What Google Actually Says About GEO in 2026

In May 2026 Google published the document the GEO industry had spent two years insisting did not exist. One sentence in it does most of the work: from Google Search’s perspective, optimising for generative AI search is optimising for the search experience, and thus still SEO.

That should have ended a lot of commerce. It did not. Search the term today and the first page still sells a separate discipline — a comparison table, a tactics list, a visibility tracker with a subscription attached. The tactics are not all wrong. They are being sold as a practice with its own leverage, and the primary documentation does not support that.

Our position: there is no shortcut called GEO. The advantage that lasts is publishing information good enough that a machine has a reason to reach for it.

What the documentation actually commits to

Google’s guide is more specific than the coverage of it suggested, and worth reading rather than reading about.

It says the AI answers are built by searching first and writing second — the system runs real searches, quietly asks itself several related questions rather than only the one you typed, and composes an answer from what comes back. Crucially, those searches run on the ordinary Search index and the ordinary ranking systems. That is the load-bearing fact. Being eligible to appear in an AI answer is the same problem as being eligible to appear in Search: can the page be crawled, is it indexed, is it worth surfacing.

On content the guidance is unusually direct. Unique perspective. Grounded in real expertise and experience. Explicitly not commodity. It also warns that spinning up many variations of the same material to widen coverage falls foul of its rule against mass-produced pages — the same old warning, pointed at a newer excuse.

Nothing there would surprise a competent SEO practitioner. That is rather the point.

The mythbusting list is a map of what is being sold

The most useful section of Google’s guide is the one listing what you do not need to do.

You do not need to publish llms.txt or any of the other text files vendors recommend for AI crawlers. You do not need to chop pages into bite-sized chunks; Google’s systems already pull the relevant part out of a page covering several topics. You do not need to rewrite content specially for machines. Chasing brand mentions you did not earn does not work. And there is no special structured data for generative features — no markup you can add that buys you a place in an AI answer.

Read those five together and something becomes obvious. It is close to an inventory of the paid GEO deliverable. If a proposal you have been sent contains three of them, you are not buying a strategy. You are buying artefacts that are cheap to produce and impossible to disprove.

One nuance worth keeping. Google says these things do not help its generative features. It does not say they are harmful. Structured data still earns rich results and still tells a search engine which number on the page is a price. Judgment, not doctrine — we still specify it on builds where it earns its place, because there it does work we can name.

Some GEO tactics really are distinct from SEO

The strongest objection to our position is not that GEO is magic. It is that Google is not the only system answering questions, and Google’s documentation only binds Google.

That objection is fair. ChatGPT, Perplexity, Claude and Copilot each have their own crawlers, their own indexes, their own licensing deals and their own way of weighting sources. Some read files Google ignores. The academic paper that gave the field its name reported visibility gains of up to 40 per cent from source-side tactics — adding citations, quotations and statistics — measured on the authors’ own benchmark, against the assistants of that period, with their own finding that the effect varied sharply by subject. That is a real result. It is not a promise about what a mid-market brand sees in 2026, and nobody quoting the number in a pitch deck mentions the caveat attached to it.

So the defensible position is narrower than either camp wants. These tactics can be genuinely different for the non-Google assistants. What they cannot yet be is attributed. The field has not agreed what to call itself, let alone how to measure itself, and not one assistant publishes anything like Search Console. Distinct is not the same as evidenced.

Measurement is where GEO programmes quietly fail

This is what decides whether a programme survives its first budget review.

Google now reports generative AI performance in Search Console — how often your pages appeared inside AI Overviews and AI Mode, broken out by page, country, device and date. It reached every site worldwide on 31 August 2026. It is the only first-party measurement any of us have, and it has a hole in the middle. It reports appearances. It does not report clicks, and it does not report the questions people asked.

A feed of appearances tells you a page showed up. It cannot tell you what that was worth. So any programme whose success metric is an “AI visibility score” from a third-party tool is reporting a proxy for a proxy — and Google’s own guidance warns against tools claiming access to numbers it does not publish. We report on pipeline, not impressions, and that constraint should decide what gets funded rather than being discovered six months in.

Three questions before you fund any GEO activity

Start with the plumbing, not the prose. Before anyone argues about content, check that a machine can reach it at all: does the text exist in the HTML a crawler receives, or does it only appear after the page runs a script — a web engineering question, not an editorial one. Then the ordinary hygiene of crawlability and speed. Then instrumentation: is the Search Console report actually connected and being read, rather than screenshotted into a deck. Teams that instrument before they optimise argue about evidence. Teams that do not argue about opinion.

Only then does the editorial question arrive, and it is the one Google’s guide is really asking: what do you know that nobody else on the results page can write. That question is worth more than any file you can put in a web root, and it gets its own argument in what makes a source worth citing.

After that, put any proposed AI-search activity — yours, an agency’s, ours — through three questions. It has to pass all three.

Is it useful without the AI story? Strip out every mention of AI Overviews and assistants. Does the work still improve the page for a person who lands on it? If the only justification is machine visibility, you are producing artefacts, not information.

Can you measure it with your own data? Name the instrument before the work starts: Search Console, analytics, CRM, revenue. A third-party visibility score is not your own data. If nobody can say what would prove the activity failed, it will never be found to have failed, and it will never be stopped.

Is it grounded in something people actually do? Point at real demand — queries people run, questions sales keeps getting asked, support tickets that repeat. Not a hypothesis about what a model prefers to quote.

The activities that pass tend to look like maintenance: fixing rendering, retiring thin pages, writing the one document only your practitioners could write. The activities that fail tend to be the ones with a product attached. If the answer for your organisation is that the fundamentals are not in place yet, that is the conversation to have first.