AI Citations and Source Selection

Three decisions that get treated as one

Source selection is not one step. An AI answer surface decides which pages to retrieve, which of those to actually use when composing the answer, and which to show the reader as a citation — a visible attribution, usually a clickable link, beneath or beside the generated text. Those three sets are not the same set. A page can be retrieved and never used, used and never cited, or cited while contributing almost nothing to the sentence it sits under.

Google AI Overviews, Google's AI Mode (the conversational search mode launched in Labs in March 2025 and rolled out to all US users that May), ChatGPT search, Perplexity and Microsoft Copilot all perform all three steps, and not one publishes the ranking function governing any of them. Each publishes an eligibility rule and a disclaimer instead. So the answerable question is not what the algorithm is but how much of the classic organic ranking system carries over into the citation set — and that has been measured, repeatedly, by parties who disagree with each other.

The eligibility rule is the only part Google states plainly

Google's AI features and your website documentation, last updated 10 December 2025 and checked on 29 August 2026, sets the precondition without hedging: to be shown as a supporting link, "a page must be indexed and eligible to be shown in Google Search." No separate AI index, no separate submission route, no opt-in. The same page states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary," and that "there's also no special schema.org structured data that you need to add." Gary Illyes said the same from a stage at Google's Search Central Deep Dive, reported 24 July 2025: normal SEO is what works for AI Overviews. Microsoft made the same architectural move in February 2026, folding Copilot and grounding into the existing Bing Webmaster Guidelines rather than issuing a parallel set.

This is the most useful documented fact in the subject and the one most often skipped, because it collapses a large volume of AI-specific advice back into indexation work that is neither novel nor billable at a premium. The cost of ignoring it is paid in budget spent on markup add-ons and file formats the platform has said in writing are not required.

What each operator will say about selection, and what it will not

Google is the fullest of the four and is still thin. Beyond eligibility it says AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response," and attributes everything else to "core information quality systems." It does not say how many candidates are retrieved, how the citation set is chosen from them, or whether scoring happens at passage or document level. The Search Central documentation on AI features is worth reading in full for how quickly it runs out of things to say.

OpenAI offers a disclaimer: ChatGPT "ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed." It documents that ChatGPT "rewrites your query into one or more targeted queries" before sending them to search providers, and names Microsoft among those providers — the documented basis for the claim that ChatGPT search runs on Bing. OpenAI describes partners in the plural and runs its own crawler, so that claim overstates the documentation.

Perplexity, whose entire product proposition is attribution, documents its selection least: it gathers information "from authoritative sources like articles, websites, and journals" and attaches "numbered citations linking to the original sources." It does not define authoritative, does not name the index it searches, and has published no ranking documentation. The silence is itself the finding.

Microsoft is the most explicit as of 2026. The Bing Webmaster Guidelines rewritten on 26 February 2026 cover "Bing search experiences, Copilot, and grounding API results" in one document, add an eligibility category for grounding results and citations, and adopt the term Generative Engine Optimization — shaping content for inclusion in AI-generated answers. Microsoft also gave two content instructions no other platform has: facts should be stated directly rather than implied, and entity names should be clear and consistent, with no ambiguous references.

Why no two overlap studies agree

Do AI surfaces cite the pages that already rank? Four published answers point the same direction and disagree on magnitude by nearly a factor of two. Ahrefs, on 863,000 keyword SERPs and four million AI Overview URLs, reported on 2 March 2026 that 37.9% of citations also appeared in the first ten organic results, with roughly 31% ranking 11 to 100 and roughly 31% outside the top 100. seoClarity, on 362,000 US desktop queries with data to 12 October 2025, reported 32% overlap with the top ten, rising to 94% against the top twenty. BrightEdge reported 54.5% in September 2025. Semrush, on 5,000 keywords in July 2025, found AI Mode overlapping the top ten at roughly 54% by domain but only about 35% by URL.

Two of those series run in opposite directions over the same period. Ahrefs, repeating its own methodology, saw top-ten overlap fall from about 76% in July 2025 to about 38% in March 2026; BrightEdge's rises from 32.3% in May 2024 to 54.5% in September 2025. Both cannot describe the same reality. They sample different query mixes, in different countries, on different dates, through proprietary parsers each vendor built.

Ahrefs, seoClarity, BrightEdge and Semrush each funded its own study and each sells products whose value proposition is affected by the answer. Vendor research is not automatically self-serving — Ahrefs published a null result on schema that cuts against selling schema tooling — but the metric chosen and the cut selected for publication are commercial choices made by an interested party. Pew Research Center is the only large study here funded by an organization with nothing to sell.

A citation is a pointer, not a warrant

The word citation borrows its authority from scholarship, where it means a checkable claim of support. In an AI answer it means something weaker: this link was displayed alongside this text. Liu, Zhang and Liang, in Evaluating Verifiability in Generative Search Engines (Findings of EMNLP 2023), found that "on average, a mere 51.5% of generated sentences are fully supported by citations and only 74.5% of citations support their associated sentence." That figure is now quoted as though it described Google. It does not. The study evaluated Bing Chat, NeevaAI, perplexity.ai and YouChat in 2023, before AI Overviews and AI Mode existed, and no equivalent audit of the current surfaces was found as of 29 August 2026 — worth saying out loud, because the absence of a replication is not evidence the problem was fixed.

Nor does a citation mean ranking: Microsoft warns that the citation counts in its own reporting do not indicate ranking or page importance, and a page can be cited because it answered one sub-query cleanly rather than because it is the best page on the topic. Nor does it mean credit. Seer Interactive's March 2026 work named the gap between content that is used and a brand that is named a ghost citation; Semrush's 2026 AI Visibility Index, released 26 June 2026 on 126 million US prompts, found the overlap between brands mentioned and domains cited on Gemini can be as low as 30%. Optimizing for citation when the real failure is that the model never names you is work aimed at the wrong step.

The peer-reviewed work points away from copywriting

Three academic results bear on whether writing a page differently gets it cited, and practitioners quote the weakest of them most often. The paper everyone cites is Aggarwal et al., GEO: Generative Engine Optimization, accepted to KDD 2024, reporting visibility gains of "up to 40%" from specific content tactics. Those gains were measured on GEO-BENCH, a benchmark the authors built, against generative engines the authors constructed — not against live AI Overviews, AI Mode, ChatGPT search or Perplexity. The finding is real inside the benchmark; the extrapolation is what gets sold.

The stronger result cuts the other way. Puerto, Gubri, Green, Oh and Yun, C-SEO Bench: Does Conversational SEO Work? (NeurIPS Datasets and Benchmarks 2025), tested ten conversational-SEO rewriting methods across two search tasks in three domains each, and added a protocol varying how many competitors adopt the same tactic. Their conclusion: "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking," and "as we increase the number of C-SEO adopters, the overall gains decrease." Getting into the retrieved set beats how the page is written once it is there, and whatever advantage a tactic holds decays as it spreads — exactly the property a tactic sold to an entire market will exhibit.

The one controlled field test agrees with them. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched controls and reported on 11 May 2026 no meaningful uplift on any platform: AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%.

Provenance: being able to show where an answer came from

If a citation is a pointer rather than a warrant, the burden of provenance moves to whoever repeats the answer, and the tooling serves that burden badly. AI answers are not stable artifacts. The same prompt on the same surface returns different text and a different citation set from one week to the next, the surfaces offer no permalink to a generated response, and none of the four operators exports what a given answer used. A screenshot with no timestamp, no prompt text and no record of which surface produced it documents nothing.

The working discipline is unglamorous, and it separates a defensible claim from an anecdote. Record the surface and the exact prompt, the date and time, the country and device context where the surface personalizes by them, the cited URL set as displayed, and a saved copy of the answer text, because the answer will not reproduce. For a fuller treatment of keeping that kind of record, The Expert Record is a companion reference on documenting AI use in professional work.

What you can actually measure about your own citations

Until 2026, no publisher could see its own citation exposure at all. Two operators have since opened a window, and both windows are partly shuttered. Microsoft moved first: the AI Performance report in Bing Webmaster Tools, announced 10 February 2026 and still labeled public preview, gives Total Citations, Average Cited Pages and grounding queries across Copilot, AI summaries in Bing and select partner integrations. As of 29 August 2026 it is the only per-page citation count any operator publishes to site owners.

Google's Search Generative AI performance report followed on 3 June 2026, initially to a subset of UK site owners. It gives impressions, pages, countries, devices and dates, and it gives no click data and no query dimension — the two things a publisher most wants. Whether it has since left UK-only preview could not be confirmed as of 29 August 2026. John Mueller confirmed on 6 August 2026 that AI Overviews and AI Mode data sit inside the general Search Console performance report. Everything else on the market is third-party sampling, which estimates a population of prompts rather than measuring your exposure.

On what a citation is worth once won, the only large non-vendor measurement remains Pew Research Center's July 2025 study of 900 US adults and 68,879 searches: users clicked a link inside the AI summary on 1% of visits, and clicked a traditional result on 8% of visits with a summary against 15% without. Google said in May 2024 that "the links included in AI Overviews get more clicks than if the page had appeared as a traditional web listing" and has never published the underlying data; its December 2025 documentation restates it in the narrower form that such clicks are higher quality. Two claims, two metrics, and only one has independent evidence beside it.

Frequently asked questions