Why your #1 ranked page is invisible in AI search

On May 15, 2026, Google published its first official guide to optimizing for AI search. The headline message: AEO and GEO are "still SEO." No separate framework needed, no special markup, no llms.txt files.

Then Ahrefs ran the numbers.

Across 863,000 keywords and 4 million URLs, pages cited in Google AI Overviews that also ranked in the top 10 dropped from 76% to 38% in seven months. BrightEdge, using a different dataset, puts the overlap even lower — closer to 17%. Surfer SEO found that 67.82% of AI Overview citations don't rank top 10 at all.

Both things are true. AI search is "still SEO." And a #1 ranking no longer predicts who gets cited.

This post is about reconciling those two facts — what actually changed inside Google's AI pipeline, what data the new behavior reveals, and what to change about your content if you want to keep showing up in the answer.

Citations stopped tracking rankings

The clearest single data point comes from Ahrefs. In a study analyzing 863,000 keywords against 4 million AI Overview URLs, only 38% of cited pages also appeared in the top 10 organic results for the same query. In Ahrefs' July 2025 version of the same study, that number was 76%.

That's a 50% drop in seven months.

The citations didn't disappear. They redistributed. Ahrefs found AI Overview citations now split almost evenly across three buckets:

• 37.9% from pages ranking in the top 10
• 31.2% from positions 11–100
• 31.0% from pages that don't appear in the top 100 at all

A separate BrightEdge analysis published February 12, 2026 puts the top-10 overlap even lower at approximately 17%, depending on methodology and dataset. Surfer SEO's December 2025 study of 173,902 URLs found that 67.82% of AIO citations come from pages outside the top 10 organic results.

The studies don't agree on the exact percentage, but they agree on the direction. Citation behavior changed materially between mid-2025 and early 2026. Ranking #1 used to be a strong proxy for citation. Today it's a coin flip at best.

Ahrefs flags one contributing factor: Google upgraded AI Overviews to Gemini 3 globally on January 27, 2026. The shift in citation behavior maps onto that timing.

What Google finally documented

For two years, the AEO and GEO community speculated about how AI Overviews picked sources. Google stayed quiet. Then on May 15, 2026, they published Optimizing your website for generative AI features on Google Search — and finally explained the mechanism.

Two systems run behind every AI search:

RAG (Retrieval-Augmented Generation) — the model doesn't generate answers from training data alone. It first retrieves real documents from Google's index, then synthesizes an answer grounded in those documents.

Query Fan-Out — the original user query gets exploded into multiple parallel sub-queries that retrieve different passages from different sources.

Google's own example in the guide: a user searches "fix lawn weeds." The system rewrites that into parallel sub-queries like "best herbicides," "remove weeds without chemicals," and "prevent weeds in spring." Each one runs against the index. Each one pulls a different passage from a different page. The final AI Overview is assembled from those passages.

iPullRank's December 2025 analysis quantified the scale. Standard AI Mode queries generate 8–12 sub-queries. Deep Search mode can fire hundreds of sub-queries per single user prompt. The same research found that AI search queries average 70–80 words behind the scenes — compared to 3–4 words for the user-typed query that triggered them. The system is doing 17–26× more retrieval work than what shows up in your logs.

This is the part nobody could see from the outside. Your "fix lawn weeds" page might rank #2 for that exact phrase and still lose every citation slot — because the AI ran a different search than the one the user typed, and your page didn't have the best paragraph for any of the sub-queries.

The contradiction that isn't

There's a second data point that, on the surface, seems to contradict the Ahrefs collapse.

iPullRank's relevance engineering team — Mike King's group, the same researchers who built Qforia for visualizing fan-out — reports that ranking position remains the strongest predictor of AI citation in their own data. Patrick Schofield, iPullRank's Lead Relevance Engineer: "Traditional ranking position is still the great gatekeeper in AI citations. Our data shows a stark drop-off in AI citations for any page ranking outside of the top 10."

So which is it? Is ranking still the gatekeeper, or is half of all citations now coming from outside the top 100?

Both. The reconciliation is that ranking shifted from sufficient to necessary.

In mid-2025, ranking #1 was a near-guarantee of citation. Today it isn't — because the system might run a sub-query you don't rank for and pull from a page on position 30 that does. But ranking outside the top 10 on the underlying query still drops your citation probability sharply.

The new shape of the game:

• Top 10 ranking is the entry ticket to the candidate pool.
• Inside the pool, the page with the best paragraph for the most fan-out sub-queries wins the citation.
• Outside the pool, you're competing for the ~31% of citations Google pulls from positions 11–100 — and your odds drop fast.

YouTube is the outlier worth flagging. Ahrefs' Brand Radar finds YouTube is now the single most-cited domain in AI Overviews, up 34% over six months, accounting for 18.2% of citations that come from outside Google's top 100. Video transcripts have become a category-defining AEO surface, and almost no SEO content programs are treating them that way.

What cited pages have in common

Across the studies, the cited pages share four consistent traits.

Length. iPullRank's research found cited URLs average 1,800 words. Non-cited URLs average 1,200. That's not a case for endless padding — thin content can't cover the breadth of sub-queries fan-out generates, but bloated content gets skipped just as fast.

Structured data. SE Ranking's analysis found ~65% of pages cited in Google AI Mode include schema markup. The number is even higher for ChatGPT at ~71%. Google's May 15 guide says structured data isn't required for AI features, and that's literally true — but the citation data suggests it correlates strongly with being picked. Use it for traditional rich results and treat the AI lift as a free byproduct.

Entity density. iPullRank found mid-tail queries saw a 292% lift in citation probability when pages were optimized for entity density — the count and clarity of named brands, tools, methods, and adjacent concepts. AI systems retrieve pages that surface the right entities for a sub-query. Pages that paraphrase around the specific terms users search for get skipped.

Multi-format coverage. Cited pages tend to mix prose, definitions, comparison tables, and FAQ blocks. Each format paraphrases well into different sub-query answers. A 1,800-word essay with no tables, no lists, and no FAQ has fewer surfaces to be quoted from than a 1,400-word piece structured around multiple answer shapes.

The unit shifted from page to passage

Each H2 section on your page now functions like its own mini-result for whichever sub-query it answers best. Optimize accordingly.

Lead every section with the answer. Don't bury the lift behind anecdote or framing. The first sentence of each H2 should be quote-ready in isolation.

Use question-shaped H2s and H3s. Headers that mirror likely fan-out sub-queries make the retrieval step's job easier. "Best CRM features for small teams" beats "Features."

Add an FAQ block that targets fan-out sub-queries explicitly. Three to five questions per page, phrased exactly how a user might prompt them, with one-paragraph answers underneath. This is the single highest-leverage edit for pages that already rank but don't get cited.

Increase entity density. Audit each page for the named brands, tools, methods, statistics, and adjacent concepts a user might search for. Mention them by name. Don't substitute generic phrases.

Use schema markup for the formats that earn it. FAQPage, HowTo, Article, Product. Google's guide says it's not required for AI; the citation data says it correlates anyway.

Stop trying to rank for everything on one page. Pick the cluster of sub-queries you want to own, build each one a dedicated section with a dedicated answer, and let the page win citations across the cluster rather than fighting for the main keyword alone.

For the deeper engine-specific tactics, see How to rank on ChatGPT and How to rank on Perplexity. The principles above apply to all four engines; the per-engine deep dives cover the differences in retrieval behavior.

Finding the sub-queries Google generates

You can't optimize for sub-queries you can't see. Three tools currently surface fan-out expansions:

• Qforia. Built by Mike King at iPullRank. Generates fan-out query expansions for any topic. The closest available view into how Google's AI Mode expands a user query.
• GoFishDigital's Gemini API + Screaming Frog workflow. A scriptable pipeline for extracting AI Overview fan-outs at scale across a site or competitor set.
• WordLift AI Visibility Fan-Out. Collaborative research interface for the same purpose.

You can also generate plausible fan-out queries yourself by prompting any frontier LLM: "Here's a search query: [X]. If you were a search system that needed to retrieve diverse passages to answer this comprehensively, what 10–15 sub-queries would you generate?" The exact queries won't match Google's — fan-out is probabilistic and varies per run — but the recurring themes will. Those themes are what your content architecture should cover.

The goal isn't perfect replication of Google's internal fan-out. It's identifying the cluster of intents your topic triggers and building content that answers each one in a discrete, citable passage.

The measurement gap

Search Console tells you which queries your pages rank for and how often they get clicked. It doesn't tell you which paragraphs got pulled into an AI Overview, or which fan-out sub-queries you got cited on. GA4 doesn't see citations that never produced a click — and most don't.

The visibility moved to a layer none of the standard tools report on.

The workaround is to query the engines directly: run your prompt set against ChatGPT, Perplexity, Gemini, and AI Overviews on a schedule, capture the full answer text, and track which prompts cite you, which cite competitors instead, and which sources the engines pulled from. That's the measurement layer that maps onto how AI search actually works.

For the methodology breakdown — what counts as a "citation," how different engines define it, and how to build a defensible prompt set — see What counts as a citation. For the monthly workflow that wraps citation tracking into a client deliverable, see Writing a monthly client report in ten minutes with AEBOT.

What "still SEO" actually means

Google's May 15 guide is right that AI search is "still SEO" in the technical sense. Same crawlers. Same authority signals. Same E-E-A-T frameworks. The page that gets cited is the page that earned the right to be in the index in the first place.

But the unit of optimization changed. A page used to compete as a whole — one ranking, one click, one sentence about you in the SERP description. Now it competes paragraph by paragraph against the fan-out queries Google generates in the background.

Top 10 is the prerequisite. The competition for citation happens inside that pool, at the passage level, against sub-queries the user didn't type.

The teams that move first are the ones who stop optimizing pages and start optimizing answers.

Want to see which prompts cite you across all four engines? Run a free audit on AuditAE — drop in your brand, your domain, and 5 prompts your buyers actually ask. We'll show you which ones cite you, which cite competitors, and where the gap sits across ChatGPT, Perplexity, Gemini, and Google AI Overviews.