For the past year and a half the marketing world has been living in a mild panic. First came AI Overviews, those generated answers at the top of the results that resolve a query before you click a single link. Then AI Mode arrived, turning search into something closer to a conversation. Meanwhile ChatGPT, Gemini, Perplexity and Claude started answering people directly too. And around all of it a new vocabulary sprang up overnight: GEO, AEO, AI SEO, “optimization for neural networks.” Along with the vocabulary came a market of vendors promising to get you into those AI answers.

Against that backdrop, a recent piece on Think with Google is plain by comparison. Brendon Kraham, Google’s VP for Search & Commerce, sums the whole thing up in one line: good SEO is good GEO. Or AEO. Or AI SEO. Or whatever you’re calling it this week.

This article sets out what Google actually said, what it did not say, and where independent specialists disagree. It owns the question of Google’s position. How ChatGPT chooses its sources is a separate mechanism, covered in how ChatGPT picks sources.

The core claim: the new search runs on the old foundation

Google’s logic is almost embarrassingly simple. The generative features (both AI Overviews and AI Mode) don’t live in a separate universe. They run on the same ranking and quality systems as ordinary search, and they pull current content from the same index. There is no secret “index for the neural networks.”

From that follows the conclusion the whole piece seems written for: your past investment in solid, foundational SEO isn’t legacy you need to rebuild, it’s the launch pad. If your site is fast, clear and genuinely useful to a human, it can show up both in the regular results and in an AI answer. Google says nothing more exotic is needed.

Google backs this with one figure from its own reporting: last quarter (per Alphabet’s Q1 2026 release) the number of search queries hit an all-time high. So the “people stopped googling” narrative is overstated. People google more, just differently.

What Google explicitly asks you to stop doing

Here Google names things that whole agencies sell. In Kraham’s telling, it is time to stop:

  • optimizing content for bots instead of people;
  • stuffing text with keywords;
  • chopping material into artificial “snippets” in the hope they’re easier to cite;
  • chasing unnatural brand mentions;
  • building special technical formats like an llms.txt file.

The argument is single-minded: modern systems understand language roughly the way a person does, so tricks aimed at the machine are useless at best. Generative features are perfectly able to find genuine, living content in blogs, videos and forums, without hints in the form of service files.

llms.txt deserves a pause. It is a proposed format, something like a sitemap for language models, that the SEO world has argued about for a year. Google’s position here is plain: don’t spend time on it. We come back to it below.

What Google suggests doing instead

The freed-up time, the company proposes, should go into the one genuinely scarce asset, your own expertise and point of view. In a world of infinite content, what sets you apart isn’t volume but authenticity: first-hand experience, in-house specialists, opinions you won’t find anywhere else.

The example in the article is ordinary on purpose. Picture a local running-shoe store that films a detailed breakdown of why one specific model falls apart over long distances. That video beats the standard “Top 10 running shoes” article, because it answers a real, narrow question from someone who has already felt the problem. Anyone can write the generic roundup (a neural network included). A breakdown of a specific failure, grounded in experience, can’t be faked.

On the practical side, Google reminds you about the Merchant Center and Google Business Profile pairing: filled-in product and company data helps you land in both AI answers and the classic results. And about basic site hygiene, correct display across devices, low latency, a clear content structure. There’s a subtle but important twist here: a user arriving from an AI answer is usually better informed and closer to a decision. So a clumsy interface or a slow load costs you more than before, you’re losing not “cold” traffic but an almost-ready customer.

How to measure success when the old metrics lie

The piece ends with measurement. Google says two things outright. First: the company doesn’t evaluate third-party SEO tools and has no access to their internal metrics, so their “AI visibility scores” deserve a healthy dose of caution. Second: the only dependable measurement strategy is to look at business outcomes. Not abstract visibility, but leads, sales and sign-ups.

There are also new reference points. Search Console now shows the number of impressions from search AI features, and the Merchant Center reports on how products surface in generative answers. It’s not a cure-all, just a baseline: a way to confirm that good content and sound technical SEO actually move the needle.

For specifics, Google names brands that have reworked themselves around the new logic, Lufthansa Group, Etsy and Royal Canin. Big names; worth keeping in mind that global corporations have resources a small business doesn’t, so copying their playbook one-to-one won’t work.

Where Google’s piece stops

Kraham’s piece is the view of a platform that benefits from you doing what you always have: writing good content on your own site. Independent search specialists read the same picture differently.

Search Engine Land’s line of argument is that authority in AI search is created less by what you publish on your own site and more by what is said about you elsewhere. According to a Semrush analysis, the sources AI cites especially often include Reddit, Quora, LinkedIn, YouTube and review platforms like G2. That does not contradict Google’s “be helpful” advice, but it shifts the emphasis: one excellent blog may not be enough if nobody talks about the brand anywhere else. The measured version of that argument, from a capture of ChatGPT’s own traffic, is in how ChatGPT picks sources.

On llms.txt, Google says skip it, while some practitioners treat it as cheap insurance. The evidence is not there for either side yet. Ours is described in the next section.

What we checked ourselves

This site has published an llms.txt since 8 June 2026. The first version was 40 lines. When we checked on 2 October 2026 it had grown to 744 lines listing every service and section page in four languages, and it had changed in 52 commits as pages were added.

We cannot tell you whether it did anything. We have not looked in our logs for AI crawlers requesting the file, and we have no before and after comparison of how often AI answers cite us. We keep it, and we do not count it as evidence in either direction.

What to check on a small site

Three checks follow directly from what Google named, and none needs a vendor:

  1. Business data. If you sell products or serve a local area, confirm your Google Business Profile and Merchant Center entries are complete and agree with your site. Google ties both to appearing in AI answers and in classic results.
  2. Speed and layout. A visitor from an AI answer is usually closer to a decision, so a slow or broken mobile page costs more than it did for cold traffic.
  3. Reporting. Open the Search Console report for impressions from AI features, and keep leads and sales as the judge. Google says it does not evaluate third-party “AI visibility” scores and has no access to their data.

Nobody, Google included, promises a place in AI answers. Google’s own argument is that the work is the same as before.

Sources