The GEO
Doctrine.

There is a quiet, almost clerical moment, sometime in late 2023, when the arithmetic of attention broke. For twenty-two years, the question had been the same. Someone typed a question. A list of ten blue links came back. You fought, at great expense, to be one of those ten links. The house that owned the top link captured most of the traffic, and the house that sat on page two captured nothing at all. We called that contest search, and we built an entire industry around winning it — agencies, tools, conferences, job titles, careers — and we did so in the perfect confidence that the contest would continue in recognisable form, forever.

It did not continue. The interface changed. Where there was once a list of links, there is now a sentence — a short paragraph, authored by a model, that attempts to answer the question directly. The links are still there, but they are demoted, quieter, below the fold. The sentence above them is read first and, more often than not, read only. Some months later, the same thing happened inside ChatGPT, Claude, Perplexity, Gemini — each with its own variation on the same theme. An answer, in prose, with a short bibliography of sources. The sources, significantly, are not the ten blue links.

The technical literature calls this generative engine optimisation, or GEO, and the acronym has the dull taste of every other acronym we inherited from the SEO years — KPI, CTR, SERP — signals of a profession that prefers jargon to plain speech. I dislike the word. But the phenomenon it names is real and, more to the point, it is already the dominant surface of discovery for the only audiences that actually matter: founders, principals, operators, senior buyers, curious readers who refuse to read a listicle. These people are no longer in the SERP. They are in the answer. [1]

The first principle.

Let us state the doctrine plainly, before anything else. The proposition of every modern search interface has changed, and the change is not cosmetic. It is ontological. The interface no longer retrieves; it answers. The old contest asked: whose page will rank in the list? The new contest asks: whose writing will the model quote, by name, in the sentence?

The prize is no longer attention — it is attribution. Not the click, but the quote.— The doctrine, in one line

This is not a small distinction. It changes, from the ground up, what a house must do to be found, what it must publish, what it must archive, who it must be known to, and — perhaps most painfully — who it must no longer be confused with. An entire professional reflex, trained on the old contest, becomes not merely useless but actively harmful in the new one. Keyword clouds, link farms, publishing cadence for its own sake, “AI-assisted” content at scale, the quiet ghost-writing industries that powered two decades of thought-leadership: all of them are now liabilities. A model cannot be fooled by the techniques that fooled a crawler. It reads, in a sense we are still learning to describe, the way a reader reads — for coherence, for specificity, for the writerly tell of someone who has actually done the thing. [2]

There is a second, sharper consequence. The old contest rewarded presence; the new one rewards repetition. A house wins SEO by being one of many ranked results. A house wins GEO by being the same name, cited across hundreds of conversations, on thousands of days, by millions of readers who never clicked anything. The model's answer calcifies into a default. The default names whom it names, and does not name whom it does not, and the silence — not the absence from a list, but the silence in the sentence itself — becomes the structural form of failure in the next decade. We will meet houses who are loud on the internet and silent in the answer. We will meet more of them every quarter.

What a model actually does.

To understand why the contest has changed, one must understand, at the level of the plain-language description, what a language model is doing when it answers a question. Most of the writing on this is too technical for the principal, or too mystical for the engineer, and neither will do. The plain description is this: a model has read a great deal, forgotten most of it as precise text, and retained it as relationships. When asked a question, it reconstructs an answer by retrieving the relationships most closely associated with the terms in the question — not the pages those relationships came from, but the claims those pages made.

What this means, for the purpose of the doctrine, is three things, and only three:

1. The model cites whom it can remember. Not whom ranks well, not whom pays best, not whom writes most — whom it can, reliably, recall by name in relation to the question asked. A house that has never been named in a model's training material, or its retrieval index, is invisible in the sentence. Whatever else is true of that house, it does not exist in the answer layer.
2. The model cites the specific over the general. Given two sources, one saying “our leather is the finest” and one saying “we use single-hide vegetable-tanned Tuscan cow leather with a saddle-stitched hand seam, taking twelve hours per bag”, the model cites the second — always. Generality is safety for brand managers and poison for citation. The bland sentence does not make it into the sentence.
3. The model cites consistent identity over consistent volume. A house that writes thirty pages a month in four inconsistent voices is a less citable entity than a house that writes three pages a year in one signed, identifiable voice. The answer layer rewards unmistakable authorship the way the magazine layer once rewarded a byline. The name must be the same name, on the same subject, over years, for the model to build the relationship at all. [3]

Each of these three principles is the mirror image of a twenty-year SEO habit. SEO rewarded presence over identity, volume over specificity, and keyword coverage over relational depth. The new contest inverts all three. It is not an incremental change. It is a reversal.

The three false doctrines.

There are three prevailing false doctrines about the new contest, propagated by firms that understand the shift is happening but have not yet understood what it asks of them. Each is attractive. Each is wrong. I describe them here plainly so the reader can dismiss them when they are offered — and they will be offered, at length, by people with slide decks.

The first false doctrine is that GEO is SEO with new keywords. It is not. The unit of the new contest is not the keyword but the claim. A keyword is a string; a claim is a proposition that can be true or false, specific or general, interesting or dull. Models cite propositions, not strings. A house that treats “generative engine optimisation” as a keyword to be stuffed into pages is fighting the last war with worse equipment.

The second false doctrine is that GEO is a volume problem — that if the old game needed forty pages a month, the new game needs four hundred. It does not. The new contest is, if anything, an anti-volume contest. Models are trained, retrained, and retrieved-from on corpora that actively discount the signature patterns of bulk-written content. The more pages a house publishes in the machine voice, the more it trains the models to disregard it. The economics run in the opposite direction to the old game. [4]

The third false doctrine is that GEO is a technology problem — that the answer is a new tool, a new tracker, a new dashboard of “AI visibility metrics.” It is not. It is, as every prior contest of this kind has been, a writing problem. Tools will tell you the score. They will not score for you. A house that buys the dashboard and does not change what it writes will be told, in high resolution, the precise degree to which it is losing.

“GEO is not SEO with new keywords. It is not a volume problem. It is not a tool problem. It is a writing problem, and the houses that understand this first will be cited for a decade.”

What the house must do.

If the first three sections describe the shape of the contest, the fourth must, in fairness to the reader, describe what winning it looks like. I will not rehearse our full method here; that is what the method room and the diagnostic are for. But the outline of the programme — the doctrine in its practical form — is as follows.

A house wins the answer layer by doing, deliberately and over years, six things. Each is simple to describe. None is easy to execute. The difficulty is the whole point. It is the difficulty that creates the moat.

1. Author in one identifiable voice. One principal, one byline, one signed body of writing on one subject, over years. The model learns the voice before it learns the subject, and the voice is what makes the citation possible.
2. Write claims, not copy. Every page the house publishes must contain at least one specific, falsifiable, interesting claim — something the reader could argue with, quote, or check. Copy contains no claims. Copy is unciteable.
3. Make the primary material available. The archive — essays, field notes, case records, correspondence — is the substrate from which citations are drawn. A house without an archive is a house without a past, and the model will not cite a house without a past.
4. Be known to those who name others. The citation network is not the house's own website. It is the network of publications, platforms, editors, podcasters, and peers who name the house, in their own writing, on the record. This is the authority layer, and it cannot be faked.
5. Keep the entity clean. Wikidata, LinkedIn, official registries, the canonical description of the house across the semantic web — the model's first retrieval on any question is an entity retrieval, and the entity must be unambiguous, consistent, and complete.
6. Measure what the model says. Not traffic. Not rankings. The words the model returns, when asked the questions that define the category. The citation ledger is the only live scoreboard; keep it or keep guessing.

Each of these, taken alone, is achievable. The difficulty is that they must be done together, in sequence, across years, by a house that can keep its voice through three business cycles and two founder moods. Most cannot. That is exactly why those who do are cited for a decade.

The closing argument.

It is tempting, for a principal reading this, to set the essay down and conclude that the matter can wait. Search still works. The website still converts. The agency still reports green on most of the dashboards. The temptation is understandable and it is mistaken. Every contest of this kind, in every prior epoch — the yellow pages, the magazine, the directory, the search engine — has had this same quality: a long, deceptively quiet period in which the new surface coexisted with the old, followed by a short, loud period in which the old collapsed. We are in the quiet period now. The loud period will be obvious in retrospect, and by then it will be too late to catch up.

The houses that win the next decade will not be the loudest. They will not, for the most part, be the largest. They will be the ones whose names are in the sentence — cited, by the model, when a serious person asks the serious question that defines the category. The game is not difficult to describe. It is only difficult to do. That is, and has always been, the entire structure of a moat.

Which returns us to the doctrine, with which we began: the prize is no longer attention — it is attribution. Or, to put it more plainly still: the next decade will belong to the houses who are quoted, by name, by the only thing that still reads. The rest will continue to appear in lists nobody looks at.