Research

How to show up in Google AI Overviews, by the evidence

The click loss from AI Overviews is measured and severe; we covered it in the disappearing click. This is the other half: what gets a brand cited inside the answer box. The correlation studies now exist, and they rank the levers in a different order than most vendors sell them.

Published 2026-07-2810 sources, every stat linked10 min read

Michael NefedovFounder

Which queries trigger an Overview at all#

Across an Ahrefs analysis of 146 million SERPs, 99.9% of keywords that trigger an AI Overview are informational.[4]Question phrasings trigger one on 57.9% of queries, “why” questions on 59.8%, definitions on 47.3%.[4] If your buyers ask questions before they buy, and in B2B they do, the Overview now sits on top of exactly the queries your content strategy was built for.

That is worth stating plainly before any tactics: you do not choose whether to compete on this surface. Google chose for you when it put an answer box on half of your category's question queries and the top result started losing 58% of its clicks to it.[5]The only open question is whether the Overview cites you or your competitor.

How Google picks the citations#

An AI Overview is not written from the ten blue links you see. Google expands the query into synthetic sub-queries, retrieves candidates for each, and assembles the answer from that wider pool. Google describes the technique itself for AI Mode, and the measured citation data shows the same behaviour in AI Overviews.[1]The consequence is visible in the numbers: of 4 million AI Overview citations Ahrefs examined in early 2026, only 37.9% ranked in Google's top 10 for the visible query. 31.2% ranked in positions 11 to 100, and 31.0% ranked nowhere in the top 100.[3]

A year earlier the top-10 share was roughly 76%.[6] The trend line matters more than the snapshot: Google is drawing its answers from the fan-out pool more and from the head-query SERP less with every measurement. Optimizing only your head-query ranking is optimizing the shrinking third of the citation pool.

Lever 1: cover the fan-out, the strongest measured effect#

The largest correlation anyone has published: pages that rank across the fan-out sub-queries of a topic are 161% more likely to be cited in the AI Overview for the head query, with a Spearman correlation of 0.77.[1] In practice that means one pillar page chasing one head term is the wrong shape. The right shape is a set of pages, each one fully answering one of the specific questions Google generates when it decomposes the topic: the comparison, the pricing question, the how-does-it-work, the is-it-worth-it.

This is the mechanism behind advice you have heard in vaguer forms (“build topical authority”, “answer related questions”). The measurable version: enumerate the sub-queries your buyers' questions decompose into, and make sure a page of yours is a retrievable, self-contained answer to each one. It is also why answer-shaped long-tail pages, cheap to produce and easy to dismiss as small traffic, keep showing up as AI Overview citations: they are what the fan-out retrieves.

Lever 2: earned mentions, the strongest brand-level effect#

At the brand level rather than the page level, Ahrefs correlated visibility factors across 76.7 million AI Overviews for the 50 most visible sites. Branded web mentions correlate at 0.664, the top brand-level factor. Mentions on highly linked pages reach 0.70. Mentions on YouTube, the single most cited domain in AI Overviews, reach 0.740.[2]

This matches what we measure across engines generally: most AI citations point at third-party sources, not at brand sites, and the only published before/after experiment on third-party distribution measured a 239% citation lift. We wrote that mechanism up in third-party citations. For AI Overviews specifically, the practical reading: the review sites, comparison posts, and YouTube explainers that mention you are not marketing nice-to-haves, they are the substrate Google's answer is assembled from.

Lever 3: extractable evidence on pages Google can read#

Two conditions gate everything above. First, mechanical: the AI Overview pipeline reads rendered HTML, so evidence locked inside client-side JavaScript is invisible to it. A pricing table that only exists after hydration cannot be quoted. Second, structural: the foundational GEO research measured up to 40% visibility gains from adding concrete statistics, quotable sentences, and cited sources to otherwise identical content.[7] Passages that read like defensible answers get lifted; prose that reads like positioning does not.

Length is not evidence. Word count correlates with AI citation at roughly 0.04, and grounding quality plateaus around 540 words in Dan Petrovic's 7,000-query study.[2] A 600-word page that answers one question completely, with a dated statistic and a named source, beats a 4,000-word pillar that answers ten questions halfway.

What does not work, despite being widely sold#

Schema markup first. The correlation studies that exist are vendor-run and point in opposite directions; the largest, SE Ranking's pass over 216,524 pages, found pages with FAQ schema earning slightly fewer ChatGPT citations than pages without it, because the visible FAQ text does the work and the JSON-LD adds nothing on top.[8]Notably, schema appears nowhere in Ahrefs' measured AI Overview visibility factors.[2]Keep it for Google's classic rich results, which is hygiene, and expect zero citation lift from it. llms.txt is weaker still: across 137,000 domains, 97% of llms.txt files receive zero requests at all, and almost none of the remaining traffic comes from AI assistants.[9] We keep both in the run-and-forget bucket and price them accordingly, at zero.

The pattern across the null results is consistent: tactics that talk to machines about your content do not move citations; changes to what the content actually says do. If an agency's AI Overviews offer is schema markup plus an llms.txt file, you are buying the two measured zeros. We went through every major vendor's claims in can GEO be automated.

Measuring whether any of it worked#

AI Overviews change without changelogs, and the same query does not reliably produce the same Overview. Peer-reviewed work on AI-visibility measurement puts the floor at roughly seven repeated runs per query before a per-brand estimate stabilizes.[10] One manual check per week produces noise wearing the costume of data.

Our loop samples Google AI Overviews and AI Mode daily from the live search surface, alongside ChatGPT, Perplexity, and Gemini through their APIs, and re-runs every tracked query after each shipped fix, so a citation gained or lost is attributable to the change that moved it. The sampling design, run counts, and confidence intervals are public at noetio.com/methodology.

Sources

1. Pages ranking across fan-out sub-queries are 161% more likely to be cited in the AI Overview; Spearman 0.77. Study of 10,000 keywords and 33,000 Gemini-derived fan-out queries by Joshua Hardwick with SurferSEO, “Query Fan-Out Impact”, 2026; independently reported by Search Engine Land.
2. Brand-factor correlations across 76.7M AI Overviews, top 50 sites: branded mentions 0.664, highly linked pages 0.70, YouTube mentions 0.740; word count ~0.04; grounding plateau ~540 words (Dan Petrovic, 7,000+ queries): Ahrefs, 2026.
3. 37.9% of AI Overview citations from Google's top 10, 31.2% from 11 to 100, 31.0% outside the top 100; 4M cited URLs, 863,000 keyword SERPs, published 2 March 2026: Ahrefs.
4. 99.9% of AI Overview keywords informational; question queries trigger at 57.9%, “why” at 59.8%, definitions at 47.3%; 146M SERPs: Ahrefs, 2026.
5. 58% lower clickthrough for the top-ranking page when an AI Overview is present; 300,000 keywords, published 4 February 2026: Ahrefs.
6. The earlier measurement placed roughly three quarters of AI Overview citations inside the top 10; the March 2026 update reports the drop to 37.9% against its own initial study: Ahrefs, “Update: 38% of AI Overview Citations Pull From The Top 10”.
7. Adding statistics, quotations, and citations lifts generative-answer visibility by up to 40% on a 10,000-query benchmark: Aggarwal et al., “GEO: Generative Engine Optimization”, arXiv:2311.09735, KDD 2024.
8. Pages with FAQ schema averaged 3.6 ChatGPT citations vs 4.2 without, across 216,524 pages on 129,000 domains in 20 niches: SE Ranking, 2026. Vendor-run studies on schema conflict; no large controlled study shows a citation lift.
9. 97% of llms.txt files across 137,000 domains received zero requests (server logs, May 2026); of files with traffic, 96% of requests were bots, mostly non-AI: Ahrefs study, reported by Search Engine Journal, June 2026.
10. Roughly seven or more repeated runs per query needed for stable AI-visibility estimates: Schulte, Bleeker and Kaufmann, “Don't Measure Once: Measuring Visibility in AI Search (GEO)”, arXiv:2604.07585, 2026.

Correlation studies describe association, not causation; where a controlled experiment exists it is labeled as such. Vendor-run studies are labeled with the vendor. Noetio's own measurement methodology is published at noetio.com/methodology.