For twenty years we asked about rankings — what GEO changed.
For twenty years the question was always the same: “Where do we rank on Google?” The question is changing. Your customer now asks a model, and the model gives a single answer.
GEO — Generative Engine Optimization — builds your site into that answer. It does not replace classic SEO; it adds a layer on top: content clear enough to be quoted, identity consistent on every page, structure that machines can read.
Ranking and being cited are not the same thing
A search engine returns ten blue links and leaves the choice to the user; an answer engine makes the choice itself, produces one piece of text and lists a few sources beneath it. That difference changes the measure. Ranking second still brings some traffic. In citation there is no second place: either you are one of those three sources or you are absent.
The second difference is subtler: a model does not read a page, it reads a PASSAGE. When it builds an answer it takes the paragraph that answers the question directly, not the whole document. So the unit of GEO is the paragraph, not the page. "This page covers the topic" is not enough; "this paragraph answers the question on its own" is.
What a quotable passage looks like
It has three properties. It stands alone: it does not borrow context from earlier paragraphs and never opens with "as we explained above". It starts directly: the first sentence answers the question with no warm-up lap. It is verifiable: where there is a number, there is a note on how it was measured. All three are also the rules of good writing — GEO invents nothing here, it just makes mandatory what used to be optional.
The counter-example teaches more: "With over 20 years of experience in the sector, placing customer satisfaction first…" answers no question at all. A model cannot quote it because there is nothing in it to quote. That kind of text never worked in search either; in answer engines it simply disappears.
Whether models can read your site is measurable
AI crawlers largely receive the page as raw HTML. Most never see content painted later by JavaScript — they do not run a browser engine. You do not have to guess: request your own site with those user-agents and read the body that comes back.
- GPTBot · OAI-SearchBot · ChatGPT-User → 200
- ClaudeBot · PerplexityBot → 200
- Google-Extended · Applebot-Extended → 200
- # nine UAs, all nine got full HTML · 2026-08-21
On this site all nine crawlers receive the complete body, because the content arrives ready from the server. The measurement takes a minute and closes the question of "can they even see it" for good.
Three steps, in order
One: build the schema layer. JSON-LD tells a model who this is, what it does and where, without making it guess from sentences. Two: audit entity consistency — the brand name, the founder’s name, the service name must be written identically on every page; three spellings mean three different entities to a model. Three: publish llms.txt, leaving a ready summary for the model to work from.
This site applies all three itself — check the source for proof. The schema graph carries eleven node types, the identity resolves to the same @id on every page, and llms.txt sits at the root listing real addresses.
Update, 1 October 2026: we now write the third step more cautiously. Google’s AI guidance states plainly that Google Search does not use llms.txt, and the piece where we gathered the 2026 data shows no citation gain attributable to the file. The file stays on this site, but as an optional map, not a lever.
What GEO does not promise
No technical change guarantees that a model will mention you; if someone guarantees it, walk away. GEO builds the PRECONDITIONS of visibility: content becomes readable, identity consistent, passages quotable. The rest depends on authority, topic and time. That is the measurable part, and for the unmeasurable part we write no numbers.
A site has two readers: a human and a model. Tell both the same truth.