GEO, or how to become the source the models cite.

What changes when an engine answers instead of listing

On a classic engine, the results page is a list: ten links, an order, and the decision left to the person. Position matters, because it determines the probability of a click. On a generative engine there is no list. The model writes an answer, then cites two, three, sometimes five sources that fed it. Everything else disappears.

The consequence is blunt and simple: between the second and the twentieth relevant source there is no longer any difference. Either you are in the answer, or you do not exist for that question. That is what makes the exercise different from classic search, where a tenth position still returns something.

Generative Engine Optimisation — GEO — is the set of practices that raise the probability of being retained as a source. The term is recent; the mechanics are less so. They borrow from search, from document structuring and from accessibility.

How a model chooses its sources

When a generative engine answers a question that calls for recent facts, it does not draw on its training memory. It runs one or more searches, retrieves pages, extracts passages, and writes from those passages. Four filters follow one another, and a page can fail at any of them.

Access

The first filter is the dumbest and the most common: the crawler has to be able to read the page. GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended are separate crawlers from Googlebot, with their own permissions. Many sites block them without knowing — a security plugin, a host setting, a line inherited from an old robots.txt. Excellent content behind a block is worth zero.

Identity

The second filter is about what you are. A model more readily cites a source it can name, place and attach to a field of competence. That is what structured data is for: declaring in JSON-LD that the site belongs to an organisation, that it works in a given trade, that it operates in a given area. Without that declaration the model has to guess — and in doubt, it cites somebody else.

Extractability

The third filter is the one most sites fail. A model takes passages, not pages. A paragraph that answers a precise question in three or four sentences, under a subheading that poses that question, is directly usable. A two-thousand-word argument circling the subject without ever stating it is not. Heading hierarchy matters for the same reason: it tells the model where an idea begins and where it ends.

Freshness and corroboration

The fourth filter separates the sources that are left. All else equal, a page that is dated, updated, and whose claims are found elsewhere — press, directories, business listings, public documents — goes ahead of an anonymous, undated page. That is also why a GEO strategy is not played out on the site alone: it is played out across every place your name appears.

What GEO does not do

This has to be said, because the market usually says the opposite. Nobody guarantees a citation: models vary from one session to the next, change version, change their sources. What can be steered is the probability, not the result.

Nor does GEO replace search optimisation. A generative engine leans very heavily on what a classic index has already found. A site invisible in Google is invisible in ChatGPT, for the simple reason that there is nothing to retrieve. The two disciplines are worked together, with levers that partly overlap.

And finally, no markup rescues empty content. Structured data describes what exists; it does not replace it. A site with no real expertise to expose will stay uncited, however good its JSON-LD.

Where to start, in this order

1. Check that the generative crawlers have access. Read the robots.txt and lift the inherited blocks. It is free, immediate, and it is the one point that can cancel everything else.
2. Declare the entity. A JSON-LD Organization or LocalBusiness block on every page: name, address, contact, fields of competence.
3. Rewrite the subheadings as questions. Take the questions customers actually ask and turn them into headings, with the answer in three sentences directly underneath.
4. Repair the heading hierarchy. One H1, no skipped level. This serves the models and the screen readers alike — both read the same structure.
5. Date it and sign it. Show the publication and update dates, mark them up, name an author.
6. Publish an llms.txt. A text file at the root summarising the business and pointing at the reference pages. The convention is young and its effect is still to be measured, but it costs an hour.

Measuring, without telling yourself stories

Two things can be measured seriously. The first is the technical state of the site: the points above are verifiable, one by one, against the code actually served. That is what our free audit does, in a minute and with no sign-up.

The second is the citation itself, and it takes a method. You build a list of questions as a customer would ask them, never naming the brand. You put them to several engines, several days apart, and you note who gets cited. A single reading proves nothing; a series over three months shows a trend. It is slow, it is manual, and it is the only honest measurement available today.

What cannot be measured: a position, a ranking, a universal "GEO score". Those objects do not exist. A score — ours included — measures how far a page meets known criteria, not a place in a league table.

The four subjects that deserve their own page

• SEO and GEO: what overlaps, what differs — the shared levers, the ones specific to generative engines, and how to split the effort depending on where you stand.
• Structured data: which schemas to declare — the five types that count, the order to lay them down in, and the three mistakes that cancel all the markup.
• Being cited by ChatGPT — the three situations where the system goes out to the web, what makes a passage get picked, and the test you can run yourself in fifteen minutes.
• llms.txt: should you publish one — what is known about its real effect, how to write it, and why the answer is still yes.

The article Is your company cited by ChatGPT? goes back over the source-selection criteria and the check you can run yourself. If you would rather start with findings, the audit is free and immediate; if it confirms the need, the full audit is set out here.