- Confirm every page you want cited is indexed and snippet-eligible, and remove any nosnippet or data-nosnippet attributes covering the passages you want quoted.
- Map the five or six follow-up questions a buyer asks after your head query, and give each one its own section that answers it in the first sentence.
- Put a unit, a date and a condition on every number you publish, so a passage can be attributed on its own without the rest of the page.
Most advice about AI Overviews sources starts from one assumption: rank well, and Google will cite you. That was a reasonable working model in 2025. It is weaker now. In a March 2026 Ahrefs study of about 4 million AI Overview URLs, only about 38% of the cited pages ranked in the top 10 for the query that triggered the overview. Ahrefs’ earlier study, from July 2025, put that overlap at roughly 76%. The newer data also contains a finding the ranking model cannot explain: about a third of cited URLs did not rank in the top 100 for that query at all.
So what does pick them? Google documents part of the answer. The rest you can infer from how the system is built and from what the citations actually look like. The rest of this piece goes in order, from the gate a page has to pass to the sentence that finally gets quoted.
AI Overviews sources: what a supporting link has to clear
An AI Overview is a generated answer, and the links beneath it are not a bibliography the model wrote from memory. They are pages the system retrieved and used while composing the answer. Google’s Search Central documentation on AI features calls these “supporting links,” and it sets a deliberately plain bar: a page must be indexed and eligible to appear in Search with a snippet. Google says there are “no additional technical requirements” beyond that.
Two things follow. There is no separate AI index to get into, and no AI-only crawler you need to satisfy. The gate is also binary and checkable. Google says you “don’t need to create new machine readable files, AI text files, or markup” to appear in these features, and that there is no special schema.org structured data to add. Its documentation does not mention llms.txt by name. If someone sells you an llms.txt file as the key to Google citations, the official guidance does not support that claim.
Eligibility is also conditional on the query. Google says AI Overviews are “only shown when our systems determine that it is additive to classic Search.” Many queries never trigger one, so a page can be perfectly eligible and still have no overview to be cited in.
Check the page, not the site: run the URL through URL Inspection in Search Console, then search the rendered HTML for nosnippet, data-nosnippet and max-snippet. Stray directives often come from templates, or from a staging noindex that never got removed. Google notes that preview-control changes can take several days to several months to take effect, so audit early.
A page that should be cited but never appears often fails this check before anything else is evaluated. Whether the citations send any clicks is a separate question, and we cover it in our analysis of AI Overviews and traffic.
Classic ranking still fills the candidate pool
Ranking does not decide citations, but it still does most of the work of finding them. The Ahrefs numbers make the point concrete:
| Where the cited URL ranked for the query | Share of cited URLs (March 2026) |
|---|---|
| Top 10 organic results | 37.9% |
| Positions 11 to 100 | 31.2% |
| Beyond position 100 | 31.0% |
Ahrefs, all SERP blocks included, 863,000 keyword SERPs. Figures rounded.
Read the table two ways. The top 10 is still the single largest group, so strong rankings give a page a real head start. But almost a third of citations come from pages that did not rank in the top 100 at all. The organic list is not the whole candidate pool. Ahrefs suggests some of those citations come from SERP features rather than blue links, which is plausible, though the data cannot settle every case.
Be careful with the trend itself. Ahrefs changed its parsing between studies, so the drop from 76% to 38% mixes a real shift with a change in measurement. The authors also note that AI Overviews moved to Gemini 3 in January 2026. Ahrefs is a vendor reporting its own crawl and tooling. Treat these figures as direction, not precision. The direction is still clear: the top 10 stopped being the whole story.
The practical upshot is that top-10 rank works as an entry ticket for your head term, not as a guarantee of citation. Google does not publish its retrieval stack in detail. Its documentation does say supporting links must be eligible to appear in Search, which suggests they come from the same index classic ranking draws on. A page that cannot rank for any part of a query’s neighborhood has a hard time getting into the pool at all.
Query fan-out: the question Google answers is not the one you typed
Google says these features may use query fan-out, which it describes as “issuing multiple related searches across subtopics and data sources.” This is the mechanism that explains most of the gap between ranking and citation, so it is worth understanding precisely.
Take a buyer who types “project management tool for a small agency.” That head query hides a dozen smaller questions: what it costs per seat at ten people, whether it supports client portals, whether it integrates with the time tracker the agency already uses, how long onboarding takes, and whether it is worth leaving a spreadsheet. A system that wants to write a useful answer cannot do that from a single result set. It issues narrower searches, gathers candidates for each subtopic, and composes from the best material it found.
Google does not expose the subqueries, so the specifics here are inference. The logic follows from the documented behavior, though, and it has a useful consequence. A page does not have to rank for the head term to get cited. It has to be the cleanest answer to one of the subquestions. A pricing page that states the per-seat cost for a ten-person team may rank poorly for “project management tool” and still be the page that supplies the cost line in the answer.
Google also says AI features can display “a wider and more diverse set of helpful links” than classic search. Fan-out is the most plausible reason. More subqueries mean more distinct pages surviving into the final answer.
Here is an illustrative case, not a measured one. A project tool publishes a long guide, and its pricing sits as one paragraph in the middle of a feature overview. The fan-out query “agency project management pricing per seat” pulls candidates. The guide’s pricing paragraph is diffuse, buried under feature discussion and missing a specific figure with a date. A competitor’s pricing page, with an unambiguous table, takes the slot. The guide ranks well for the head term and gets no citation for the pricing claim. Same site, same ranking, different outcome. The unit that got selected was a passage, not a page.
That is the key shift. You are competing for subquestions, not only for the head term. AutoSEO Stack’s research step starts from the same kind of list, built from what buyers actually search rather than from guesses about what they might ask.
Passage selection decides which sentence gets quoted
Google has ranked individual passages since 2021, so a page has never been a single unit of relevance. Generative answers add one more step. The system needs a span it can attribute a specific claim to. At that stage the question is no longer “is this page about pricing?” It is “is there a passage here that states this claim clearly enough to support a sentence in the answer?”
That changes what a good section looks like. Four traits make a passage easy to select, and none of them is exotic.
It answers under a heading that names the question. A heading such as “How much does the Team plan cost?” followed by the answer in the first sentence gives the system a clean boundary to work with.
It carries its own scope. Compare “Pricing starts at $29 per user per month on the Team plan, billed annually, as of October 2026” with “Affordable plans for every team size.” The first can be attributed. The second cannot, because there is nothing in it to check.
It is self-contained. Pronouns such as “this” and “it” that point back three paragraphs leave a span unattributable the moment it is lifted out of the page.
It is specific without being inflated. A precise figure with its method stated reads as credible. A round number with no basis reads as marketing, and a system composing an answer has no reason to trust it over a more careful source.
Consensus is the other half of the story. When several retrieved sources state the same fact, an answer tends to restate it and cite one or more of them. When sources conflict, answers either hedge or lean toward the source with the most specific, checkable detail. That is an inference from how generative answers usually behave, not a documented rule. It is still the behavior to design for. Being the source that says the specific thing plainly beats being one more source that says the generic thing.
Don't write for extraction alone. Google's guidance is people-first, and a page of stilted, question-shaped blocks written to be lifted is the same thin, mass-produced pattern that scaled-content guidance targets. Write one clear answer per section for a reader first. Clear answers tend to be extractable anyway.
Signals that shape selection, and what to do about each
Google does not publish a list of citation signals, and anyone who offers you one is usually selling a score. What follows are the factors that the documentation and the observable citation patterns support, with what each one means in practice.
Eligibility and crawlability
Everything in the first section is a signal too. A page that is blocked from crawling, that keeps its important content inside images, or that sits behind a nosnippet directive is excluded before any quality judgment happens. Google’s basics are unglamorous: allow crawling through robots.txt and any CDN in front of your site, make important content available as text, and link internally so the page is findable. Most of that work is invisible when it is done right, which is why it is the first place to look when a page should be cited and is not.
Specificity and verifiability
The strongest content signal in citation patterns is simple: specific, checkable claims get cited more often than generic ones. Google’s quality guidance asks for content with real expertise and first-hand detail, and in practice that means a number with its unit and date, a named tool or standard, the method behind a result, and an honest limit.
Compare two versions of the same claim. “Teams see faster onboarding” cannot be attributed to anything. “In our onboarding test with 40 accounts, setup took a median of three days with a template and nine days without” gives a system something to quote and a reader something to verify. Every number has to be real, of course. Fabricated precision is the fastest way to become uncitable, because a claim that falls apart under a single check poisons the rest of the page.
Freshness on time-sensitive queries
For queries where recency matters, Google has long weighted freshness, and AI answers inherit that bias. Pricing, product features, regulations and “best of” lists go stale quickly. A system composing an answer in October will prefer a page dated this year to one whose figures are from 2023. Stale pages lose citations quietly, with no error message to tell you why.
The fix is not to republish daily. Update the claims that changed, change the visible date when you do, and say what changed. A dated changelog line in a pricing section does more for trust than a fresh timestamp on an unchanged page.
Entity clarity and visible authorship
Citation requires attribution, and attribution requires knowing who is speaking. Pages where the author, the publishing organization and the date are visible and consistent give a system something to attribute to. Google’s quality guidance asks raters to judge experience, expertise, authoritativeness and trust. A byline with real credentials does not guarantee citation. But an anonymous page with an undated claim is the hardest kind of page to attribute, and difficulty is a reason to skip it.
Structured data as consistency, not a switch
Google says no special schema is required for AI features. That is a relief, not a reason to skip markup. Google’s guidance is to keep structured data consistent with the visible text. Schema that describes a price the page does not show, or an author the page does not name, creates a mismatch that helps nobody. Use markup where it describes what is on the page accurately, and leave it out where it does not.
A checklist for becoming a cited source
Work through these in order. The first two are gates; the rest shape whether you win the slot.
- Confirm eligibility on the pages you care about. Each should be indexed and snippet-eligible, with no nosnippet or data-nosnippet covering the passages you want quoted.
- List the subquestions. Start from your head query and write the next five or six questions a buyer asks: cost, integrations, setup time, limits, alternatives, and what switching involves. Those are your fan-out candidates.
- Give each subquestion its own section. Use a heading that states the question, answer it in the first sentence, and support the answer in the paragraphs that follow.
- Put scope on every number. State the unit, the date and the condition. Where you have a method or a source, state it in the same paragraph so the passage stands alone.
- Keep the head term competitive too. Top-10 rankings still supply the largest share of cited URLs, and the candidate pool starts there.
- Refresh time-sensitive claims on a schedule. Update the visible date whenever you change a figure, and note what changed.
- Measure what you can. Google says sites appearing in AI features are counted in Search Console’s Performance report under the Web search type. That data does not isolate AI Overview citations, so you will need a separate citation check to see which queries actually cite you.
What this means for your content
The honest summary is less tidy than most advice you will find. There is no single ranking factor to chase for AI Overviews sources. Eligibility decides whether a page can be considered at all. Classic ranking fills the candidate pool, and in the March 2026 data it still supplies a bit under two-fifths of cited URLs. Fan-out multiplies the number of chances, which makes the subquestions matter as much as the head term. Passage selection decides which sentence wins, which is why specific, self-contained, dated claims punch above their weight.
You cannot control the model. You can control whether your pages clear each gate, and that is most of the work: a page that answers a real subquestion, states its claims with scope, and stays current. The same passage logic applies outside Google as well. If you are working on ChatGPT, Gemini and Perplexity too, our guide to getting cited by AI engines covers the other engines.
That discipline is what AutoSEO Stack is built around. It researches the subquestions buyers actually ask, writes pages that answer them in a human voice, and publishes on a schedule, so the quality bar is part of every page rather than a check made at the end. To see how that works on your own site, start with AI visibility.