What the Agents Read

On May 15, 2026, Google closed the debate on GEO. What remains, for territories that have something to say, is a narrow window in which situated substance replaces the keyword.

On May 15, 2026, in the Google Search Central documentation, under a newly created tab that would later be named Generative AI fundamentals, a document of a few thousand words appeared, titled Optimizing your website for generative AI features on Google Search. Neither a mood piece nor a product announcement: a doc page, filed next to the SEO Starter Guide, that cleanly closes two years of uncertainty. Four days later, in Mountain View, Google I/O on May 19 and 20 rolled out AI Mode at scale, extended Personal Intelligence to nearly two hundred countries, and wrote information agents into the background of search. The following day, Google announced a new Core Update, which, as I write, is beginning its rollout. This sequence is not incidental. It shows where traffic will go over the next eighteen months, and therefore what it's still good for to write on the web at all.

What this sequence changes for Swiss businesses, and in particular for Valais actors who have, in the shift now underway, a short window to seize (as the book argues), comes down to a handful of observations and a handful of instructions.

What the document says

The guide rests on three propositions, stated plainly. The first is technical. Google Search's AI features (AI Overviews at the top of results pages, AI Mode as a conversational interface, agents that crawl the web in the background to prepare concrete actions) do not run on a separate index. They draw on the standard Search index through retrieval-augmented generation and query fan-out: a language model breaks the query into sub-questions, retrieves the relevant passages from pages already ranked by the usual systems, and composes the answer from those extracts. If your content isn't indexable, crawlable, and judged high-quality by yesterday's systems, it won't be by today's either.

From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

The second proposition is a marketing one, and it takes direct aim at agencies and their market. Everything sold since late 2024 under the name GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization), as a framework distinct from SEO, rests, according to Google, on nothing technically specific. The same sentence, repeated from different angles throughout the document, closes the door on parallel frameworks. There is no new discipline; there is an old discipline whose demands have just been raised a notch.

The third proposition is blunter still. A section titled Mythbusting generative AI search names the techniques to abandon. llms.txt files (an open standard pushed by part of the technical community to signal to models what they may ingest) are treated by Google like any other text file, with no indexing privilege. Artificially breaking content into chunks meant to ease ingestion by LLMs achieves nothing: Google's systems extract on their own what they retain from a multi-topic page. AI-specific Schema.org tags, rewrites designed to please models rather than readers, paid mentions on third-party sites meant to "train" the AIs: all of it is expressly filed under the useless. The guide doesn't ask anyone to give these up on principle. It simply notes that they don't serve the goal being pursued.

In their place, the document puts four requirements back at the center. Content that is valuable, unique, non-commodity — first among its keywords. First-hand expertise: lived, situated, signed by an identifiable human being. A technically sound site, accessible, fast, readable by humans as much as by machines. And the E-E-A-T framework, expanded in 2022 (Experience, Expertise, Authoritativeness, Trustworthiness), whose first letter, Experience, becomes the real differentiator now that generative models frame how reading happens.

What is closing

To understand what this announcement closes, it helps to look at what had opened in the first place.

Over the past twenty years, search engine optimization gradually industrialized. A significant share of the content published on the web existed for the sole purpose of capturing a position on a keyword. Aggregator sites, content farms, pages mass-produced around variations of a query, FAQs stuffed with questions no one ever asked, thousand-word articles whose first eight hundred exist only to hit some presumed minimum length before finally delivering the answer: all of this constituted an industry. That industry employed tens of thousands of people, fed agencies, sustained a whole ecosystem of tools, and trained an entire generation of writers to produce text that doesn't address human beings.

The arrival of accessible language models in 2023 supercharged that industry. Producing a thousand SEO pages cost next to nothing. The result was predictable: the web filled up with plausible, empty text, production costs fell, competition rose, and rankings were fought over in a race that the most disciplined players no longer won.

What agencies sold in 2025 under the name GEO was, for the most part, the same race shifted up one level: producing content not to climb the blue links, but to appear inside the answers synthesized by LLMs. Google has just written that, from Google Search's point of view, this race has no separate track.

SEO as a discipline therefore survives the announcement. What closes is the idea that one could, by industrializing content production, durably capture a share of organic attention without holding any real substance.

What is opening

What opens in return is more demanding, and fairer.

The retrieval systems feeding AI Overviews must, by construction, choose which passages to cite. For every query, they select a handful of sources out of millions. They no longer bother with linear rankings: there is no page one, page two, page three anymore. There is the answer, and the two or three sources that fed it. To exist in this new geography of attention, a site must produce what SEO literature now calls information gain: content that adds something other sources don't have. An original analysis, an unpublished figure, a situated observation, a lived experience, a framework for interpretation that exists nowhere else on the web.

This standard remains out of reach for industrial production. It is, however, well within reach of practitioners who speak from inside their subject.

This is where the connection to this essay's thesis becomes direct. The competitive shift I describe in chapter five (the partial erosion of skilled work's urban concentration, driven by AI tooling) has an editorial counterpart. For twenty years, industrial SEO had neutralized part of the advantage that expert practitioners held over content producers with no substance of their own. A large agency could publish a thousand pages on Valais tax law without ever having advised a single local client, and still capture a substantial share of the corresponding queries. That contest played out at the scale of search engines: on rankings, on internal linking, on accumulated domain authority. The four-person trust firm, custodian of real expertise but with no SEO budget, didn't figure in it.

What the sequence from May 15 to 21, 2026 sets in motion is, as I read it, the partial reversal of that mechanism. Generative models synthesizing answers need sources where the substance runs dense. At equal technical quality, they favor content that contributes something irreducible: a lived case, a piece of local data, an original line of reasoning, an identifiable voice. The trust firm that has spent three years publishing signed notes on the taxation of second homes in tourist zones now finds itself, in 2026, in a position its resources never would have afforded it. It can be cited — not for its size, but for its substance.

One should still resist too comfortable a reading. The example of the trust firm publishing for three years assumes precisely what most SMEs lack: a writing discipline established before the window opened. Generative models don't distinguish, in a single day, the situated practitioner from the circumstantial newcomer. They pick up, through authority signals distributed over time, those who have been writing about their subject for a while. The window this sequence opens therefore presupposes a running start, not a mechanical switch. SMEs that begin in 2026, signed and patient, will see the benefit in eighteen to twenty-four months. Those that wait for others to prove the return will arrive, as so often, once the positions have already set.

What businesses need to do — and stop doing

The first front is content production. Anything outsourced to an agency whose deliverable is called a "1,500-word SEO article" needs rethinking. Not that these agencies are dishonest — many do clean work — but the commodity they produce is exactly what Google now signals it no longer wants to promote. The opposite direction is clear: have practitioners write, in their own voice, about what they actually know, under their own name. This means reorganizing the internal production chain. It also means accepting that one publishes less, and longer: a reallocation of budget, more than a comfort.

The second front is editorial discipline. Content that AI systems will favor in 2026 answers to a handful of testable criteria. It is visibly dated, because a recent answer outranks an old one; signed by an identifiable person, because models assess the producer's Experience; sourced, because they distinguish assertion from rumor. It takes a position when doing so is honest, because the information gain of a text that never commits to anything is zero, and it owns its own voice, because a text smoothed over by a rewriting tool ends up losing the very mark that models read as a signal of authenticity. It avoids, finally, the patterns agencies have overused these past eighteen months: artificial FAQs, bullet lists dropped into contexts that don't call for them, marketing chunking, opening lines that announce what's about to be said instead of just saying it.

The third front is measurement. Classic SEO tools (rankings, search volumes, backlinks) remain useful but no longer suffice. It's time to start instrumenting visibility inside AI answers: tracking citations in Google's AI Overviews, in the answers given by ChatGPT, Claude, Perplexity, Mistral. Several platforms now offer this kind of tracking. None of them is perfect yet. That's no excuse not to look. The metric that matters is no longer just the click, which erodes as AI answers display directly on the page, but the citation, and the authority that citation builds through repetition.

Separately, there remains work of another order: editorial sovereignty. The sequence of May 15 carries a subtext. Google explains that llms.txt files are of no use to it, which is technically accurate for Google Search. But this standard functions less as a ranking lever than as a documentary act of declaration: a way for a site to state what it considers readable, stable, citable. Whether the major models honor it tomorrow or not, the act of writing it and publishing it openly belongs to whoever edits the site. The argument is political well before it is technical. And this site, which publishes one, knows it.

What this means for Valais

I have written elsewhere that AI's territorial window would open to the benefit of actors who know how to orchestrate the shift rather than suffer it. The sequence of May 15, 2026 is one of its clearest manifestations, applied to the visibility of businesses on the web.

For Valais SMEs (trust firms, lawyers, doctors, wine cellars, craftspeople, tourism operators, business consultants), a path reopens. Not by imitating the SEO agencies of Zurich or Geneva. By doing the opposite. By publishing what they actually know, in a voice that belongs to them, at a sustainable pace. One note a month, signed, dated, situated, on a subject they genuinely master, is now worth more than ten generic articles mass-produced. Far from a consolation prize for small players, it is their comparative advantage reasserting itself.

For the canton itself, as an institution, the stakes run parallel. Public content — from administrations, universities, the bourgeoisies, professional associations — is exactly what AI models cite readily, because it carries situated expertise and a domain authority few commercial actors can match. There is an underexploited asset here, one that doesn't call for a conventional communications budget but for rigorous editorial discipline and an explicit policy of openness toward the AI crawlers of the publishers one agrees to be read by.

In the background remains the political question this site tries to keep open. When AI agents fetch what we write, they hand it to their users without those users ever visiting our sites. The business model of "free content funded by click-through advertising" is going dark, page by page, as answers form outside our pages. What territories and businesses now draw from their content is less monetizable traffic than distributed authority — an authority that has value but has not yet found its price.

That is precisely this site's wager, and the reason its robots.txt names the agents it allows to come and read it. Being cited, in 2026, counts for less as an advertising win than as an entry in the conversation these models hold in our place. At this level of stakes, gaming the tags leads nowhere anymore. What matters is having, on one's subject, something to say that no one else has written.