirements for a page to appear as a supporting link in AI Overviews or AI Mode. A page must be indexed, eligible to appear in Search with a snippet, and compliant with Search requirements. Google specifically recommends making content crawlable, internally discoverable, available in textual form, supported by useful images or video when appropriate, and backed by a good page experience.

Does a top-10 Google ranking guarantee an AI Overview citation?

No. It helps establish Search eligibility and relevance, but citation overlap varies sharply by study. Ahrefs' March 2026 analysis of 863,000 keyword SERPs and four million AI Overview URLs found about 37.9% of cited URLs in the first 10 result blocks. BrightEdge, using a different dataset and methodology, reported roughly 17% top-10 overlap across its tracked set. These figures should not be blended into a single trend line because the units and methods differ.

The useful conclusion is narrower: page-one visibility and AI Overview citations are related, but they are not interchangeable measurements. Track both rather than assuming one proves the other.

How Gemini Apps choose and show sources

Gemini Apps should not be treated as the same product as Google AI Overviews. Google Search AI features have their own Search-specific retrieval and eligibility documentation, while Gemini Apps can answer from model knowledge, connected apps, or web-related information depending on the request and product state.

Documented Google says Gemini Apps may show a Sources button or inline related links for some responses. Separately, the Double-check response feature uses Google Search to find content that is similar to or different from claims in Gemini's answer. Google explicitly warns that a page surfaced by Double-check is not necessarily a page Gemini used to generate the response.

How does Gemini decide which brands to cite?

There is no public Google formula that says “brand X gets cited when it reaches score Y.” The safest answer is to separate discoverability from evidence. A brand needs accurate, accessible information on its own site and across credible independent sources; then the specific prompt still determines which pages are relevant enough to appear as sources or related links. For Google Search's AI features specifically, Google says standard Search eligibility and people-first SEO remain the foundation.

That distinction matters for brand monitoring: brand mention rate, visible citation rate, and Google Search ranking are three different metrics. Measure them separately.

How Claude chooses sources when web search is used

Documented Anthropic says that when Claude receives a request that would benefit from current information or specialized knowledge, its web-search capability can generate a targeted search query, retrieve relevant results, analyze the information, and provide citations back to source material. Anthropic also says web-sourced responses include citations so users can verify the information.

This does not support a blanket claim that Claude inherently “rewards bullet points” or always prefers long-form pages. Clean structure is sensible for users and retrieval, but a platform-wide weighting for headings or bullets has not been publicly disclosed by Anthropic. Treat such claims as hypotheses unless a study clearly defines its sample and method.

October 2026 SERP snapshot: what is ranking for this topic now?

Observed To keep this guide grounded in the current search landscape, Search Counsel checked U.S. English Google SERPs on October 7, 2026 for five source-selection query variants using Ubersuggest SERP analysis. This is a point-in-time ranking snapshot, not a universal claim about what Google will show in every country, device, account state, or future date.

The most useful finding is not that one particular domain ranks. It is that the head SERP is not locked exclusively by very large publishers. The current results reward exact-intent specialist explanations, while platform-specific searches often elevate official documentation. That creates a realistic path for this page: be more precise than the specialist pages, more useful than documentation alone, and more transparent about evidence than generic “AI ranking factor” articles.

What 2026 citation studies actually show

The strongest competitive pages in this topic do something the old version of this article did not do consistently: they show their methodology, sample size, observation window, and limits. That matters because AI citation studies often produce different-looking numbers while measuring different things.

The repeated signal across these studies is not “copy the same formatting trick.”

Why credible AI citation studies can disagree

Two serious studies can reach different conclusions without either one being “wrong.” The result depends heavily on what was asked, when it was asked, which product surface was tested, and what the researchers counted as a citation. That is why statements such as “Perplexity prefers Reddit” or “ChatGPT prefers Wikipedia” should always carry a sample and observation window.

This is why this guide separates documented mechanics from measured preferences. A platform's own documentation can establish that query rewriting or web search exists. A third-party study can show that a source type appeared frequently in its sample. Neither, by itself, reveals a permanent private ranking formula.

Source selection also changes with the intent behind the query

Optimizing a page for “AI citations” without considering intent is too broad. The engine still needs a source that fits the sub-question it is trying to answer. The table below is a practical content-design guide, not a claim that every engine applies the same weighting.

For this page, that means every platform section should answer the platform-specific question in its opening sentence, then explain the documented mechanics, then separate observational research from advice. That structure is useful for readers, ordinary search, and any retrieval system looking for a passage that answers one sub-question cleanly.

What to do if you want your content cited more often

1. Fix retrieval blockers first. Keep important pages indexable and available in server-rendered HTML. For ChatGPT Search, review OAI-SearchBot access. For Google's AI features, verify ordinary Search indexability and snippet eligibility. See our AI crawler access guide and machine-first architecture guide.
2. Write for a specific information need. Use a direct answer immediately after the relevant heading, then explain the evidence, limits, examples, and exceptions. Our answer-first content guide shows the structure.
3. Prefer primary evidence over recycled claims. Link to official product documentation, government or academic sources, original datasets, and clearly defined studies whenever possible. Our GEO evidence guide explains why documented, research-supported and merely observed claims should not be mixed together.
4. Add information that a summary cannot replace. Original tests, worked examples, screenshots, decision tables, methodology, downloadable templates, and clear comparisons give readers a reason to visit after the AI has answered the basic question.
5. Keep facts time-stamped. AI products change quickly. A visible update date plus explicit observation windows makes the page safer to cite and easier for readers to evaluate.
6. Build independent corroboration where it is legitimate. Reviews, industry coverage, partner documentation, community discussions, and other third-party references can shape how a brand is represented. Do not manufacture consensus or spam communities. See our guide to third-party mentions and AI citations.
7. Measure per surface. Test the same important prompt set across the engines you care about. Track mentions, visible citations, cited URLs, referral sessions, and conversions separately. Use our AI visibility tracking guide for the measurement layer.

What not to do

• Do not present vendor studies as platform rules. A study can describe its sample; it cannot reveal a private ranking algorithm.
• Do not conflate Gemini Apps with Google AI Overviews. They share Google technology, but their interfaces, sourcing behavior, and documentation are not identical.
• Do not add special “AI schema” because someone promises citations. Google explicitly says no special AI markup is required for AI Overviews or AI Mode.
• Do not optimize only for citations. A citation without a meaningful brand mention, click, or business outcome may have limited value.
• Do not refresh dates without refreshing facts. Recency labels should reflect real editorial review.

A 30-minute source-selection test you can run on your own site

You do not need a giant research platform to start measuring whether AI systems are actually finding and citing your content. A small, consistent prompt set is more useful than occasional screenshots because it lets you separate one-off appearances from repeatable visibility.

1. Choose five prompts that matter commercially. Include one definition, one comparison, one problem/diagnostic prompt, one commercial-investigation prompt, and one current or freshness-sensitive prompt.
2. Run the same prompt set across the surfaces you care about. Keep the country, language, account state, and wording as consistent as practical. Record whether live search or web retrieval appears to be active.
3. Separate mentions from citations. Log whether your brand is named, whether your domain is linked, and which exact URL receives the link. Those are three different outcomes.
4. Classify the source that beat you. Was it the vendor's own documentation, an independent publisher, a review site, community discussion, video, academic source, or another page on your own site?
5. Repeat the test. Re-run the set on a fixed schedule. Citation visibility is volatile; one appearance is not the same as stable selection.

After several runs, calculate three separate rates: mention rate (how often the brand appears), citation rate (how often your domain gets a visible source link), and URL share (which pages on your site earn those citations). Then compare that with ordinary rankings and referral traffic instead of assuming one metric explains the others. Our AI visibility tracking guide covers the measurement layer in more detail.

FAQ: how AI engines choose sources

Sources and methodology

This guide separates platform-documented behavior, third-party observational research, and Search Counsel's point-in-time SERP observations. Platform documentation is used for statements about how a product says it searches, crawls, or displays sources. Third-party studies are used only to describe measured citation patterns in their stated samples. Our October 7, 2026 SERP snapshot is used only to describe what appeared for the tested U.S. English queries on that date.

1. OpenAI Help: Searching the web with ChatGPT — automatic web search, query rewriting, source review, and search-provider information.
2. OpenAI: Overview of OpenAI crawlers — OAI-SearchBot eligibility and crawler controls.
3. Perplexity Help: What is Pro Search? — multiple searches, source types, synthesis, and citations.
4. Google Search Central: AI features and your website — AI Overview/AI Mode eligibility, query fan-out, SEO requirements, and measurement.
5. Google Search Central: Optimizing for generative AI features — current SEO guidance and warning against AI-search hacks.
6. Gemini Apps Help: Sources and Double-check — source links and the distinction between Double-check results and generation sources.
7. Anthropic: Web search on the Anthropic API — targeted queries, retrieval, synthesis, and citations.
8. Conductor: How AI engines choose and cite sources — seven-month cross-engine observational study.
9. PromptScout: How AI answers choose their sources — 5,436 completed answers across 135 tracked prompts, with explicit separation of returned source URLs from displayed citations.
10. Similarweb: How AI chooses sites to cite — analysis based on nearly 600,000 AI citation events.
11. CiteGraph: How ChatGPT chooses its sources — 45,309-citation buyer-intent sample.
12. Ahrefs: 38% of AI Overview citations pull from the top 10 — March 2026 overlap analysis.
13. BrightEdge: AI Overviews at the one-year mark — citation-overlap tracking across industries.
14. 5W: AI Platform Citation Source Index 2026 — consolidated cross-study citation-source synthesis.
15. Search Counsel SERP snapshot, October 7, 2026 — Ubersuggest SERP analysis for five U.S. English source-selection queries. Used only as a point-in-time ranking observation, not as a search-volume estimate or ranking-factor study.