--- title: "Four Pages Took 91.5% of Our AI-Answer Fetches. Here’s What They Had in Common." description: "In 30 days of groas.com edge logs, four pages drew 401 of 438 verified live-answer fetches. A closer look at their dates, numbers and answer-ready structure." image: "https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7542f4015c8bd75d15b79_83375a7f-5a4f-4346-acac-05258148f197.png" --- September 26, 2026 • min read # Four Pages Took 91.5% of Our AI-Answer Fetches. Here’s What They Had in Common. ![Young man with curly hair wearing a black shirt outdoors against green foliage background.](https://cdn.prod.website-files.com/6821efca072e48f6f495a47e/68562d390107b3921a6e3d68_1743932904108.jpg) **Alexander Perleman**, Head Of Product @ groas Ex-Goldman Sachs and Stanford Computer Science alex@groas.ai [LinkedIn](https://groas.com/post/four-pages-drove-91-of-our-ai-answer-fet#) ![Cover image for: Four Pages Took 91.5% of Our AI-Answer Fetches. Here’s What They Had in Common.](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7542f4015c8bd75d15b79_83375a7f-5a4f-4346-acac-05258148f197.png) 401 out of 438. That is how concentrated our AI-answer fetches were over 30 days of groas.com edge logs. Four pages took **91.5% of verified live-answer fetches**; the other 19 pages that got pulled at all split the remaining 37. Our blog did not get consulted evenly. A few pages did almost all the work. The pattern was not *best writing wins*. The pages that drew fetches made it easy to find a current claim, lift a specific number or follow a direct answer. That is useful if you are trying to get cited in ChatGPT, Perplexity or an AI Overview. It is also easy to overstate. An edge request can show that a page was fetched under conditions consistent with a live answer being assembled. It cannot show that the final answer quoted the page, displayed a citation or sent a reader back to us. I am not talking about raw bot hits, either. Those outnumbered the set below by about 40 to 1, and most tell you very little. A user-agent that says GPTBot is a claim, not proof. Here is what I counted, what the numbers say and where they stop. #### What counted as a live-answer fetch I split requests to the edge with an AI-crawler user-agent into two buckets: 1. **Raw hits:** every such request, including background crawls, prefetches, repeat pulls and requests whose user-agent could be spoofed. This is the noise bucket. 2. **Verified live-answer fetches:** requests that passed reverse-DNS and ASN identity checks, pulled full page HTML rather than robots.txt or a head request, and arrived in the timing and query pattern associated with live answer retrieval. These are the requests counted below. The source for every fetch count in this article is the same: groas.com edge logs from the 30-day window, filtered by that second rule and grouped by URL. I call the resulting requests *verified live-answer fetches* because the identity and request checks narrow the set substantially. The timing pattern still gives us an inference about purpose, not a recording of the finished answer. **A fetch is not a confirmed citation.** Keep that distinction in view, especially when a dashboard offers one tidy line labelled “AI citations.” Bucket one was roughly 40 times larger than bucket two. If I graphed all those raw hits and called the result visibility, I would be measuring an impressive amount of activity without knowing whether any page was even a candidate for a live answer. I used to make a version of that mistake with log files while managing ecommerce accounts: more crawls felt like more demand. It was a comfortable story. It did not tell me what a person saw. So the smaller bucket is the useful one for this analysis: **438 verified fetches across 23 touched pages**. It lets me compare where these requests went. It does not let me calculate how often an assistant cited groas, how many people saw an answer or what happened after they did. #### Four pages took 401 of 438 fetches Here is the URL-level breakdown from that filtered 30-day set. The shares use all 438 verified fetches as the denominator, not raw crawler traffic. | Page | Verified live-answer fetches | Share of 438 | | --- | --- | --- | | YouTube Ads in 2026: frequency and policy changes | 148 | 33.8% | | Google Ads benchmarks by industry | 122 | 27.9% | | GA4 changes and what to fix | 84 | 19.2% | | SaaS campaign structure guide | 47 | 10.7% | | All other 19 touched pages combined | 37 | 8.4% | *Source: groas.com edge logs, 30-day window; identity-checked, full-HTML requests matching the live-retrieval pattern, grouped by URL. Shares are rounded.* Those top four sum to 401. The first page alone drew 148; the entire 19-page tail drew 37. That is the shape of the finding. It is not a claim that the other pages are worthless. Some get crawled, and some rank in classic search. They just did not attract many requests under this particular filter during this window. I used to tell clients that a bigger blog meant more chances to be cited. In this data, the extra pages mostly meant more pages sitting idle while four did the work. The practical question is not “How do we publish more?” It is **“What can page five borrow from the four that got pulled?”** ![Bar chart showing four tall bars beside a row of much smaller bars for the remaining touched pages](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab754304015c8bd75d15bae_93222876-5a51-4e31-b3de-1b924eb8fc16.png) #### The common pattern was an answer you could lift The four winners do not read like our best writing. Two are dry reference pages I would not put in a portfolio. I looked instead at what a system trying to assemble a short answer could use without digging through six paragraphs of setup. Three traits stood out, with an important qualification: they appear in different combinations. This small set does not isolate which trait caused a fetch. ##### 1\. A current claim was easy to find The YouTube page puts 2026 in the title and deals with frequency and policy changes. The benchmarks page gives industry-specific figures; the GA4 page identifies changes and what to fix. These pages offer a reader, or an assistant, something more time-specific than “this topic matters.” The SaaS structure guide is the useful check on an overly neat theory: its draw is a concrete campaign-planning answer, not a 2026 headline. That distinction matters. **The lesson is to date claims that depend on a date, not to paste a year onto every title.** A current change is easier to identify when the page says when it applies. A structure guide can still be useful because it answers a narrow question with an example. I would rather make a genuinely current claim visible than manufacture freshness in a heading and leave the substance untouched. ##### 2\. Specifics survived compression The YouTube page names frequency-cap behaviour. The benchmarks page gives CPC and conversion ranges by industry. The GA4 page says what broke and puts the fixes in order. The SaaS guide gives budget splits, including how to think about brand versus non-brand testing for a $20k/month account. That is a different kind of material from “optimise your campaigns regularly.” A range can fit in a table row. A fix order can become a short list. A budget split can be repeated with its context intact. General advice usually needs paraphrasing before it becomes an answer; a well-labelled figure needs less work. **Make the useful unit easy to quote without making it easy to misunderstand.** A number without the industry, account type or date that gives it meaning is not a better answer. It is just a tidier mistake. I cannot tell from these logs whether any particular number appeared in a finished answer. I can say the frequently fetched pages contain specifics a short answer could carry, while many pages in the tail spend more room setting the scene. That is a pattern worth testing, not a citation tally. ##### 3\. The headings sounded like the questions The leading pages use headings close to the questions a buyer might ask: what changed with YouTube ad frequency in 2026, what a good CPC is for an industry, what broke in GA4 and how to fix it. I checked those headings against the questions implied by the fetch patterns. The match looked tighter on the four leaders than across the tail. Our opinion pieces use clever headings. The fetched pages tend to use obvious ones. If someone asks what changed, a heading that says *what changed* gives them a place to start. Retrieval can use that phrasing overlap, and a human skimming the page benefits too. I would not claim the heading alone won the fetch; the logs cannot separate it from the page underneath. But if the answer is buried under a witty label, rewriting the label is cheaper than commissioning another long guide. The larger point is about packaging. The winning pages offer extractable blocks: a definition, a dated fact, a small table or steps in order. The answer arrives before the argument about why the topic matters. **Put the answer where a skimmer can find it on the first screen.** Then use the rest of the page to explain the limits and the work behind it. #### What did not line up with fetches Word count did not give us a useful ranking in this set. Our longest guide in the window runs just over 3,200 words, with custom tables and screenshots. It drew 2 verified fetches. A 900-word benchmarks update drew 122. That is not proof that short pages beat long ones. It is a reason to stop treating length as a substitute for a findable answer. Backlinks did not neatly sort these pages either. The SaaS structure page has fewer referring domains than three opinion pieces that got zero pulls, yet it drew 47 verified fetches, or 10.7% of the total. Classic search traffic told a different story again: two of the top four pages sit outside our top ten for organic sessions. They are not simply our most-read pages. Publishing date was a poor shortcut for freshness. One of the four leaders was published months before the window. What still looked current was the material inside it: a 2026 reference, a current number or a fix for a problem the page still addresses. “Published recently” and “contains a usable current answer” are not the same test. I lined up the 19 touched pages in the tail against the four leaders the way I used to line up ad-copy losers against winners. The tail had more introduction, context and voice. The leaders got to the change, the figure or the fix sooner. I would not strip every opinion piece down to a table; some pages have a different job. But if a page is meant to answer a specific buyer question, its first screen should do that job before it asks for patience. Here are the limits in one place: - **Scope:** 30 days, one domain and 438 requests that passed the chosen filter. These are not results for the wider web. - **Outcome:** the logs show filtered fetches, not confirmed citations, answer wording or revenue. - **Cause:** the comparison is observational. Dates, figures, headings and topic demand can move together; this set does not prove which one drove an individual request. That does not make the concentration uninteresting. It tells us what decision the evidence can support: test the structure of the pages being fetched. Do not promise that copying their topics, or adding a year to a title, will reproduce their counts on another site. #### Rewrite one answer block, then check the logs Pick one existing page that ought to answer a question your buyers ask. Rewrite the top, not the whole article: headline, first 100 words and first table or list. Leave the rest alone for this pass. It gives you a manageable change to inspect rather than a new post whose subject, length and format all changed at once. Work through the page in this order: 1. **Name the question plainly.** Use the wording a buyer would recognise, not the clever version. If the answer concerns a particular year or change, say so in the heading. Do not add a date to an undated claim merely for decoration. 2. **Answer before explaining the background.** Open with what changed, what the figure means or what to fix. Move the “why this matters” paragraph below the answer if it still earns its place. 3. **Give the specific a usable label.** Put a cap, CPC range, conversion range or budget split beside the context needed to interpret it. If there is an order to the fix, number the steps. If there is a comparison, use a small table. 4. **Read the block on its own.** Could someone lift the first answer without losing the condition that makes it true? If not, the block is not finished. A quotable mistake is still a mistake. That is smaller than a content strategy deck and more useful than changing every title on the site to include “2026.” It also responds to what the logs actually show. Four pages drew most of the filtered requests. Their first screens give a reader something to use. The test is whether making another page similarly direct changes its fetch pattern, not whether the edit sounds like *AI optimisation*. ![Close-up of hands editing a printout to add a date, a number and a direct question heading](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab754304015c8bd75d15bb1_83a5610c-cfd8-454c-93d4-bcdf40255e3d.png) Most teams asking how to get cited do the steps backward. They buy a tracker that graphs mentions across ChatGPT, Perplexity and AI Overviews, then wonder why the line never moves. Tracking never moved CPA for me when I ran ads either. It told me where to look. The page still needed work. If you are assessing a tool meant to help you get featured in Google AI Overviews, judge it by that work, not just the dashboard. Can it make the answer readable, fix a block a crawler cannot use and show you what changed afterward? That is the job [groas Earned Search](https://groas.com/earned-search) is built for: improve the page, make the answer citable, then inspect the fetch change. Monitoring tells you the four pages won. Someone still has to build page five. ![Overhead photo of a desk with two reports, one showing a circled monitoring graph and the other a circled edited page draft](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab754304015c8bd75d15bc7_61097040-2c5e-484b-859e-eec0d4b1ea2d.png) The decision the numbers support is narrow. Do not publish 20 new posts this month just to make the blog bigger. Pick three questions your buyers already ask an assistant and give each one a page with a direct answer near the top. Where the question calls for it, add a real date and a contextualised number. I would start with pricing or benchmarks, a what-changed post for 2026, and one fix-it guide with steps in order. Publish the edited block, wait 14 days, then check verified fetches again using the same filter. Keep what earns pulls; revisit what does not. That check will not tell you whether an answer cited the page, but it will tell you whether the page joined the set being retrieved. For a longer rewrite checklist, use this [guide to getting cited in Google AI Overviews](https://groas.com/post/how-to-get-cited-in-google-ai-overviews). **Make the answer current where currency matters, specific where a number helps and easy to find.** Everything else is editing for applause. ## Related Posts [![Cover image for: Why AI Assistants Cite Some Businesses and Skip Others](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7546d364bd9c55c3ad111_aa7733cf-8698-4d54-ace8-b0b1128cd278.png) ##### Why AI Assistants Cite Some Businesses and Skip Others A first-principles look at how ChatGPT, Perplexity and Google AI Overviews find sources, choose passages and cite businesses, using one running example you can apply to your own pages. September 26, 2026 • min read Written by](https://groas.com/post/how-ai-assistants-decide-which-business) [![Cover image for: Agency AEO Tools: The Questions to Ask Before You Buy Another Dashboard](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7545fb9190b5f6646c28d_88af2294-213d-46f9-9b31-9b2630ccb221.png) ##### Agency AEO Tools: The Questions to Ask Before You Buy Another Dashboard Blunt answers for agencies buying AEO software: what prompt tracking measures, what delivery requires, and how to judge white-label tools by the work they remove. September 26, 2026 • min read Written by](https://groas.com/post/reader-q-a-the-aeo-questions-agencies-ke) [![Cover image for: The AEO Sales Call: $4,000 a Month to Watch Your Citations](https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7542a4015c8bd75d15a0a_2ddcec87-223f-4c72-8582-fe300ec8aa2d.png) ##### The AEO Sales Call: $4,000 a Month to Watch Your Citations A satirical AEO sales call translates prompt tracking, citation networks and monthly reports into the work an agency will actually do. September 26, 2026 • min read Written by](https://groas.com/post/the-aeo-sales-call-transcribed-how-your) ![](https://cdn.prod.website-files.com/6821efca072e48f6f495a47e/6823c03839b2b618430d5ae6_Light-p-500.png) [![White stylized owl eyes with green background icon.](https://cdn.prod.website-files.com/6821efca072e48f6f495a47e/6a68a036a5dd5627409d3579_Groas%20Icon.png)](https://groas.com/post/four-pages-drove-91-of-our-ai-answer-fet#) [contact](mailto:scale@groas.ai) Explore [Home](https://groas.com/)[Philosophy](https://groas.com/our-philosophy)[Blog](https://groas.com/blog) What We Do [Paid Search](https://groas.com/paid-search)[Earned Search](https://groas.com/earned-search) Who We Serve [Businesses](https://groas.com/for-businesses)[Agencies](https://groas.com/for-agencies) MCP [ChatGPT](https://chatgpt.com/plugins/plugin_asdk_app_6a9c2d7108f08191beb5b3d736be8655?q=groas)[Claude](https://claude.ai/directory/mcp-groas-ai) Legal [Privacy Policy](https://groas.com/legal/privacy-policy)[Terms of Service](https://groas.com/legal/terms) Get Started [Apply](https://groas.typeform.com/to/xC1bQNUT) © 2026 groas 🇺🇸 [![](https://cdn.prod.website-files.com/62434fa732124a0fb112aab4/62434fa732124a389912aad8_linkedin%20small.svg)](https://www.linkedin.com/company/groas/) ```json { "@context": "https://schema.org", "@graph": [ { "@type": "", "@id": "https://www.groas.com/post/four-pages-drove-91-of-our-ai-answer-fet#article", "headline": "Four Pages Took 91.5% of Our AI-Answer Fetches. Here’s What They Had in Common.", "description": "In 30 days of groas.com edge logs, four pages drew 401 of 438 verified live-answer fetches. A closer look at their dates, numbers and answer-ready structure.", "url": "https://www.groas.com/post/four-pages-drove-91-of-our-ai-answer-fet", "image": { "@type": "ImageObject", "url": "https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7542f4015c8bd75d15b79_83375a7f-5a4f-4346-acac-05258148f197.png" }, "thumbnailUrl": "https://cdn.prod.website-files.com/6823bbd57170ea42b357cf81/6ab7542f4015c8bd75d15b79_83375a7f-5a4f-4346-acac-05258148f197.png", "datePublished": "2026-09-26T05:12:17.520Z", "dateModified": "2026-09-26T05:12:17.520Z", "inLanguage": "en-US", "isAccessibleForFree": true, "articleSection": "", "keywords": "Google Ads, ", "mainEntityOfPage": { "@type": "WebPage", "@id": "https://www.groas.com/post/four-pages-drove-91-of-our-ai-answer-fet" }, "author": { "@id": "https://www.groas.com/author/#person" }, "publisher": { "@id": "https://www.groas.com#organization" }, "about": [ { "@type": "Thing", "name": "Google Ads" }, { "@type": "Thing", "name": "Pay-per-click advertising" }, { "@type": "Thing", "name": "Performance marketing" } ], "mentions": [ { "@type": "Organization", "name": "groas", "url": "https://www.groas.com" }, { "@type": "Thing", "name": "Google Ads" } ], "speakable": { "@type": "SpeakableSpecification", "cssSelector": ["h1"] } }, { "@type": "Person", "@id": "https://www.groas.com/author/#person", "name": "", "description": "", "jobTitle": "", "email": "", "url": "https://www.groas.com/author/", "image": { "@type": "ImageObject", "url": "" }, "sameAs": [ "" ], "knowsAbout": [ "Google Ads", "Performance Max", "AI Max", "Pay-per-click advertising", "Conversion tracking", "Bid management", "Search advertising" ], "worksFor": { "@id": "https://www.groas.com#organization" } }, { "@type": "Organization", "@id": "https://www.groas.com#organization", "name": "groas", "alternateName": "groas.com", "url": "https://www.groas.com", "logo": { "@type": "ImageObject", "url": "https://cdn.prod.website-files.com/6821efca072e48f6f495a47e/6834a970e03a7016560e3515_Logo%20Design%20256x256.png" }, "description": "Dedicated strategists run your account and a proprietary engine trained on $500B in profitable ad spend optimises the execution underneath them", "sameAs": [ "https://www.linkedin.com/company/groas" ], "knowsAbout": [ "Google Ads", "Performance Max", "AI Max for Search", "Pay-per-click advertising", "Autonomous campaign management", "AI advertising agents" ], "areaServed": "Worldwide" }, { "@type": "BreadcrumbList", "itemListElement": [ { "@type": "ListItem", "position": 1, "name": "Home", "item": "https://www.groas.com" }, { "@type": "ListItem", "position": 2, "name": "Blog", "item": "https://www.groas.com/blog" }, { "@type": "ListItem", "position": 3, "name": "", "item": "https://www.groas.com/category/" }, { "@type": "ListItem", "position": 4, "name": "Four Pages Took 91.5% of Our AI-Answer Fetches. Here’s What They Had in Common.", "item": "https://www.groas.com/post/four-pages-drove-91-of-our-ai-answer-fet" } ] } ] } ```