Generative Engine Optimization Agencies

On building a generative engine optimization agency, and on being spoken of when you are not in the room

Late on a weeknight, a woman sits at her kitchen table and asks her phone which accounting firm a small export business should hire.

She does not get ten blue links. She gets a paragraph. Three firms are named, each with a sentence of explanation, the way a well-informed friend might answer over coffee. She chooses the second one. She never scrolls, because there is nothing to scroll.

Across town, a fourth firm spent eight years and a considerable budget earning the top Google position for almost exactly that question. Its partners will never learn that they lost her. No dashboard records a conversation it was never part of. The loss arrives the way many modern losses arrive: as an absence too quiet to notice.

This is the new condition of commerce. Some of the most consequential sentences ever written about your business are now composed by a machine, for a stranger, in a room you will never enter.

Someone is going to build a business around that fact. The interesting question is what kind of business it will be.

I. The finding most people will step over

The name arrived before the industry did. In late 2023, researchers from Princeton and several partner institutions published a paper introducing generative engine optimization: the study of how content becomes visible inside AI-generated answers rather than on a page of search results. They tested a range of tactics across thousands of queries. The old reflex of stuffing a page with keywords did little. What helped, in their experiments, were habits a careful editor would recognize at once: citing sources, including quotations, adding statistics. Visibility rose by as much as forty percent.

Hold onto that result. Most of the people now rushing into this field will step over it on their way through the door.

Around the same time, Gartner forecast that traditional search volume would fall by a quarter by 2026 as people turned to AI assistants. That deadline has now arrived, and the exact figure matters less than the direction, which is hard to argue with. People are asking machines instead of searching indexes. The ones who once ranked are discovering that ranking isn't the same as being mentioned.

II. The librarian who could not read

To see what is changing, remember the old game at its worst.

Early search was a contest against a librarian who could not read. Algorithms counted things such as links, keywords, and how often a phrase appeared, because counting was what machines did well. An entire profession grew up around feeding the counter. There were pages written for robots and unreadable by people, white text hidden on white backgrounds, networks of sites that existed only to link to one another. The best practitioners always did honest work, but the incentive structure was plain: find the gap between what the machine measured and what was actually good, and make a living inside it.

Generative engines narrow that gap. They do not close it entirely, and we will come to the people trying to pry it open. But a system that has absorbed a meaningful fraction of the written internet does not tally your keywords. It forms something closer to an impression. It notices who is mentioned where, by whom, in what context, and whether the story holds together.

The librarian has learned to read. That changes the job of everyone who used to whisper in her ear.

III. The witness

It helps to stop thinking of a language model as a search engine and start thinking of it as a witness.

A witness testifies from what it has seen. Every trial lawyer learns early that you cannot put words in a witness's mouth. More precisely, you can, and the law calls it tampering. What you may do, legitimately and with great effect, is make sure the witness has seen the evidence.

A generative engine is a witness that has read an enormous amount and remembers it unevenly.

Ask it about your firm and it gives a composite account. That account draws on your website, but also on the trade journal that quoted your founder three years ago, the forum thread where an unhappy customer vented, the directory still listing your old address, and the comparison article that happened to name your competitor first. Where the record is thick and consistent, the testimony is confident. Where it is thin, the witness hesitates, leaves you out, or fills the silence with something it made up.

This reframing is not decoration. It defines the entire business. A GEO agency does not sell rankings, because there are no fixed rankings to sell. It sells testimony: the shaping of a public record so that a fair witness, asked a fair question, says true things that happen to be favorable.

IV. What the work actually is

The first service is a deposition. Before changing anything, you question the witness. You ask the machines hundreds of questions a real customer might ask, across several models, and you ask them repeatedly, because the answers are probabilistic and the same question can return different names on different nights. Then you map the results: where the client appears, how it is described, which competitors are named instead, and which facts are wrong.

Clients are often stunned by this document. It may be the first time they have heard themselves described by something with no reason to flatter them. Many agencies will treat the audit as a lead magnet. It is better understood as the moment the client finally sees the problem.

The second service is clarity. Models extract claims, and vague marketing gives them nothing to extract. "Innovative solutions for a changing world" is invisible to a reader that is searching for meaning. "We prepare export documentation for small manufacturers shipping to the EU" is a sentence a machine can carry into an answer. The agency's editors force clients to say specific, checkable things about whom they serve, what they do, what it costs, and how they differ, and to say them consistently everywhere their name appears. Most companies have never done this, because human readers forgave the fog. This reader doesn't.

The third service is original evidence, and this is where the Princeton finding grows teeth. If statistics, sources, and quotations make content more visible, the deepest version of that tactic is to produce knowledge that does not exist anywhere else: a proprietary survey, an annual benchmark, a dataset an industry comes to depend on. Machines cite what is citable, and the most citable thing in the world is something only you know. The serious GEO agency becomes part research shop, part newsroom.

The fourth service is corroboration: press coverage, podcast appearances, reviews, expert commentary, mentions in the lists and discussions where an industry talks to itself. This is public relations, the oldest discipline in the building, wearing new clothes. Reputation was always distributed across many mouths. The difference now is that a single reader listens to all of them at once.

The fifth service is honest measurement. You track how often the client appears across a fixed set of questions, how it is described, and how that changes over months. You report it as a share of answers, never as "position one." You sell probabilities, and you say plainly that you are selling probabilities. In a field that will fill quickly with confident nonsense, candor about uncertainty may be the strongest differentiator available.

V. The temptation

There will be a darker version of this industry, and it is already forming.

Journalists and researchers have shown that hidden text planted on a webpage can sway what some AI search tools say about a product. People are mass-producing articles to seed favorable claims, buying reviews, and coordinating forum posts designed to read as spontaneous. Each tactic is an attempt to coach the witness.

Some of it will work for a while. The last time an industry was built on exploiting a machine's blind spots, the reckoning came in the form of algorithm updates that wiped out sites overnight, and the clients went down alongside the firms that had taken their money. Generative systems will be patched faster and with less mercy, because their makers' credibility depends on it.

There is a more serious cost, though. A model's knowledge is a shared memory, drawn from a commons that everyone uses. An agency that poisons that well is drinking from it too. A client caught manipulating the record does not simply lose visibility. The record itself turns. The scandal becomes part of what the witness has seen.

In the old game, cheating cost you rankings. In this one, it costs you your character, recited by a machine to everyone who asks.

VI. The business beneath the business

So what does it mean to build one of these agencies well?

Staff it oddly. You will need technicians who understand structured data and how models retrieve information, but your most valuable people will be editors, researchers, data analysts, and former journalists. These are people who know what makes a fact worth repeating and a source worth trusting. The tools for monitoring AI answers will multiply and become cheap. Judgment will not.

Choose a territory narrow enough to know intimately. An agency that understands every question a prospective patient might ask about orthodontics in one region, or every question a CFO might ask about mid-market payroll software, will beat a generalist that knows a little about everything. The question landscape is the map, and maps are drawn one territory at a time.

Price the work as what it is, a continuing relationship with a moving record. Monitoring and editorial work suit a retainer. Original research suits a project fee. Never guarantee placement in an answer you do not control.

Above all, understand the economics that will reshape everything around you. When machines can produce fluent prose for almost nothing, words stop being scarce. What becomes scarce is what words are supposed to carry: verified facts, original findings, credible people willing to put their names to a claim. When words become free, the price of a fact goes up. The agency that builds its business on that scarcity is building on rock. The one selling volume is building on a tide that is already coming in.

Look closely and the paradox at the center of this industry comes into view. The most effective way to optimize for a machine that has read everything is to become more genuinely useful to the humans it read. The technique runs back into substance. The shortcut leads, eventually, to doing the work.

Which means the best GEO agencies will spend a surprising amount of their time telling clients things they do not want to hear. Your positioning is meaningless. Your claims cannot be verified. Nobody outside your building has ever said anything specific about you. The machine is not ignoring you out of malice. It simply has nothing to repeat.

Go back to the kitchen table, and to the firm she chose.

Perhaps its partners have never heard the phrase generative engine optimization. Perhaps they simply spent a decade answering client questions in public, publishing a yearly survey of export costs that trade reporters kept citing, keeping their address and services accurate, and doing work specific enough that customers described them in specific sentences. When a stranger asked a question late at night, the machine repeated what the world had been saying all along.

She will never know why the second name sounded right.

It sounded right because, for years, it had been true.