How AI Search Chooses Sources to Cite

How do AI search engines decide which sources to cite?

AI search engines cite a source after 2 gates: the page must first be retrieved from the engine's index for the query, then a passage on it must support a specific sentence in the generated answer. Pages that pass only the first gate are read but not cited.

A citation is a link from an AI answer to a web page, shown inline, as a numbered marker or in a source list. A mention is different: the brand or product is named in the answer text, with or without a link. This page is about citations — the links — and the process that decides which pages get them.

That process has two separate decisions. The first is retrieval: for each query the engine runs, it pulls a candidate set of pages from its index. The candidate set, or retrieved set, is every page the model gets to read for that answer. The second is citation: after the model writes, the engine attaches links to the passages that support particular sentences. A passage is a self-contained part of a page — a paragraph, a list, a table — and it is the passage, not the page, that earns the link.

Operators describe both gates in their own words. Google’s AI optimization guide says a page “must be indexed and eligible to be shown in Google Search with a snippet” to appear at all, which is gate 1. Google’s AI features documentation says the answer shows “prominent, clickable links to relevant web pages that support the information in the response”, which is gate 2. OpenAI’s web search tool documentation separates the two in its output: inline citations “show only the most relevant references”, while a sources field “returns the complete list of URLs the model consulted.”

The two-gate structure explains most of what site owners see. A page that never appears in any AI answer usually fails gate 1. A page that shows up in an engine’s consulted sources but never as a citation fails gate 2. The fixes differ, so diagnosing the gate comes first.

Four recent studies measured parts of this process. Each figure below belongs to one study, with its own scope; none is an average across studies.

• arXiv 2604.27790 · April 2026

  51.5% and below 0.2

  Grossman et al. found AI Overviews generated for 51.5% of representative real-user queries, and an average Jaccard similarity below 0.2 between generative engines' sources and Google's traditional results.

• arXiv 2508.00838 · August 2025

  34% and 24%

  Strauss et al. found 34% of Gemini responses and 24% of GPT-4o responses were generated without fetching any online content.

• arXiv 2603.29979 · March 2026

  17.3%

  Yu et al. measured a 17.3% rise in citation rate from structural content features across 6 generative engines.

• arXiv 2604.25707 · April 2026

  21,143 citations

  Zhang, He and Yao analysed 602 prompts on 3 platforms and found ChatGPT cites fewer sources, with higher average influence per fetched page.

What gets a page into the retrieved set?

Crawl access, index eligibility and query relevance get a page into the retrieved set. Google retrieves only indexed, snippet-eligible pages and ranks them with its core Search ranking and quality systems; ChatGPT and Perplexity retrieve pages their own search crawlers fetched.

Gate 1 is the better-documented gate. Every major operator publishes which crawler feeds its search index, and Google publishes the eligibility rule and the ranking basis for its AI features. The table lists only signals an operator states in its own documentation.

The table shows that gate 1 is mostly ordinary search engineering. For Google, a page enters the retrieved set the same way it enters classic results: Googlebot can crawl it, it is indexed, snippets are allowed, and the core ranking systems judge it relevant to one of the queries the model issued. The generative AI setting in Search Console adds one switch on top, so check that the site is included before diagnosing anything else.

For ChatGPT search and Perplexity, the gate is the operator’s own search crawler. OpenAI states that OAI-SearchBot is used to surface websites in ChatGPT’s search features, and Perplexity states the same for PerplexityBot. A robots.txt rule that blocks either one removes the site from that engine’s retrieval pool, while the same rule has no effect on Google. The reverse also holds: blocking Google-Extended, which governs use of content for Gemini models, leaves AI Overviews and AI Mode untouched.

Query relevance is the part of gate 1 that changes with every question. Because engines expand one question into several sub-queries, a page is retrieved when it is relevant to any of them. Freshness enters through the same route: Google’s own definition of grounding asks for “up-to-date” pages, so on time-sensitive questions an older page competes at a disadvantage even when it is otherwise relevant.

What makes a retrieved passage worth citing?

A retrieved passage earns a citation when it directly supports a statement the model writes. Google links pages "that support the information in the response". A 2026 study across 6 generative engines measured a 17.3% citation-rate gain from structural content features.

Gate 2 happens after the model has written. The engine maps each sentence of the answer back to the passages that support it and attaches links to those. Google states the purpose directly: the links are to pages “that support the information in the response.” A passage that states the fact behind a sentence, in words close to the sentence, is the easiest one to map.

Google adds a second property. Its AI features documentation says the models “identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links … than with a classic web search.” A diverse set means a passage that repeats what three other retrieved pages already say has less reason to be linked than one that adds something they lack. This is the logic behind information gain: original facts, first-hand data and specific numbers give the engine a reason to cite one page over its neighbours.

Measured evidence points the same way. Yu et al. (arXiv 2603.29979, March 2026) tested structural content features across 6 generative engines and measured a 17.3% increase in citation rate. The original GEO paper (Aggarwal et al., arXiv 2311.09735) found the effect depends on position within the retrieved set: adding cited sources raised the visibility of the fifth-ranked source by 115.1% and lowered the first-ranked source’s by 30.3%. Both results concern pages already retrieved. They show that wording and structure change gate 2, not that they get a page through gate 1.

Reranking and set selection

Between retrieval and writing sits a reranker, a model that rescores retrieved passages against the query so only the best reach the answer model. The standard design scores each passage on its own; Nogueira and Cho’s BERT reranker (arXiv 1901.04085, 2019) is the common baseline. Lee et al. (arXiv 2507.06838, July 2025) proposed SETR, which selects passages as a set, judging whether together they cover the information the question needs. On multi-hop question-answering benchmarks, set selection outperformed both proprietary and open-source rerankers.

Set selection matters for citation because it rewards complementarity. When passages are chosen as a set, a passage that covers a part of the question the others miss has an advantage over one that duplicates them. No engine names the reranker it uses, so SETR is evidence of the direction research is taking, not a description of any engine.

Why do AI answers leave out pages they read?

AI answers show only the most relevant references, not every page consulted. OpenAI's web search tool returns inline citations for the strongest sources and a separate full list of consulted URLs. Some answers retrieve nothing: one 2025 study found 34% of Gemini responses fetched no content.

OpenAI’s developer documentation is the clearest operator statement that retrieved and cited are different sets. Its web search tool returns inline citations as url_citation objects, each with a url, a title, and a start_index and end_index marking the span of answer text it supports. Separately, the sources field lists every URL the model consulted. The inline set is the subset that earned a link; the rest were read and left uncredited.

Three reasons account for the gap. A consulted page may duplicate a passage from a page that was cited, so crediting both adds nothing. A consulted page may support no specific sentence, because its information was background the model used without stating. And a consulted page may lose to a passage that states the same fact more directly.

There is also a step before any of this: some answers retrieve nothing. Strauss et al. (arXiv 2508.00838, August 2025) found that 34% of Gemini responses and 24% of GPT-4o responses were generated without fetching online content, answered from the model’s training alone. The same study found Gemini gave no clickable citation in 92% of its answers. When no search runs, no page can pass either gate, however well it is built. That makes the question part of the diagnosis: citation work pays off only on the questions where an engine actually searches, so tracking has to record whether each answer shows sources at all.

Do AI engines cite the pages that rank in Google?

Mostly not the same set. A 2026 study of Google Search, Gemini and AI Overviews measured an average Jaccard similarity below 0.2 between sources used by generative engines and Google's traditional results, while AI Overviews appeared for 51.5% of real-user queries.

Jaccard similarity measures the overlap of two sets: the number of items they share divided by the number of distinct items across both. A score of 1 means identical sets; 0 means no shared item. Grossman et al. (arXiv 2604.27790, April 2026) compared the sources behind generative answers with Google’s traditional results for representative real-user queries and found an average below 0.2 — the two sets shared only a small fraction of their pages for the same query.

The study’s scope matters when reading the number. It covered Google Search, Gemini and AI Overviews only; it did not measure ChatGPT search or Perplexity. It reports its own metric, and figures from other studies that use different samples and measures should not be merged with it.

The low overlap is consistent with the mechanism. AI Overviews retrieve from the same Google index as classic results, but they search with fanned-out sub-queries, not only the user’s query, and they cite at passage level. A page ranked on page two for the original question can supply the passage that answers a sub-query. Ranking still matters — the core ranking systems decide what enters the retrieved set — but ranking for the head query and being cited in the answer are measured on different questions.

Compare how Google, ChatGPT and Perplexity pick citations

Google AI Overviews retrieve from the Google Search index, ChatGPT search from pages OAI-SearchBot surfaces, and Perplexity from pages PerplexityBot surfaces. A 2026 study of 21,143 citations found ChatGPT cites fewer sources than other platforms, with higher influence per cited page.

The table shows the documentation is uneven. Google states a ranking basis and a purpose for links, OpenAI states that its citations are a relevance-filtered subset, and Perplexity documents its crawler but not how it chooses among the pages it retrieves.

The ChatGPT finding comes from Zhang, He and Yao (arXiv 2604.25707, April 2026), who ran 602 prompts across 3 platforms and analysed 21,143 citations. They report that “ChatGPT cites fewer sources but shows substantially higher average citation influence among fetched pages.” In practice, a ChatGPT citation is harder to win and worth more to the answer once won.

For Perplexity, where the operator publishes less, the practical route is observation: run the questions that matter, record which pages are numbered as sources, and compare them with the pages that were not. That procedure is set out in the guide to how Perplexity picks its citations.

Does brand recognition decide AI citations?

No operator documents brand recognition as a citation signal. Google's guide lists foundational SEO, non-commodity content, technical structure and local business or ecommerce details. Brand effects on AI answers are measured in observational studies, not confirmed by engines.

Google’s AI optimization guide is organised under headings that name what it recommends: apply foundational SEO best practices, create valuable non-commodity content, build and maintain a clear technical structure, and optimize local business and ecommerce details. None of them is brand fame. The closest item is the last one, where Google says that “using products like Merchant Center (such as Merchant Center feeds) and Google Business Profiles can help your products and services to be visible in both AI responses and other Google Search results.” That is a statement about structured business data Google holds, not about how well known a brand is.

The distinction between citations and mentions matters here. A model can name a well-known brand from what it learned in training without retrieving or linking the brand’s site. Recognition can therefore raise mentions while leaving citations unchanged, and the two should be measured separately.

Signals the engines control are fixed; the question left is which inputs a publisher controls.

Which source-selection signals can a site owner change?

A site owner changes 4 inputs: crawler access, index and snippet eligibility, passage wording that supports specific claims, and original information absent from other pages. Reranking, the answer model and the final citation count stay under the engine's control.

1. Crawler access — allow each engine's search crawler in robots.txt and at the server or CDN: Googlebot, OAI-SearchBot, PerplexityBot and Bingbot. Check that firewalls and bot protection do not block them silently.
2. Index eligibility — confirm each target page is indexed, is not restricted by nosnippet or a tight max-snippet, and that the site is included in generative AI features in Search Console.
3. Claim-level passages — state each answer in a self-contained passage near the top of its section, with the fact, the number and the named entity in the same sentence, so a supporting span is easy to map.
4. Original information — add what the other retrieved pages lack: first-hand data, a measured result, a worked example or a dated source, so the page is not one more copy of the consensus.

The first two inputs act on gate 1 and are binary: a page either passes or it does not, and each can be checked in minutes. The last two act on gate 2 and are comparative: the page competes against every other passage in the retrieved set for the same sentence. Work in that order. Improving wording on a page that no engine can retrieve changes nothing.

What stays outside a site owner’s reach is just as clear. The sub-queries the model writes, the reranker that scores passages, the model that writes the answer and the number of citations shown are all engine decisions, and no on-page change alters them. Those four inputs are the working core of earning AI citations.

What is Grow AI Search testing about citation selection?

Three planned experiments test gate-2 inputs on live AI engines: statistics and quotations, answer-first passages, and structured data. None has started, so no results exist yet, and none is reported on this page.

• Planned Exp 02 — Do statistics and quotations raise citation rate?
• Planned Exp 04 — Answer-first vs story-first: which passage do AI Overviews lift?
• Planned Exp 05 — Does structured data change citation rate?

Each experiment will change one input on a set of live pages, keep matched pages unchanged as controls, and track citations over time with a public log. Until then, the studies above are the evidence. To find out which gate your own pages fail on each engine today, a citation audit runs your buyer prompts and traces every missing citation back to its cause.