Reference

What generative engine optimisation (GEO) is

The work of being mentioned when someone asks an AI what to buy. Here is where the term came from, what the research found actually moves it, and what nobody can promise you.

Last updated 29 September 2026

The short answer

Generative engine optimisation is the work of getting your brand named, and your pages cited, in the answers AI systems write. When a buyer asks ChatGPT, Gemini, Claude or Perplexity which supplier to use, the model composes a recommendation out of what it can find. GEO is the attempt to be in it.

It sits where SEO sat, with one structural difference. SEO competed for a position in a list of ten links, and the list was mostly a function of your own site. An AI answer names two or three options and often no links at all, and it is assembled from what other people have written about you. That makes GEO less a technical exercise on your own pages and more a question of what evidence about you exists for a model to read.

Where the term came from

The phrase was coined in a paper, not by an agency. GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande went up on arXiv on 16 November 2023 and was published at KDD ’24. It is worth knowing because a good deal of writing on the subject credits a venture-capital thesis from May 2025 instead, and the paper is both earlier and more useful: it ran an experiment rather than describing a trend.

The authors built GEO-bench, 10,000 queries split 8,000 for training and 1,000 each for validation and test, then rewrote source content nine different ways and measured which rewrites made a generative engine quote them more.

What the research found works

Of the nine methods tested, the three strongest were editorial rather than technical: adding quotations, adding statistics, and citing sources. From the paper:

  • Citing sources produced “a substantial 115.1% increase in visibility for websites ranked fifth” — the largest effect reported, and it landed on the pages that were losing.
  • Adding statistics achieved “a relative improvement of 30-40% on the Position-Adjusted Word Count”, the paper’s measure of how much of an answer a given source accounts for and how prominently.
  • Adding quotations was the single best-performing method overall, around 40% on the same metric.
  • Keyword stuffing was one of the nine. It was not among the ones that worked.

Read that list again as a description of what a generative engine is doing. It is not ranking pages. It is looking for something quotable, attributable and specific enough to repeat without getting it wrong. The tactics that won are the ones that make a page easier to quote responsibly.

The limits of that result matter as much as the result. The main experiments ran on GPT-3.5-turbo, against the authors’ own benchmark and their own visibility metric, on models from 2023 and 2024. A smaller real-world check on Perplexity showed gains up to 37% on a subjective impression measure. Nobody has replicated it across today’s engines at scale, ourselves included. Treat it as the best evidence available, not as a settled figure.

Why anyone is spending money on this

Because the answer increasingly replaces the click. Pew Research Center tracked the real browsing of 900 US adults through 68,879 Google searches in March 2025, of which 12,593 produced an AI summary. Where a summary appeared, users clicked a traditional result in 8% of visits; where none appeared, they clicked “nearly twice as often (15% of visits)”. Clicks on a link inside the summary happened in “just 1% of all visits”. Sessions ended on 26% of pages carrying a summary against 16% of pages without one.

That is Google AI Overviews and US users, not ChatGPT or Perplexity, and one month of one year. It is still the most direct measurement published of the thing everyone is worried about: the answer satisfies the question, and the source gets nothing. If your brand is not in the answer, there is no second chance further down the page.

What GEO is not

There is a market in files and markup you are told an AI needs. Google published guidance on its own generative features in May 2026 and disagrees, in plain terms: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, and “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.”

So an llms.txt file is not a prerequisite, and neither is AI-specific schema. That guidance covers Google’s features and not ChatGPT, Claude or Perplexity, none of which publish anything equivalent. But when the one engine that has told us anything says the requirement is the ordinary one — a crawlable site and content worth quoting — the burden of proof sits with whoever is selling the file.

What can honestly be promised

Not a citation. The engines decide what they recommend, and no supplier — us included — has a mechanism to make one name you. What can be done is narrower and real:

  • Find out what the answers currently say about your category, which competitors they name, and which sources they were built from. That is measurement, and it is what MediaDesk sells. How we measure publishes the engine list, the matching rules and the score formula.
  • Work out which of those sources you are missing from, and fix the ones you can. Usually a third-party listing that is wrong, or a comparison page in your category that does not mention you.
  • Watch whether it moved, with enough repetition to tell a change from noise. Which is harder than it sounds: why AI answers change when nothing else did.

On the vocabulary: GEO, AEO and AI search optimisation are used to mean roughly the same work by different people. AEO and GEO sets out the one distinction that is actually load-bearing.

Sources

Try it on your own brand

Run a free AI visibility check. It asks four buyer questions in your category and shows you what came back, including the answers that name someone else.