What is GEO (Generative Engine Optimization)?

The definition

Generative engine optimization (GEO) is optimising a website, content assets and external sources so that generative AI systems name, quote or recommend a company in their answers to user questions.

The term was introduced in November 2023 by researchers from Princeton University, IIT Delhi and Georgia Tech, and published at the KDD conference in 2024. They defined a generative engine as a system that combines information from several sources into a single answer, with or without citing sources.

You will find the same movement under other names. AEO (answer engine optimization), AIO (AI optimization) and LLM SEO. In practice they describe the same work from a slightly different angle. GEO is the term with an academic origin.

Why it matters now

Two numbers tell the whole story together, and they do not point the same way.

• 68.01% of Google searches ended without a click in the first four months of 2026. Ten years ago that was about 45%. The answer is more and more often on the results page itself, or in a chat window (SparkToro, June 2026, based on Similarweb data).
• AI assistants still send hardly any click traffic today. Measured across 107,000 websites, Google accounted for 26.4% of referral traffic and all AI assistants together for 0.39% (Ahrefs, June 2026).

If you only quote the first number, you are selling panic. If you only quote the second, you miss the point. The right reading is this. The traffic AI sends is small, but the decision is formed there. The B2B Buyer Experience Report by 6sense (November 2025) shows that 94% of B2B buyers use language models, among other things to summarise reviews and analyse data, while they still have 16 interactions with the winning supplier on average.

In other words, the shortlist forms in the conversation with AI, and the purchase happens with you afterwards. If you are not in that conversation, you are not on the shortlist.

How a generative engine chooses a source

A language model draws on two things. What it learned in training, and what it fetches from a search layer at the moment of the question. That second route is where GEO steps in.

• Fetching.The engine turns the question into one or more searches and fetches a set of pages. If a crawler cannot read your site, it ends here.
• Selecting.From that set, the model picks the passages that carry the answer. Short, factual, well-defined paragraphs beat long marketing copy.
• Combining.The model writes an answer and sometimes names sources with it. Companies that appear again and again in several independent sources get named sooner than companies that only appear on their own site.

That is also the biggest misunderstanding. GEO is not a trick on your homepage. A model only starts naming you once others name you.

What demonstrably works

The KDD study tested nine edits on thousands of questions. In the authors' words, the best-performing methods, Cite Sources, Quotation Addition and Statistics Addition, delivered a relative improvement of 30 to 40% on the visibility measure and 15 to 30% on the subjective assessment.

In practice that means three things.

• Cite your sources.A claim with a reference gets picked up more often than the same claim without one.
• Add numbers.Concrete figures with a year and a source are easier to quote than adjectives.
• Let experts speak.A literal quote from a named person is a block a model can take over whole.

And a negative result that is at least as useful. Keyword stuffing performed worse than doing nothing. The classic SEO reflex works against you here.

An honest nuance from the same study. In a test on a real engine (Perplexity) the gains were smaller than in the benchmark, and the effects differ a lot per field. So treat 40% as a ceiling, not a promise.

The five levers we use

• Website optimisation for AI.Letting crawlers in, schema.org markup, a readable information structure and an llms.txt, so a model knows who you are.
• Quotable content assets.Pages that answer a question in the first paragraph, with definitions, numbers and sources.
• Reviews and listings.Your presence on the platforms models look at when someone asks who is best.
• Mentions outside your own domain.Trade media, press releases, guest contributions. This is the slowest and heaviest lever, and the deciding one.
• Measurement.Fixed questions, several engines, again every month. Without a baseline you do not know if anything works.

How to measure whether it works

GEO has no ranking position and no Search Console. What you can do is write a fixed set of 20 to 30 questions your buyers really ask, put them to several engines every month, and track whether you get named, how you are described and which sources the engine cites.

That last part is the most useful. The sources quoted today for your core question are the places you want to be next quarter.

Watch out for one trap. A single screenshot is not a measurement. Models give different users different answers at the same moment. Only a repeated measurement across several engines tells you something.

Sources

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, GEO: Generative Engine Optimization, KDD 2024. · Rand Fishkin, SparkToro zero-click research, June 2026. · Ahrefs AI traffic research, 2025-2026. · 6sense, B2B Buyer Experience Report 2025, November 2025.

Read more