Google just published the GEO playbook. It’s the SEO playbook.

On Friday, Google published a help document titled *Optimizing your website for generative AI features on Google Search*. It is the first time Google has formally written down what to do — and more importantly, what not to do — for AI Overviews and AI Mode.

The document is short. The myth-busting section at the end is shorter. And between them, they end about eighteen months of well-funded industry argument about whether GEO is its own discipline.

It isn't. Google just said so.

The interesting question is what happens next, because a sizeable chunk of the consulting industry, the tooling industry, and the conference-circuit industry has spent the last year selling something the platform now says doesn't exist. And the people selling it are not going to stop selling it just because Google published a help document. They will adapt. The pitch will change. The acronyms will multiply. The invoices will keep going out.

What Google actually said

The document covers what you'd expect from a Google SEO doc: write useful non-commodity content, organise it for readers, mind your technical foundations, follow crawling best practices, provide a good page experience. Nothing in that section would have looked out of place in 2018.

The interesting part is the myth-busting list. Google explicitly says you don't need:

• LLMs.txt files
• Other special markup
• Content "chunking"
• Rewriting content for AI systems
• Inauthentic mentions
• An over-focus on structured data

That list is, almost item for item, the GEO consulting playbook that's been doing the rounds at conferences and in agency decks since early 2025. Llms.txt was sold as the new robots.txt. Chunking was sold as the new heading hierarchy. Schema-everything was sold as the new on-page SEO. Brand mention services — many of which are functionally astroturfing — were sold as the new digital PR.

Google has now publicly labelled all of it as not required, and in the case of inauthentic mentions, actively risky.

The industry response is the tell

Read the reactions Barry Schwartz collected and you can sort the SEO industry into three buckets in about ninety seconds.

The GEO industry has spent eighteen months selling tactics the platform now says aren't needed. That's not a marketing problem. That's an existence problem.

The first bucket — Lily Ray, Aleyda Solis, Glenn Gabe, Pedro Dias — said variations of *finally*. These are the practitioners who have been arguing since AI Overviews launched that the fundamentals didn't change, that GEO was mostly a rebadge, and that the chunking and llms.txt advice was a distraction. The document validates a position they've held publicly for over a year.

The second bucket went quiet. These are the people who built businesses, courses, conference talks, and tooling on the premise that GEO is a distinct discipline requiring distinct work. There's no clean way to respond when the platform you optimise for publishes a document that contradicts your pitch. You either pivot or you double down.

The third bucket is going to double down. You can already see the early signals: *Google says one thing publicly and another in the algorithm*. *The document is for beginners, not advanced practitioners*. *Trust the data, not the documentation*. These are all things people will say, and some of them aren't even wrong in principle — Google's public guidance has historically lagged its actual behaviour — but the gap between what Google says and what Google does has rarely been *this* wide on a topic with this much commercial activity around it.

The GEO industry has spent eighteen months selling tactics the platform now says aren't needed. That's not a marketing problem. That's an existence problem.

Why this isn't a victory lap

If you've been reading this site for any length of time, you'll know I've been pretty consistent on this: most GEO tactics are SEO tactics with a new label, and the honest playbook is the SEO playbook with structured data and earned media on top. Lily Ray's 220-site case study analysis made the data case. The Conductor report made the organisational case. The Duane Forrester diagnosis framework made the strategic case.

So this should feel like vindication. It doesn't. Two reasons.

First, the document is correct about what doesn't matter, but it's noticeably thinner on what does. It says create non-commodity content with a unique point of view. Fine — true, but unhelpful in the same way *make better content* is unhelpful. The hard part of AI search isn't knowing that quality matters. The hard part is the measurement problem: working out which of your content actually gets cited, by which models, for which prompts, with what user behaviour downstream. Google's document doesn't touch this, because Google has no incentive to make AI search performance measurable in a way the rest of the ecosystem can act on.

Second, the document removes the bad answers without replacing them with a process. *Stop doing llms.txt. Stop chunking. Stop chasing brand mentions inauthentically.* OK. Now what does an in-house marketing team actually do on Monday morning? The honest answer is *the same SEO work you should already be doing, with more attention to brand-building and original research,* which is correct but commercially difficult to sell as a programme of work to a board that has been told for a year that GEO is a new discipline requiring new budget.

What this means for the tooling layer

A significant portion of the AI search tooling category was built on the premise that AI search requires new instrumentation, new optimisation targets, and new content workflows distinct from SEO. Some of that tooling is genuinely useful — citation monitoring, prompt-response tracking, log-file analysis surfacing AI crawler behaviour. Some of it is rebadged keyword tracking with a coloured ChatGPT logo.

The genuinely useful tools are fine. Citation monitoring is going to matter more, not less, as AI search share grows, and the people building proper measurement infrastructure have a long runway ahead of them.

The rebadged tools are in a harder position. If Google has just said you don't need to chunk content, don't need llms.txt, and shouldn't over-focus on structured data, the value proposition of a tool whose primary feature is *audit your content for AI-readiness using our proprietary chunking algorithm* gets thinner by the week. Some of these tools will quietly pivot. Some will keep selling to buyers who haven't read the document.

The market will sort it out eventually. It just won't sort it out quickly. There is a lag between when a platform publishes guidance and when the consulting and tooling layers downstream actually internalise it, and that lag is often measured in years rather than months.

The interesting tension Google didn't address

Set the document next to the *other* Google story from this week — the new *Updated by AI X minutes ago* label appearing on live search results — and you get a quiet contradiction Google hasn't explained.

On one hand: *AI search rewards the same things SEO rewards. Don't optimise differently for it. Treat AI as a distribution surface for the work you're already doing.*

On the other hand: *We're going to publicly label content as AI-touched in search results, which implies the AI-ness of content is now a user-facing signal we care about enough to surface.*

These aren't direct contradictions. Google can hold *AI content can be fine if it's quality* and *we'll disclose when AI was used to update something* simultaneously. But it tells you that the freshness-and-provenance layer of search is becoming more visible, and that *who or what created this content* is becoming a thing the search result UI cares about, even if the ranking algorithm formally doesn't.

The implication for content programmes is worth thinking about. If labels start appearing widely, then user trust signals around AI-generated content become a click-through-rate factor whether or not they're a ranking factor. CTR feeds back into ranking through behavioural signals. The route by which AI content provenance affects performance becomes indirect rather than direct, but the effect is still there.

This is the kind of second-order thing the help document doesn't address, and the kind of thing that's going to matter more than llms.txt ever did.

The honest limits

A few things this argument doesn't cover, and where reasonable people will push back.

Google's public guidance has historically been incomplete and occasionally misleading. The document says don't over-focus on structured data, but every working SEO knows structured data correlates with rich result eligibility and, increasingly, with AI Overview citation rates. *Over-focus* is doing a lot of work in that sentence. The document is best read as *don't make schema the centre of your strategy*, not *schema doesn't matter*. People reading it as the latter will under-invest in a thing that does matter.

The document is Google-specific. ChatGPT search, Perplexity, Claude, and the agent layer beyond them are not bound by Google's guidance and may reward different things. Mike King's JavaScript-cloaking argument — controversial, technically sound, ethically thorny — is a reminder that the non-Google AI crawlers operate differently enough that uniform advice across all of them is impossible. Google publishing a document tells you what Google wants. It doesn't tell you what OpenAI wants. The two are not the same thing and the gap is widening.

And there's the gap between guidance and behaviour, which has always existed. Google says *don't seek inauthentic mentions*, but the systems clearly still respond to mention volume and entity association. A literal reading of the document would have you stop digital PR entirely, which would be a terrible decision. The document is a guide to what *not* to obsess over, not a complete account of what the algorithm rewards.

What to actually do on Monday morning

Stop paying for tooling that exists to solve problems Google has now said aren't problems. If the tool's primary value is auditing your content for chunking, llms.txt compliance, or AI-readability scoring, the value proposition has just been formally undercut by the platform. Reassess.

Keep doing the work that was always going to matter. Technical hygiene. Core Web Vitals. Genuinely useful content with a point of view. Brand-building through earned media and original research. Internal linking and information architecture that helps both humans and machines understand your site. None of this has changed. None of it will.

Pay closer attention to the measurement gap than to the optimisation gap. The interesting question isn't *how do I rank in AI Overviews*. It's *how do I know whether I already do, for which prompts, with what downstream behaviour*. That problem is unsolved, the tooling is immature, and most of the budget being spent on the wrong half of the equation should be redirected to the right half.

And if you've been hearing pitches from agencies or vendors that lean heavily on the GEO-as-new-discipline framing — chunking strategies, llms.txt audits, brand mention services, AI-specific content rewrites — bring Friday's document to the next meeting. Ask them how their methodology changes in light of it. The answer will tell you whether you're working with people who care about what's true or people who care about what's sellable.

The two are not the same. They've never been the same. Friday's document is just one of the rare moments when the gap is impossible to hide.