Google's generative AI guide names five GEO tactics you can ignore

What Google published, and when it published it

The document is Google's guide to optimizing for generative AI features, and it describes itself as being for website owners looking for official best practices from Google Search on how to succeed in generative AI features in Google Search, naming AI Overviews and AI Mode as the examples. That framing matters for how much weight to give it. This is not a Googler's conference answer or a post on a company blog. It is reference documentation on the same site as the robots.txt specification and the crawler pages, which means it is the thing Google will point at when asked what it recommends.

Dating it takes a little care, because two dates on the page disagree about what changed. Google's documentation changelog records an entry for 15 May 2026 titled Adding a new guide on optimizing for generative AI features, whose description names mythbusting common AEO and GEO misconceptions among the new sections. A second entry, dated 15 June 2026 and titled Clarifying guidance on llms.txt files, records a note being added to that same guide. The page footer, when the page was read on 1 August 2026, carried the line Last updated 2026-07-10 UTC.

Those are not the same claim. The changelog's entry for 10 July 2026 concerns the canonicalization troubleshooting guide and says nothing about generative AI features, so the footer timestamp is evidence that the page was touched, not evidence that its substance changed in July. The honest reading is that the guide is a May document with a June addition, still current at the beginning of August. Anyone reporting that Google updated its AI guidance in July is reading a footer as though it were a changelog entry, which is a page timestamp doing work it was never built to do.

One neighbouring entry is worth naming because it explains the tone of the guide. On 5 June 2026 the changelog records adding Google Search's guidance on using third-party SEO tools, services, and advice. The AI guide points at that page from several places, including the paragraph about AEO and GEO and the section on measuring in Search Console. The two documents were built to be read together, and the second one is about the people selling the first one's subject matter.

What the guide says a page must have to be eligible at all

Strip the advice out and the guide contains one hard statement of eligibility, and it is short. Under the heading about technical structure, the first bullet reads that to be eligible to be shown in generative AI features on Google Search, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. The companion page on AI features and your website puts the same condition in terms of a supporting link in AI Overviews or AI Mode and adds that there are no additional technical requirements beyond that.

Immediately after it, the guide adds a second condition that is easy to miss: in addition to the technical requirements for Search, a site must be included in Search generative AI features in Search Console to be eligible for display. That is not a property of your HTML. It is a setting in an authenticated account, which is precisely the point of an earlier post here about the AI Overviews opt out: a control that lives in Search Console cannot be observed by fetching the site, so no external scanner, this one included, can tell you its state. The guide has now made that setting a stated precondition rather than only a way out.

The third statement is the one most often left off the summaries. The guide says plainly that just because a page meets all requirements, best practices, and complies with the policies, doesn't mean that Google will crawl, index, or serve its content, and that indexing and serving aren't guaranteed. Eligibility is a floor, not a promise, and any tool that converts a checklist into a predicted outcome is adding a claim Google's own documentation refuses to make.

What the guide asks for around that floor is ordinary crawl-layer hygiene. It asks you to follow crawling best practices and says the reason directly, that Google Search generative AI models use publicly accessible, crawlable content to learn patterns and provide relevant, grounded responses. It links its JavaScript SEO best practices and says Google is able to process content within JavaScript as long as it isn't blocked, which is a materially different position from the one other archives take: Common Crawl's July archive was collected by a crawler that executes no JavaScript at all. Both statements are true of their own crawler, and a site that renders in the browser only is readable by one and not the other. That gap is what prose parity exists to measure.

The five tactics the Mythbusting section says to ignore

The section opens by naming the field. It observes that terms like Answer Engine Optimization (AEO) or Generative Engine Optimization (GEO) are common online, and that many suggested hacks aren't effective or supported by how Google Search actually works. Elsewhere the guide says that from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO. Whether you find that persuasive or self-serving, it is now the documented position, and it is a direct answer to the question answer engine optimization pages across the industry have been posing.

Five items follow. On files, you don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search including its generative AI capabilities, as Google Search itself doesn't use them. On chunking, there's no requirement to break your content into tiny pieces for AI to better understand it, and the guide adds that there's no ideal page length. On rewriting, you don't need to write in a specific way just for generative AI search, because AI systems can understand synonyms and general meanings. On mentions, seeking inauthentic mentions across the web isn't as helpful as it might seem. On markup, structured data isn't required for generative AI search, and there's no special schema.org markup you need to add.

Read the scope carefully, because it is narrower than the headline. Every one of those sentences is a statement about Google Search. The guide is explicit on the point where it would most matter, saying it is completely fine to create and maintain llms.txt files for other services or systems that use these files. So the guide rules a tactic out for one engine, and the engine it rules it out for is the one whose generative features are, by its own account, rooted in its core Search ranking and quality systems. It says nothing about ChatGPT, Claude or Perplexity, each of which fetches pages under its own AI crawler tokens and answers to nobody's documentation but its own.

There is also a limit worth stating in the other direction. A vendor saying it does not use a signal is strong evidence about that vendor and weak evidence about effect sizes in general, which is the gap the survey of 45 GEO studies spent its length on. Google's list is authoritative about Google. It is not a literature review.

Two of the five are things Lantad ships or scores

The first is llms.txt. Lantad publishes an llms.txt generator and a glossary entry on llms.txt, and it publishes them while already having said the file has no measured effect. The post on what the evidence says about llms.txt reported a third-party measurement of how rarely those files are fetched at all. Google's statement is a different kind of evidence about the same file: not how often it is requested, but whether one named consumer uses it. It says maintaining one will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them.

There is a sharper version of that awkwardness than shipping a free tool, and it is worth stating rather than leaving for a reader to find. Lantad's score does award points for the file. STRUCTURE_CHECK_WEIGHTS in core/src/config.ts lists llmsTxtPresent at 0.5 alongside six checks weighted 1, inside a Structure sub-score that the scoring weights put at 15 points of 100. Half a weight out of six and a half, applied to 15 points, is a little over one point. That is small on purpose and the comment above it says so, describing llms.txt as minor within Structure. It is still not zero, and Google has now stated that the consumer most readers care about ignores the file entirely.

Both of those figures are settings rather than findings, which is the distinction this blog keeps returning to. Nobody measured that llms.txt is worth one point of visibility. Somebody decided it was worth half a check, and a decision made before a vendor published its position is a decision worth re-reading afterwards. Neither source argues for withdrawing a free generator, because the guide itself allows that other systems read the file and a site owner may reasonably want one. Both argue against selling it as a visibility lever, and a tool that marks your site down heavily for lacking one is asserting something neither source supports.

The second is structured data, and here the tension is with Lantad's own scoring rather than with a free tool. The AI Visibility Score weights schema at 10 points of 100, alongside 50 for parity, 25 for access and 15 for structure. That is a configured weight in the scoring config, which makes it a decision somebody made rather than a finding anybody measured, and this blog has been strict about the difference before. Google's guide says structured data isn't required for generative AI search, and in the same bullet recommends continuing to use it because it helps with being eligible for rich results.

Those two positions are compatible, and pretending otherwise would be more comfortable than it deserves to be. Lantad's own reporting has argued the same limit from the other end: reading one page's JSON-LD two different ways produced a perfect score and a zero at once, and counting schema.org's published usage file found that only 16 types reach ten million domains while most of the vocabulary is almost unused. A 10 point weight on structured data is defensible as a decision about rich results and entity clarity. It is not defensible as a claim that markup buys you a place in an AI Overview, and Google has now written down that it does not.

Google's warning about third-party tools applies to this one

In its Search Console section the guide says something that a tool vendor quoting this page has an obvious incentive to leave out. It advises readers to be wary of third-party tools that promise ranking success or claim to use internal Google metrics, states that no third-party tool has access to Google's internal ranking or AI systems, and suggests using such tools where they help a workflow while evaluating their advice against official guidance.

Lantad is a third-party tool. It has no access to Google's ranking systems, no access to Google's AI systems, and no privileged data of any kind about how Google assembles an answer. What it has is the outside of a website: it fetches pages the way a crawler does, compares what a crawler receives against what a browser receives, parses robots.txt against a registry of published crawler tokens, and reads the markup that comes back. That is the whole of it, and the methodology page exists to say so at length.

The gap is wider than a disclaimer, and it has been documented here before. Lantad's crawler registry, read on 1 August 2026, holds fifteen tokens across nine vendors and Googlebot is not one of them, which is the subject of an audit of that exact blind spot. A scanner that does not model Googlebot cannot tell you anything specific about AI Overviews eligibility beyond what is true of crawlers generally. Saying that plainly costs a sentence of marketing and buys the only thing worth having, which is that the numbers this site does publish mean what they say.

There is a version of this warning that cuts the other way too. Google's own measurement surface for these features is the generative AI performance report in Search Console, which reports impressions rather than causes, and which the earlier post on the opt out already covered. First-party impression data is better evidence of presence than any external scan can produce, and it is still not an explanation. The published research on this subject keeps arriving at the same place: counting citations is not the same as measuring influence, and the stage most tools skip is the one that happens before either.

What you can check on your own site against this guide today

Take the guide at its word and most of its advice is not auditable by anybody. Nobody can scan your site and tell you whether your content provides a unique point of view, and the guide is right that this matters more than the rest. What is left is a small set of technical conditions, all of which sit in the crawl layer, and all of which can be checked in a few minutes without an account anywhere.

Start with whether a crawler gets your words. The guide's requirement is that content be crawlable and that JavaScript not block it, so the first question is whether the served HTML contains your prose or a loading shell. Fetching your page as a crawler does answers it, which is what the crawler view tool is for, and the stack guides go through the usual causes on specific frameworks, such as fixing a Next.js site where the content arrives after hydration.

Then check what your rules actually permit. A robots.txt file is evaluated per user agent, and the answer for one token says nothing about another, which is why the robots.txt tester evaluates a path against each published crawler separately. Google's requirement is about Googlebot, but the same file decides whether any other engine reads you, and the two layers that decide access do not always agree: an edge rule can block a request your file allows, which is the subject of the post on the two layers.

Last, be honest about the scoreboard. A page can pass every technical condition in this guide and never appear in an AI Overview, because Google says indexing and serving are not guaranteed, and can appear in one while your AI visibility dashboard shows nothing, because impressions live in an account you may not have connected. The value of a crawl-layer check is not that it predicts the outcome. It is that it rules out the failures that make the outcome impossible, which is the only part of this a scan was ever able to do. If you want the platform-specific version, the guide on getting cited by Google AI Overviews covers the same ground for that one surface.