SEO remains foundational for Google AI. This is the historical statement I read in its documentation on August 7, 2026: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” It rejects the need for optimization tricks; current eligibility also depends on the technical criteria and inclusion control explained below. Yet the statement is insufficient as a general answer for three checkable reasons: its scope covers one of the four engines I measure; it addresses production, not measurement; and the other three do not draw on Google’s index.
I am Mateus Gomes, operator of murmur.marketing, a SWAS GEO (Generative Engine Optimization) operation in Brazil. This page belongs to the series on whether GEO is a fad or actually works. What distinguishes it is where the objection comes from: neither an independent critic nor a paper, but the owner of one of the surfaces, stating in official documentation that no special work is necessary. That is as authoritative a source for an objection as one could ask for. It deserves an answer about scope, not a louder argument.
Summary
- Disclosure before the figures. We deliver GEO as SWAS, combining proprietary measurement software with specialist execution. The August 6, 2026 baseline predates the current guide library and does not assess today’s operation.
Disclosure of interests, before the figures
Murmur delivers GEO as SWAS: proprietary software for measuring and tracking SOV, paired with specialists in strategy and execution. This commercial interest does not change the documentation’s scope: SEO fundamentals remain relevant, and measuring answers across different products addresses a complementary need.
On August 6, 2026, we ran 100 Portuguese questions across four engines, with 800 captures, all with screenshots. The classifier recorded zero qualifying appearances, and the textual count outside the ten brand-naming questions was 0/712. This historical baseline predates the current guide library. It is a methodological reference, not an assessment of today’s SWAS strategy.
1. What Google literally says — and why it is right
Google Search Central’s documentation on AI features is brief and direct. These are the two relevant statements, as I read them on August 7, 2026:
“There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”
“robots.txt directives for Googlebot is the control for site owners to manage access to how their sites are crawled for Search.”
The implication offers no comfort to my side of the argument: Google’s AI surfaces use the same infrastructure as Search. The crawler is Googlebot; access is governed by Googlebot’s robots.txt directives. Google-Extended serves a different purpose, controlling use of content in the training of other systems, rather than inclusion in AI Overviews.
Eligibility update: special schema or a magic file is not required to improve rankings. That does not remove access requirements. The current guide requires a page to be indexed and snippet-eligible; the Search Console Search generative AI control, available worldwide since August 31, 2026, must permit inclusion. Its default is Include. Exclusion affects eligibility for AI Overviews, AI Mode and generative features in Discover. This is a participation control, not a ranking signal or a guarantee of display.
I will concede more than the statement requires. On August 6, 2026, I examined the public site structures of the most frequently cited Brazilian GEO names using raw curl responses. The result supports Google’s argument: the leading name has nine pages in its GEO directory within a sitemap of 1,366 URLs, or 0.7%, and publishes less than one GEO post per month. Its accumulated presence lies elsewhere: original research hosted by Poder360 and E-Commerce Brasil, two Band articles placing it first among ten, and its founder on five third-party podcasts. That is entity authority and public relations, not an AI technique. The original audit’s caveat remains: the Band listicle looks like syndicated PR placement; whether it was paid could not be confirmed.
2. Scope: Google speaks for Google, one of four engines
The statement covers AI Overviews and AI Mode. I measure four engines, and three do not pass through Mountain View.
| Engine | Index or crawling it draws on | Renders JavaScript? | What Search Console sees | What browser capture sees |
|---|---|---|---|---|
| Google AI Overview | Google’s index, served by Googlebot | Yes | Property visibility in the Generative AI report; not complete answer text | Answer text and the domains it cites |
| ChatGPT | Search grounded in Bing’s index plus OpenAI crawling | No — GPTBot · OAI-SearchBot | Nothing | The same, with a screenshot for each capture |
| Perplexity | Its own index and crawling | No — PerplexityBot | Nothing | The same |
| Claude | Its own crawling, with substantial reliance on third parties | No — ClaudeBot · Claude-SearchBot | Nothing | The same |
ChatGPT’s row has the largest practical implication: its search is grounded in Bing’s index. That makes Bing Webmaster Tools an overlooked lever in the Brazilian market, absent from many discussions about “AI optimization” because everyone is looking at Google.
The strongest technical difference has an external source. Across approximately 1.3 billion AI-crawler requests in one month, about 28% of Googlebot’s volume on the same network, Vercel reports that the major AI crawlers observed did not render JavaScript. They download JS files without executing them: 11.50% of requests for ChatGPT, 23.84% for Claude. Googlebot renders them.
The consequence is concrete: a site whose content exists only after JavaScript runs is potentially renderable by Googlebot, but unavailable through the non-rendering crawler paths observed. On this particular requirement, SEO and GEO differ materially. Google’s documentation remains correct because it never claimed to describe those other engines.
3. Search Console already measures AI — what answer capture adds
Production guidance and measurement capabilities need separate consideration. Google’s current guide explicitly recommends Search Console’s Generative AI report.
Updated September 23, 2026: Search Console’s Generative AI report has been available worldwide since August 31, 2026. It shows impressions and URLs in AI Overviews, AI Mode and generative features in Discover, with countries and trends over time; device data is available for Search results.
It does not provide every answer’s complete text, a controlled prompt-level run or SOV across Google, ChatGPT, Claude and Perplexity. Independent capture answers a different question: which brand was recommended, cited or mentioned within a declared question universe? In murmur’s SWAS, official data complements diagnosis, interventions and SOV tracking; it is neither discarded nor presented as equivalent to cross-engine measurement.
Our own named case shows that the engines are not interchangeable. In Fly Vet, my veterinary marketing agency, the June 2026 campaign used 100 questions per engine and the citation_kind classification rule. ChatGPT recommended it in 30/100, Claude in 41/100, and Google recorded zero recommendations in this sample, while using it as a source in 81/100. The same company and period, measured by the same rule, produced three different behaviors.
These figures justify measuring each surface and distinguishing citation from recommendation. They do not prove SEO irrelevant or identify why the engines differed. In that June campaign, Google most frequently used the site as a source (81 of 100) and recorded zero recommendations; this describes the sample, not the product’s current or permanent behavior.
Server data also shows that these surfaces exist, with a mandatory caveat. On a client subdomain, a log recorded 1,074 requests over roughly 3.6 days, from June 10–13, 2026: Googlebot 327 · OpenAI approximately 184 — OAI-SearchBot 71, GPTBot 54, ChatGPT-User 59 — PerplexityBot 62 · ClaudeBot 9 · bingbot 1. These are user-agent counts, not verified-IP counts. On another subdomain, checking actual IPs reduced a raw 88 ClaudeBot hits to 79, 10 OpenAI hits to 3, and 6 Perplexity hits to zero. That caveat must travel with the counts.
4. What I measure — and what I do not
Challenging Google’s scope must not become an overclaim about my own:
- The pipeline’s
googleengine captures AI Overview. I do not measure the Gemini app; this pipeline has no Gemini harvester, and this page makes no claim about it. - Our August 6, 2026 campaign had no treatment arm. It measured
murmur.marketingwith zero published pages. It supports the statement “the baseline suggested entity ambiguity”, not a general theory about content. - Our competitor detector covers 29 names out of at least 447 extracted by the judge from the same 800 captures. That is why we publish absolute counts and denominators rather than shares normalized across competitors.
5. Where I concede more than Google asks
The literature pushes Google’s argument further than the company does. Olivier Martinez’s critical survey, arXiv 2607.14035, submitted on July 15, 2026, reviews 45 studies and concludes that none of the reviewed techniques demonstrates a stable, longitudinal, cross-platform causal effect. It also finds that topic relevance and position in context are the most reproducible levers, generic heuristics transfer poorly, and citation-oriented rewriting can harm retrieval.
Read that alongside Google’s statement and the fit is uncomfortable for my industry: topic relevance is exactly what search practitioners already advocated. Much of what is sold as an AI-specific technique either renames that principle or can actively harm performance. Google’s statement is closer to the published evidence than most GEO vendor decks, without invalidating diagnosis, execution or outcome tracking.
The operational distinction does not depend on an absence of official data: Search Console measures property visibility on Google’s surfaces. Complementary measurement captures answers, separates recommendations from citations and compares engines within a defined universe. SWAS connects those observations to prioritization and execution, without claiming access to Google’s internal ranking systems.
6. How to test the statement without relying on me or Google
Three procedures can be checked in an afternoon:
- Open Google Search Central’s AI-feature documentation and read the full statement in context. It concerns AI Overviews and AI Mode, and it means what it says.
- Ask the same purchase question from your sector five times on each of the four engines, starting clean chats and saving every screenshot. Differences can arise from indexes, source selection, models or session state. In our 100-question, four-engine campaign, the hiring family contained 48 captures: the engines disagreed about the leading name, and 8/48 answers contained none of the 29 brands on the declared list.
- Disable JavaScript in your browser and open your site. What remains is approximately what GPTBot, ClaudeBot and PerplexityBot take away.
If those tests give you the same result throughout, Google’s statement has answered your whole question, and you do not need me. If they give different results, the issue concerns coverage and measurement instruments, rather than a special page technique. That difference can be measured.
Frequently asked questions
Is Google lying when it says AI Overviews needs no special optimization?
No. The quoted statement was read on August 7, 2026 and rejects special optimization techniques. Current eligibility requires an indexed, snippet-eligible page that is not excluded by Search Console’s Search generative AI control. That control defaults to Include; it governs participation, not ranking, and does not guarantee display. Search Console also measures visibility in Google’s generative features.
If Google’s page guidance is unchanged, what changes outside Google?
For Google’s surfaces, its documentation gives the same page-production guidance. The published evidence points in that direction too: the 45-study survey identifies topic relevance as one of the few reproducible levers. Outside Google, an externally measured technical difference matters: Vercel reports that none of the major AI crawlers renders JavaScript, whereas Googlebot does. A site that exists only after JS runs is potentially renderable by Googlebot, but unavailable through the non-rendering crawler paths observed. Search Console already reports Google AI visibility; complementary capture identifies answer text and brand states by question and engine.
If Google uses the same index, why measure the engines separately?
Because the observed behavior differs substantially. In Fly Vet’s June 2026 campaign, using 100 questions per engine and citation_kind, ChatGPT recommended the brand in 30/100, Claude in 41/100, and Google recorded zero recommendations in this sample, while using it as a source in 81/100. In that sample, the engine citing it most often recorded zero recommendations. The observed results were not interchangeable. Measuring only one can misrepresent the direction of the outcome, not merely its size.
Does Google-Extended remove my site from AI Overviews?
That is not its purpose. Search Console also provides the Search generative AI participation control, which defaults to Include. Google’s documentation separates the controls: access to Search surfaces, including AI Overviews and AI Mode, is governed by Googlebot’s robots.txt directives; Google-Extended controls content use in other systems. Blocking Googlebot to control AI also removes the site’s access to Search, trading a hypothetical gain for a definite loss. Our own scope is narrower too: we do not measure the Gemini app and have no Gemini harvester in this pipeline.
If ChatGPT relies on Bing’s index, does AI optimization become Bing optimization?
Partly. This is an overlooked lever in the Brazilian market: because ChatGPT’s search is grounded in Bing’s index, Bing Webmaster Tools matters even while most attention goes to Google. But that addresses one retrieval chain, not all four. Perplexity has its own index and crawling, Claude crawls independently, and ChatGPT also depends on OpenAI’s crawlers — GPTBot, OAI-SearchBot and ChatGPT-User. They do not execute JavaScript, consistent with Vercel’s observations of major AI crawlers: they download JS without running it, in 11.50% of ChatGPT requests and 23.84% of Claude requests, while Googlebot renders it. Search Console’s Generative AI report does not replace reading and classifying the answer for each question. In Fly Vet, with 100 questions per engine, ChatGPT recommended the brand in 30/100, Claude in 41/100, and Google recorded zero recommendations in this sample, but cited it as a source in 81/100.
Does Search Console’s Generative AI report replace SOV measurement?
No. It measures property visibility in Google’s AI features. Complementary measurement records answers and brand states for declared questions and compares engines. Use both according to the decision; do not interpret URL impressions as recommendations or share among brands.
Who wrote this
I am Mateus Gomes, operator of murmur.marketing, a Brazilian engine focused solely on GEO, rather than a full-service or SEO agency. I read Google Search Central’s AI-feature documentation on August 7, 2026, ran the August 6, 2026 campaign cited here, and conducted the named Fly Vet measurement.
I have a commercial interest as a SWAS GEO operator. The August baseline predates the current guide library and does not assess today’s strategy. The service connects diagnosis, interventions and SOV tracking; targets and any SOV-growth guarantees depend on the contract model, scope and conditions.
Conclusion
Google tells the truth within its scope. That scope is what the popular interpretation overlooks. SEO remains foundational for AI Overviews and AI Mode; also check indexing, snippet eligibility and the Search Console inclusion control. Do not confuse that control with a ranking signal or a secret optimization technique.
Beyond that scope, three checkable facts create a gap. ChatGPT, Claude and Perplexity do not draw on Google’s index and do not render JavaScript, whereas Googlebot does. Their observed behavior differs enough that, for the same company in the same period, one engine recommends it in 30/100, while another recorded zero recommendations in that sample and instead cites it as a source in 81/100. Search Console already measures Google AI visibility, but it does not replace prompt-level answer capture or cross-engine SOV. The approach is complementary: official data, diagnosis, SWAS interventions and outcome tracking. To discuss measurement for your case, with a denominator, execution date and screenshots, contact Mateus Gomes on LinkedIn.