Query fan-out: 288 of 980 sites carried under half the home page's prose one click down

What Google announced on 12 November 2026

The post is Google Search gets a major upgrade with Gemini 4 Hydrogen, credited to Robby Stein, VP of Product for Google Search, and dated 12 November 2026. Its opening claim is availability: Hydrogen is now available globally in Google Search for all users, delivering what the post calls more helpful, visual and intuitive responses. The sentence that carries the retrieval change reads that Hydrogen's advanced reasoning supercharges our query fan-out technique in Search, conducting simultaneous queries across all the relevant subtopics and connecting you to a broader range of helpful links than before. A second sentence repeats the direction: the improved reasoning capabilities connect you to a wider range of relevant websites and helpful links.

Three things are worth noting about what the post does not do. It attaches no number to the broadening, so there is no base, no comparison period and no way for a site owner to tell whether a wider range means five more links or five hundred. It does not mention AI Overviews anywhere, naming AI Mode only once and in connection with languages rather than retrieval. And it is an announcement rather than documentation, which matters because the two are maintained differently. This blog has read Google's link announcements before, when Google began placing more links inside its AI answers rather than only beneath them, and it read the reporting side of the same week when Google documented one Referer value for all of Search. The pattern is consistent: the surface changes are described, and the measurement is left to the site owner.

What is query fan-out, and which crawler serves it?

Google defines the mechanism in its own Search Central documentation rather than in the announcement. The page on AI features and your website describes fan-out as issuing multiple related searches across subtopics and data sources, and says that is what allows Google to show a wider and more diverse set of links alongside a generated response than a classic web search shows. The same page settles the question site owners actually need answered, which is who comes knocking. It names Googlebot as the crawler, and states that robots.txt directives for Googlebot are the control site owners have over access. Google-Extended, on that page's own account, governs AI training and grounding in some of Google's other systems rather than Search, and nosnippet is offered as the way to limit what a page shows in Search.

That has a practical consequence worth stating plainly, because it is the opposite of what a lot of advice implies. There is no fan-out crawler to allow or block. Access for this retrieval path is decided by the same rules that have always governed an AI crawler of Google's in your robots.txt, and the snippet directives decide how much of the page may be shown once it is retrieved. Lantad measured those directives across the corpus in October, when 132 of 1,083 home pages sent the directive that governs AI Mode and 3 limited anything. The retrieval consequence is the part this post is about, and it is also the part that has been visible in third party research for months: in the study behind nearly a third of AI Overview citations not appearing on page one, a fan-out across subtopics is the mechanism that best explains a citation pool diverging from the ranked result. What that means for a site is covered at more length in the platform guide on how to get cited in Google AI Overviews.

How we measured what sits below the home page

The corpus is the stratified sampling frame this repository commits for its crawlability study: 392 hostnames across ten platform strata and 1,027 across eight industry strata, 1,419 in total. Each host was requested once at its root. 1,078 answered 200 with an HTML content type and are the denominator for everything that follows; of the 341 that did not, 229 answered 403, 55 answered 503, 33 failed at the transport, 8 answered 202, 6 answered 429, 5 answered 404 and one each answered 400, 401, 406, 498 and 500.

Interior pages were chosen mechanically rather than editorially, because an editorial choice would decide the result. Every anchor in the home page's own HTML was read in document order, reduced to those resolving to the same host with a path deeper than the root, with asset extensions, Cloudflare's /cdn-cgi/ paths and duplicate paths removed, and then the first, the middle and the last of that deduplicated list were requested. 89 of the 1,078 home pages offered no usable internal link at all, which is the same defect Lantad reported when 84 of 1,091 home pages offered a crawler no path at all; the median home page offered 48. That left 989 sites, 2,936 interior requests and 2,843 readable responses: 901 sites returned all three pages, 61 returned two, 18 returned one and 9 returned none, giving the 980 sites that carry every comparison below. Among the 93 failed interior requests were 35 answering 403, 26 answering 404 and 7 answering 429.

Words are counted the way the product counts them, not by stripping tags. Each response was passed through the same extractor the scanner uses, and the figure taken is the main text, which is the corpus behind prose parity and excludes navigation, headers, footers and asides. The method behind that split is written up on the methodology page. One incidental finding fell out of reading every anchor: 206 of the 1,078 home pages carried at least one href that is not a URL reference at all but a template placeholder, a fragment of script or unescaped markup, 1,660 of them in total.

How much prose a site loses one click down

Across the 980 sites the median home page carried 589 words of main text, and the median of each site's interior pages carried 461. The direction is not uniform and the counts say so: 536 sites carried less main text below the home page, 404 carried more and 40 came out level. The size of the fall is the more useful figure. 288 of the 980 carried under half the home page's word count on their interior pages. 100 sites came in under 100 words, 8 of them at zero, and 44 sites paired a home page of 300 words or more with interior pages under 100. On 27 sites every interior page that answered, and at least two did, came in under 100 words; 13 of those 27 had a home page that cleared 100 and 6 had one that cleared 300. The single sharpest case in the corpus paired a home page of 3,785 words with three interior pages of 42, 53 and 26, a site that puts everything it has on the one page a subtopic search is least likely to pick.

Put beside the home page figures, the distribution shifts down rather than collapsing. 192 home pages carried 1,200 words or more against 130 sets of interior pages, and at the thin end 96 home pages came in under 100 words against 100 sites' interior pages. The thin tail is not new ground here. Lantad has measured it directly on page types fan-out is likely to reach, finding that 43 of 220 documentation pages gave a crawler under 100 words and that 418 of 1,419 sites gave a crawler no words at all on the page behind Google's supplier unit. What the figures here add is the comparison: the page a site is judged on by a visitor, and the pages a subtopic search is likelier to reach, are not the same page, and on more than half of this corpus the second is the thinner one. That gap is what an AI visibility measurement has to look at rather than past.

Which platforms and sectors thinned out most

The corpus is stratified, so the question of where this concentrates has an answer. Sites built on Wix or Squarespace were the worst affected stratum by a clear margin: 54 sites, a median of 302 words at home against 141 one click down, and 21 of the 54 with interior pages under 100 words. Ecommerce was next with 16 of 53, then no-code builders with 3 of 19. News sites show the largest absolute fall without the thin tail, dropping from a median of 1,550 words at home to 464, with 50 of 59 carrying less prose below the home page and 9 under 100 words; a newspaper front page is a dense index of headlines, and its article pages are long, so a per-site median across three mechanically chosen links lands on section and category pages rather than articles. That is a limit of the sampling, not a finding about newspapers.

Two strata moved the other way and are worth reporting for that reason. Static documentation sites carried a median of 320 words at home and 441 below it, and finance sites 544 against 593, both gaining prose as a crawler goes deeper, which is what a site organised around content rather than around a shop window looks like. Shopify storefronts also came out roughly level, 549 against 556, with 2 of 34 under 100 words, which is better than the stack's reputation and consistent with what the Shopify fix guide describes. Framer sites had no interior page under 100 words at all. Where a stack renders its interior text in the browser rather than on the server, a crawler reading the HTML sees the shell, which is the effect Lantad measured when JavaScript supplied 7.6 percent of the prose and all of it on 11 of 271 pages. Counting what a site declares against what it serves is the same exercise as finding 985 linked pages against 200,712 declared in sitemaps.

What this measurement does not show

The largest limit is the one that cuts against the finding, so it goes first. Googlebot renders JavaScript and this measurement did not. Every figure above is what a crawler reading the delivered HTML receives, which is the right model for a retrieval path that reads text rather than running an application, and it is the wrong model for Googlebot specifically. On the Wix and Squarespace stratum, where the median interior page carried 141 words, some of that prose almost certainly arrives when a browser runs the page. The honest reading is that these counts bound what is available without rendering, and the gap between that and the rendered page is itself the thing worth checking on your own site, which is what what GPTBot sees exists to show.

Four smaller limits. LantadBot is not Googlebot, and 229 of the 1,419 hosts answered our request with a 403, so the sites most willing to refuse an unfamiliar agent are underrepresented in every figure; whether a given rule would stop Google is a separate question that the robots.txt tester answers per crawler. Three pages chosen by document order are not a site, and they are certainly not what fan-out selects: Google picks pages by relevance to a subtopic it generated, and no external party can see that choice. A page can also answer 200 and still be the wrong page, which is the failure mode behind 125 of 1,094 sites answering a URL that does not exist with HTTP 200, so some thin interior pages here may be soft errors rather than thin content. And a word count is not a grade: the full scoring Lantad applies, in which 972 of 1,419 corpus sites were graded and parity took 60.7 percent of every point they lost, weighs more than length.

One last thing is a record rather than a measurement. The engine registry in core/src/engines.ts lists eight answer engines and marks three of them as exposing the searches they ran; the Google AI Overviews surface is not one of the three. That is a setting in this repository describing what each vendor publishes, not a finding about Google. But it is the reason a post like this has to measure a site's own pages: when the engine that just widened its fan-out shows nobody which queries it fanned out to, the only observable side of the exchange is what your server sent.