Generative engine optimisation, without the tricks

Generative engine optimisation is the practice of making a business retrievable, trustable, and quotable by systems that compose answers: ChatGPT, Gemini, Claude, Perplexity, and the AI modes inside search engines. Where classical search returns ten links, a generative engine returns one response, and it names a small number of brands and sources inside it.

The term entered the literature through a 2024 study by researchers at Princeton and collaborators, who tested which attributes of a source raise its presence in generated answers. The name stuck because it describes a genuinely new surface: a response that is assembled rather than ranked, drawing on the engine's trained memory and whatever it retrieves live.

That dual mechanism is the key to the whole discipline. Trained memory rewards the long record: years of consistent description and third party mention. Live retrieval rewards the quotable page: a direct answer a system can lift today. GEO that works addresses both; GEO that chases one usually decays.

When a buyer asks an assistant who should reposition their firm before a funding round, the system does something recognisable. It retrieves candidate entities from its trained impression of the world, optionally runs a live search to refresh and verify, then composes a response that names a few brands and cites a few sources.

Each stage has its own gate. Retrieval requires the brand to exist as a clear entity attached to the category and situation in question. Verification requires the live record to agree with the trained impression. Composition requires passages plain enough to quote and safe enough to repeat: hedged, vague, or salesy text gets paraphrased away or dropped.

A business can fail at any gate invisibly. The most common failure is the first one: the brand was never recorded as an answer to the situation being asked about, so retrieval has nothing to find. Writing that records the brand against real buying situations is therefore the core GEO activity, which is also a fair description of a positioning content programme.

The KDD study tested nine interventions on thousands of queries and measured how each changed a source's share of the generated answer. The findings are refreshingly specific. Adding quotations, adding statistics, and adding citations to credible sources each raised inclusion substantially, with gains up to around forty percent on the study's metrics. Keyword stuffing did little. Fluency edits alone did little.

Read plainly: generative engines reward pages that look like evidence and ignore pages that look like advertising. A page carrying named sources, real numbers, and quotable sentences reads as a record worth repeating. A page of adjectives reads as noise, however well it once ranked.

The study also found effects varied by domain, with smaller and lower ranked sites gaining the most from the evidence interventions. That is worth underlining for any small firm: this surface is the first one in decades where the incumbents' link advantage counts for less than the quality of the record.

Four assets are fully in your hands. The entity sentence: name, category, audience, in one line, repeated verbatim everywhere the business appears. The answer library: one page per real buying question, each opening with its conclusion. The markup: organisation, person, article, FAQ, and definition schema connecting every page to the entity. The vocabulary: a glossary giving each term the practice uses its own stable URL.

Two habits multiply all four. Dates that move: engines and their retrieval layers favour records that show maintenance, so update pages and say when. And internal agreement: every page using the same terms for the same ideas, because a site that argues with itself reads as an unreliable witness.

None of this requires new tooling. It requires editorial discipline, which is cheaper and rarer.

Corroboration is the half of GEO you can only earn. Models weight agreement between independent sources far above anything a brand says about itself: reviews, press mentions, directory listings, podcast appearances, and client write ups that repeat the same description of the business.

Forums deserve specific attention. The places categories get discussed, Reddit threads, Quora answers, industry communities, are heavily represented in what assistants read and cite. A founder answering real questions there, under a real name, with the patience to be useful before being findable, is writing the training data of the next model release. The same behaviour done as disguised advertising gets downvoted by humans first and filtered by machines second.

The practical craft is making the echo easy: an entity sentence short enough for a journalist to quote whole, project stories a client can republish, and answers so clear that a stranger citing you gets the description right by accident.

Every new surface breeds tricks, and this one already has a full menu: text written for models and hidden from people, fake statistics inserted to trigger the evidence preference, prompt injection buried in pages, review astroturfing at scale. Each works briefly, on one model, until the next release closes the gap, and detection now carries reputational cost in the record itself.

The decay argument is structural rather than moral. A trick targets the current model's weighting; the weighting is the one thing guaranteed to change. Conditions target memory itself, one clear entity, real answers, independent corroboration, and memory mechanics survive every release because each new model relearns the world from the same record.

This is the old distinction between campaign and brand, replayed on a new surface. Campaigns spike and vanish. The brands that keep getting named are the ones that spent years being one consistent, corroborated thing.

Days one to fifteen: write the entity sentence, install it everywhere, and run the baseline. Twenty buyer prompts across the major assistants, recorded in a spreadsheet: which brands get named, which sources get cited, where you appear and where you are absent.

Days fifteen to sixty: publish the answer library. One page per buying question from the baseline, each opening with a direct answer, carrying question markup, real numbers, and named sources. Fix the schema and the glossary in the same pass.

Days sixty to ninety: start the echo. Two or three client write ups published under real names, directory entries corrected to the entity sentence, and honest participation in the two forums where your category actually gets discussed. Then rerun the baseline and compare. Expect movement in live retrieval assistants first; trained memory follows over quarters, and it follows the record you have now started keeping.