# How AI Search Engines Select Sources, and How Publishers Get In *Most pages cited by AI Overviews no longer rank in Google's top 10. Here is how Google, Perplexity and ChatGPT actually pick their sources, what gets a publisher into the pool, and why being there is the new distribution contract.* Published: 2026-10-12 | Read time: 5 min read | Author: Smalk AI Research | Source: https://www.smalk.ai/blog/how-ai-search-engines-select-sources-publishers-get-in --- The page an AI engine cites from your site is increasingly one that never ranked. Ahrefs' March 2026 study found that 62% of AI Overview citations come from pages outside Google's top 10, and 31% from beyond position 100. For a publisher, that changes the job. For twenty years the contract was crawl, rank, get the session. The new contract is be retrieved, be cited, and, so far, get nothing. Understanding how AI search engines select sources is now the publisher's distribution strategy, and it is the precondition for ever being paid as one. - AI engines select sources by retrieving passages from a wide pool and crediting whichever page answers a sub-question best, not by reading the top of the SERP. - Google AI Overviews, Perplexity and ChatGPT each run their own selection pipeline, with top-10 overlap ranging from 37.9% to around 12%. - Publishers get into the pool with dated original evidence, full sub-question coverage and crawl access for AI Search bots, not with formatting tricks. - Being in the sources is the new distribution contract: it drives the answer, rarely drives the click, and has no payment clause yet. - The publishers who map and sell their retrieved pages will own the inventory the next ad market runs on. ## What AI source selection is AI source selection is the process by which a generative engine decides which pages to retrieve, ground its answer in, and credit with a citation. Google AI Overviews, Perplexity and ChatGPT all use a form of retrieval-augmented generation: the engine breaks a query into related sub-questions, pulls candidate passages from its index or search partner, writes an answer grounded in that text, and cites the pages each passage came from. The unit the engine evaluates is the passage, not the page's rank for the original keyword. ## The crawl-and-rank contract vs the retrieve-and-cite contract Under the crawl-and-rank contract, a publisher knew how to be found and how to be paid: rank well, receive the session, monetize the pageview. Under the retrieve-and-cite contract, being found means being pulled into a retrieval pool a publisher cannot see, and being paid is not part of the deal. The same pages do the work in both economies. Only the first one attached a session, and therefore a price, to that work. ## What Analytics Insight reported, and what we verified Analytics Insight's October 11, 2026 publisher guide argues that rank no longer predicts citation, that each engine selects differently, and that publishers win with original evidence rather than format checklists. The numbers check out. Ahrefs' March 2, 2026 analysis of 863,000 keywords and 4 million AI Overview URLs found 37.9% of cited URLs in the top 10 result blocks, 31.2% at positions 11 to 100 and 31% beyond position 100. Ahrefs' August 2025 study of 15,000 prompts found ChatGPT, Gemini and Copilot citations overlapped Google's top 10 only 12% of the time on average, with Perplexity at 28.6%. One caveat the source rightly flags: Ahrefs changed its parser between the July 2025 figure of 76% and the 2026 figure, so the drop is directional rather than a clean trend line. ## How Google, Perplexity and ChatGPT each pick their sources ### Google AI Overviews: core ranking plus query fan-out Google's Search Central guidance states that AI Overviews and AI Mode build on its existing ranking systems, with no special markup required. The decoupling comes from fan-out: Google splits one search into several sub-questions and cites whichever pages answer each one with precision. Ahrefs attributes the 2026 shift partly to the January 2026 move of AI Overviews onto Gemini 3, which leans harder on fan-out. A page that never ranks for the head term can win a citation by owning one sub-question. ### Perplexity: its own pipeline, closest to Google Perplexity runs its own retrieval and ranking rather than reading Google or Bing results, yet its citations land in Google's top 10 about 28.6% of the time, the highest alignment of any assistant Ahrefs tested. In practice Perplexity rewards pages that answer the question directly and stay current, which makes it the engine where classic authority signals still carry the most weight. ### ChatGPT: search partner, own retrieval and licensed sources ChatGPT draws on a search partner, its own retrieval layer and licensing agreements with specific publishers. That mix defines which pages are eligible at all, and ChatGPT and Gemini showed the lowest overlap with Google's top 10 in Ahrefs' data. For a publisher, eligibility is the first gate: a page the crawler cannot reach is not a candidate, whatever its quality. ## What gets a publisher into the sources - Answer the main question and its likely sub-questions in full on one page, because fan-out retrieval rewards topical coverage over the headline keyword. - Lead with original data, a documented method or a clearly sourced explanation; a paraphrased summary gives the engine nothing specific to extract and credit. - Keep facts and figures dated and current, since engines favor fresh evidence on subjects that move. - Find the sub-questions already sending citations to competitors, locate the missing evidence on your page, and fill that gap instead of rewriting blind. - Keep AI Search crawlers in: Cloudflare now distinguishes AI Search bots from AI training crawlers, and blocking the former removes the page from the pool entirely. - Track citation frequency and referral traffic as separate metrics; the source is right that one never predicts the other. ## Why being in the sources is necessary but not a business model ### The strongest case that source selection is the whole game The optimistic reading, which Analytics Insight and Google's own guidance both support, is that the engines are meritocratic: build evidence, cover the sub-questions, stay crawlable, and citations follow regardless of rank. On that view the publisher's problem is editorial, and the publishers who master source selection across three engines will compound visibility the way early SEO winners compounded traffic. ### Why a citation without a payment clause is distribution without revenue The source itself concedes the gap: citation frequency and referral traffic measure different things. Pew Research found in July 2025 that users click a link inside an AI summary in about 1% of visits, so a page can win every sub-question and still deliver no session to monetize. Mastering source selection gets a publisher into the answer. It does not get the publisher paid, because the retrieve-and-cite contract has no payment clause, and no editorial improvement writes one. ## What this means for publishers and for the brands that need them ### For publishers: map the retrieved archive, then sell it Treat source selection as the distribution layer and measure it as such: which pages each engine retrieves, for which commercial sub-questions, how often. Then price that retrieval as inventory rather than keep pricing the few pages that still rank and still draw sessions. The archive fan-out now pulls from is the asset. It needs a buyer, not another traffic recovery plan. ### For CMOs, media buyers and agencies: the sources are where you get in If the engines select from pages a brand does not own and cannot rank its way onto, the only purchasable route into the answer is presence on the publisher pages being retrieved. Audit which sources each engine cites in your category and budget for presence on those pages as a media line, because the paid-search bid no longer tells you whether you reached the answer. ## Three signals to watch before 2027 First, Google moved AI Overviews onto Gemini 3 in January 2026 and Ahrefs reads the citation shift as a result, so selection will keep moving deeper into the archive with each model update. Second, Cloudflare began defaulting ad-supported pages to block mixed-use crawlers on September 15, 2026, giving publishers a lever over who may retrieve them for the first time. Third, Amazon wired its DSP into ChatGPT on September 10, 2026, putting a demand pipe beside the answer while the pages feeding it still have no exchange. Selection is getting more opaque, access is getting more controllable, and demand is arriving; the publisher side still needs the rail that connects the three. ## Conclusion Hold on to this: knowing how AI search engines select sources is now a publisher's distribution strategy, and being in the sources is the new contract, one that pays in citations and not yet in revenue. Smalk AI writes the missing payment clause as Generative Engine Advertising, an AI Search ad network that places native ads for AI agents on the publisher pages engines retrieve, so publishers earn from the archive fan-out now draws on and brands buy presence where the answer is built. What to watch next: whether retrieval access, which Cloudflare just handed publishers, becomes the leverage that turns being in the sources into being paid as one. ## FAQ ### How do AI search engines select which sources to cite? They use retrieval-augmented generation: the query is split into sub-questions, candidate passages are retrieved from a wide pool, an answer is written from those passages, and the pages they came from are cited. Selection happens at the passage level, which is why pages that do not rank for the main keyword still get cited. ### Does my page need to rank in Google's top 10 to be cited in AI Overviews? No. Ahrefs' March 2026 analysis of 4 million AI Overview URLs found 37.9% of cited pages in the top 10 result blocks, 31.2% at positions 11 to 100 and 31% beyond position 100. Ranking helps, but most citations now come from outside the top 10. ### Do Google, Perplexity and ChatGPT pick sources the same way? No. Google AI Overviews build on its core ranking plus query fan-out; Perplexity runs its own pipeline and overlaps Google's top 10 about 28.6% of the time; ChatGPT combines a search partner, its own retrieval and licensed publisher content, with the lowest overlap. Publishers should track retrieval per engine. ### What is the most effective way for a publisher to get into AI sources? Identify the sub-questions in your topic that already send citations to competing pages, find the evidence or explanation your page lacks, and add it. Original, dated data and full sub-question coverage outperform formatting tricks, and AI Search crawlers must be allowed in for the page to be eligible at all. ### If my content is cited by AI engines, will I get traffic from it? Rarely. Pew Research found in July 2025 that users click a link inside an AI summary in about 1% of visits. Citation and referral are separate metrics, so being in the sources delivers influence inside the answer without delivering the session that display advertising monetizes.