Google publishes its own requirements for AI Overviews...the list runs thinner than the optimization industry admits. A page has to be indexed and eligible for a snippet. Google's list does not go past that. Where the documentation stops the evidence splits hard, and two firms with real crawl infrastructure cannot both be right about whether ranking first correlates with citation. Their numbers move in opposite directions and do not agree on the level either.

Most writing on this carries no sample size and no date...both are attached to every study below.

Google's Own Stated Requirements

Entry is being indexed...the requirement ends there. Search Central's page on AI features says a page "must be indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link...there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary".

Three more lines from that page do real work.

  • Schema: Google states there is "no special schema.org structured data that you need to add".
  • Files: New machine readable files and AI text files are not needed either.
  • Guarantee: Meeting every requirement "doesn't mean that Google will crawl, index, or serve its content".

A separate guide on generative AI features goes further.

  • Hacks: Site owners are told to "prioritize effective SEO strategies over 'AEO/GEO hacks'".
  • Chunking: Breaking content into small pieces for AI systems is named as a tactic to ignore, alongside llms.txt files.
  • Plumbing: The generative features run on the core Search ranking and quality systems, so there is no second door to find.

The Schema Test Nobody Cites

Ahrefs published the best-controlled test on 11 May 2026, tracking 1,885 pages that added JSON-LD on their own between August 2025 and March 2026, each matched against three control pages with similar prior citation levels, 4,000 in all, over the thirty days either side. Citations in AI Overviews fell 4.6%, small but statistically significant. AI Mode rose 2.4% and ChatGPT rose 2.2%...neither rise clears random variation.

Limits matter, and Search Engine Journal's write-up carries them. Every tested page already held a hundred or more citations...the finding describes well-cited pages only. Among its caveats Ahrefs lists a SearchVIU experiment reporting that AI systems read only visible HTML during retrieval. Search Atlas cites that same experiment as confirming schema gets processed. Both readings cannot describe one experiment...the original page is not reachable to settle it.

Google's LocalBusiness documentation is blunt about the general case, saying the markup "does not guarantee that features that consume structured data will show up in search results". That documentation and the best-controlled test land in the same place on schema. Only one first-party statement points the other way, and it did not come from Google...at SMX Munich in March 2025 Microsoft's Fabrice Canel said schema helps Microsoft's LLMs understand content, and no data was published with it.

Fan-Out

An Overview can cite a page that does not rank for the query typed...the mechanism is documented. Google says AI Overviews and AI Mode "may use a 'query fan-out' technique", defined as issuing "multiple related searches across subtopics and data sources", and its AI Mode help page describes the same move.

So the query the reader typed sets off a fan of sub-questions nobody typed...each written by Google on the reader's behalf, and pages get compared against those. Ahrefs took 15,000 long-tail prompts collected in July 2025 and found 12% of cited URLs ranking in the top 10 for the original query, with roughly 80% not ranking in the top 100 at all.

Documentation for AI Mode also describes a confidence gate, saying the feature "will provide a set of web links if there's not high enough confidence in the quality or helpfulness of an AI response"...that is a documented off switch, and no page on the SERP for this question mentions it.

How Often Overviews Appear

Two large samples give two different prevalence figures...neither one is wrong.

  • Conductor: 25.11% of roughly 21.9 million American Google searches between 15 September and 12 October 2025.
  • Semrush: 6.49% of queries in January 2025, 24.61% in July and 15.69% in November, across 10 million keywords.
  • By industry: Conductor put Health Care highest at 48.75% and Real Estate lowest at 4.48%.

Both numbers describe 2025, and they do not cover the same weeks. Prevalence nearly quadrupled and then fell by a third inside ten months...a single prevalence number describes a month, not a feature.

Local Intent Flips The Number

Everyone repeats 68% for local search...the study behind it does not say that. Whitespark ran 540 queries across Houston, Phoenix and Denver in six industries in the second quarter of 2025, split by intent, searching from outside all three cities by their own account. That 68% is an average across all three intent types...the split underneath is the instruction.

  • Local intent: Overviews appeared on 15% of queries and the local pack on 93%.
  • Informational: Overviews appeared on 92% of queries and the local pack on 6%.
  • Hybrid: Overviews appeared on 97% of queries and the local pack on 17%.

Search Engine Journal gave that 68% as the rate for local business queries, and it did not link the study. Whitespark's own local intent figure is 15%. For "plumber near me" the map pack still decides it...the priced and comparison questions around it go to Overviews.

Does Ranking First Still Help

Two firms with crawl infrastructure have measured the overlap between top-ten rankings and Overview citations. In March 2026 Ahrefs re-ran its study across 863,000 keyword result pages and 4 million Overview URLs. 38% of cited pages also rank in the top 10, and 31.0% sit beyond position 100. Most cited pages are not in the top 10 at all. An earlier run in July 2025 had found roughly 76%.

BrightEdge ran a different measurement and watched it strengthen. Across nine industries from May 2024 to September 2025, its own Generative Parser put overlap with organic rankings rising from 32.3% to 54.5%. Citations traced to the top 10 specifically accounted for 16.7%, most of the growth coming from positions 21 to 100. Put the two top-10 figures side by side and they cannot both be true, 38% and falling against 16.7% and rising.

A third figure circulates from a public relations release that names an analysis it never links. How it was produced stays unstated, so there is nothing to check. None of this is a rounding problem...each firm sells into the uncertainty it is measuring.

Four Click Numbers, Two Answers

Click evidence looks contradictory...most of it holds together. Ahrefs put the top-ranking page 34.5% lower on clickthrough when an Overview is present, drawn from 300,000 informational keywords split evenly between Overview and non-Overview sets, on Search Console desktop data for March 2024 against March 2025. Position one fell from 0.073 to 0.026 against a forecast of 0.040.

  • Later run: The same Ahrefs design on December 2023 against December 2025 returns 58% lower clickthrough.
  • Field data: Pew Research Center tracked 900 American adults across 68,879 searches during March 2025, logging 8% clicks with a summary against 15% without.
  • Zero-click: Semrush watched that rate fall from 33.75% to 31.53% on keywords that gained an Overview, over 200,000 keywords from January to November 2025.

Semrush and Ahrefs do not measure the same thing. Read the denominators. One counts the clickthrough rate of the top result, the other the share of searches producing no click at all...those two can move opposite ways in the same market.

Pew found the deflating numbers too, and Search Engine Land put a rough price on a citation.

  • Summary links: Clicks on the links inside a summary happened on 1% of visits.
  • Session end: 26% of pages carrying a summary ended the browsing session, against 16% without one.
  • Citation value: Across more than 20,000 ranking queries in September 2025, a first citation was worth roughly the clicks of organic position 6.

Cited And Wrong

Here is the risk with a price attached...courts on two continents have now taken it up. Wolf River Electric, a Minnesota solar company, sued Google after an AI Overview said the state Attorney General had sued it for deceptive practices. No such suit existed.

Its complaint itemizes the lost business...these are allegations in active litigation, not findings.

  • 3 March: $39,680.
  • 4 March: $26,400.
  • 5 March: $150,000.
  • 11 March: $147,400 plus $26,644.12.

The complaint also alleges that Google "cited numerous sources in support of its false assertions" and that none of the referenced materials in fact contained the information Google claimed they did. Eugene Volokh wrote on the filing on 11 June 2025. In January 2026 a judge remanded the case to Minnesota state court.

In Munich a regional court granted two publishers a preliminary injunction in May 2026, finding that Overviews generate "independent, new, and substantive statements" rather than listing third-party content. No publisher can stop the repetition...only Google can modify the models producing them.

So a business can be the cited source...the text above it states a price, a policy and/or a timeline that page never gave.

What Search Console Reports

Search Console documentation reports less than the announcements suggested. It states that "an AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position"...the position number describes the box, not the page inside it.

  • Clicks: Following an external link out of an Overview counts exactly as it always did.
  • Impressions: A link counts only once it has been scrolled or expanded into view.
  • Labs: Search Console excludes data from experiments still running in Search Labs.

Announced in June 2026, the generative AI performance report covers AI Overviews and AI Mode...impressions are the metric. No clicks, no clickthrough rate and no position are documented, the two surfaces cannot be separated from each other, and it is still rolling out to a subset of site owners.

None of that reaches the other half of the question. Those reports cover Google's own surfaces...whether ChatGPT or Claude name a business sits outside their reach. vizmeta HQ is a free tool for that gap, checking monthly and logging the answers.

Worth Doing, Not Worth Doing

Documented levers are few and fairly dull.

  • Indexing: Eligibility is being indexed and eligible for a snippet, and nothing beyond that.
  • Snippet tags: The four controls Google names are nosnippet, data-nosnippet, max-snippet and noindex.
  • Question pages: Informational and hybrid queries drew Overviews far more often than local ones did.
  • Monthly reading: Run the priced and comparison queries and read what the summary actually says.

What the evidence leaves unsupported runs longer.

  • Special schema: Google denies any is needed...the best-controlled test found no lift.
  • AI files: llms.txt sits on Google's own list of tactics to ignore.
  • One number: Prevalence at 25.11%, 15.69% and 68% measures different samples on different dates.
  • Single checks: A 2026 preprint found answers varying across runs, prompts and time.

What stays open is the part every vendor claims to have closed. Whether ranking first causes citation, whether a citation is worth what a click used to be, whether any of it survives the next model update...none of it is settled. Those two clickthrough runs do not share a window. Ahrefs moved its own between them, from March against March to December against December...the firm published that change openly, and the window is the thing to read before quoting the figure.