What is generative engine optimization?

The short version

Generative engine optimization is the work of getting named inside answers that are written rather than listed. The name comes from a 2023 research paper; the practice overlaps heavily with SEO. Three things are genuinely different: there is no position to win, the answer changes between identical asks, and an answer is assembled from several pages at once, of which yours is the one you write. Everything else is SEO you already know how to do.

What is generative engine optimization?

A generative engine answers a question by writing a sentence, not by returning a list. Ask one who should I use for accounting in Manchester and you get two or three lines naming three or four firms.

GEO is everything you do to be one of those firms. In practice:

• knowing whether you are currently named, and in which kinds of question;
• knowing what the engine read before it answered;
• making sure your business is described accurately in those sources;
• making your own pages answerable rather than promotional;
• measuring again, honestly, to tell a real change from drift.

Where did the term come from?

“Generative engine optimization” was coined in an academic paper published in November 2023. The authors tested how changes to source text affect whether a generative engine cites it, and reported visibility gains of up to 40% from content-side changes alone.

Two things are worth taking from that and nothing more.

First, the question is researchable. This is not a field where nothing can be known. Source text demonstrably affects citation.

Second, the conditions were controlled. The gains were measured on retrieval-and-cite systems in a test setting, not in the wild against ChatGPT’s memory. Treat 40% as evidence that text matters, not as a number anyone can sell you.

The industry adopted the phrase faster than it agreed on a definition, which is why you now see GEO used for three different things: appearing in chat answers, appearing in Google’s AI Overviews, and — loosely — any SEO aimed at machines. We use the first sense.

What is actually different from SEO?

Three things. Everything else transfers.

There is no position

A search result is an ordered list; an assistant answer is a handful of names. You are on it or you are not, and the useful metric is the share of answers that named you across repeats. A tool reporting a “rank” for a generated answer is reporting an invention.

The answer is unstable

Ask the same question five times and the lists overlap but do not match. This has a practical consequence that people skip: a single check proves nothing. Measurement means repetition, and reporting a share rather than a yes or no.

An answer is assembled from several pages at once

This is the one that reorders budgets. In one local market we measured, a single answer rested on eleven pages on average — the business’s own pages together with directories, review sites, local roundups and trade listings.

Two consequences, in order. Your own pages are the only ones you write, so they have to state the facts a buyer asks about plainly and early; the homepage is rarely the page that gets quoted, and the service page usually is. Then the sources read alongside them have to say the same thing, because a contradiction between four descriptions of you is a reason to use somebody else.

What carries over from SEO?

More than the “SEO is dead” posts suggest. Generative engines read the open web. Pages that were worth ranking — clear, specific, corroborated, technically reachable — are the same pages worth citing.

Specifically: crawlability, page speed, clean structure, factual accuracy, topical depth, internal linking and a coherent site architecture all transfer directly. If you have been doing competent SEO you are not starting from zero; you are starting from a position you cannot yet see.

The full comparison is in GEO vs SEO.

What does GEO work consist of?

Ranked by what tends to shift things, not by what is easiest to sell.

1. Measure before changing anything. Without a baseline, the number you get in six weeks answers no question, because compared to what? has no answer.

2. Find the question types where you are absent. Visibility is rarely uniform. A business is often named constantly in general recommendation questions and never in price questions. Working on the ones you already win changes nothing.

3. Identify the sources behind those answers. Where citations are shown, write them down. That list is your to-do list, and it is usually shorter and duller than people hope: four directories, two roundups, a review site.

4. Get the facts about you consistent across those sources. Engines cross-check. A business with three different service lists in three directories gives a model less reason to name it than one described identically everywhere. This is unglamorous and it is most of the work.

5. Make your own pages answerable. Buying question as the heading, answer in the first two sentences, plain words. Generative engines extract passages; a page whose answer arrives in paragraph four does not offer one.

6. Publish something checkable. Original numbers, stated method, named examples. A source that says something specific gets cited over one that says something impressive.

7. Measure again and be honest. Movement under about five points is drift.

How do you measure GEO?

The same way you would measure anything noisy: repeat, and report a range rather than a point.

Ask the buying questions, not keywords. People type keywords into search boxes and sentences into assistants.

Ask each one several times per engine. Five is our default. Fewer and you cannot distinguish “usually named” from “named once”.

Report per engine, never averaged. When four assistants were asked the same questions about one market, they named 63 businesses between them, and only four were named by all four; 46 appeared in a single assistant only. An average would have produced a middling figure describing none of them.

Keep the raw answers. Any number should trace back to the text that produced it. If it cannot, it is not checkable, and an unverifiable number invites an argument you cannot win.

Our full method is written out here.

What GEO cannot do

It cannot guarantee a mention. The lists are not published, not stable and not for sale, and anyone offering placement is guessing. This is not modesty; it is the reason we will not sign a contract that promises it.

It cannot be judged on a single check. A screenshot of one good answer is worth very little, in either direction.

It cannot be fast. Your own pages are quick — written, approved, published. What takes time is bringing the sources that quote them into line, because a directory correction sits in somebody else’s queue.

It cannot fix a business that is genuinely hard to recommend. If a category is thinly covered on the open web, there may be nothing for an engine to read about anyone, and the honest finding is that the opportunity is to become the source rather than to compete for citations that do not exist.

Who is GEO worth doing for?

Worth it where customers ask for a recommendation before choosing: clinics, professional firms, trades, local services, and software people evaluate before buying.

Not yet where nobody asks an assistant about your category at all. The cheapest test is to ask your own buying questions and see whether the engine names anybody. If it declines to recommend in your category, there is nothing here to win this year.

Carefully in regulated categories. Healthcare, legal and financial questions are where assistants hedge most, and where a confident claim in your own copy can cost more than the visibility is worth.

GEO, AEO and AI search optimization

Broadly the same work under different labels: AEO is usually framed around direct-answer engines, GEO around generated text, “AI search optimization” is the umbrella. The term you use matters less than whether you are measuring.

• Answer engine optimization — the same work, framed around direct answers.
• AI search optimization — the full guide.
• GEO vs SEO — what transfers and what does not.
• How to improve AI visibility — the step-by-step version of the work above.

A worked example

Here is what the first month looks like when it goes normally. The numbers are from a measurement we ran on a local market; the business is anonymised.

The baseline. Twelve buying questions, asked five times each, across four assistants. The business was named in 38% of general recommendation answers and 0% of price answers. Two competitors appeared in nearly every answer on every assistant.

What the citations showed. Across the price questions, the same four sources came back repeatedly: two local directories, one review platform, and a “best in the city” roundup written by a regional publisher. The business was listed on one of the four, with an out-of-date service list and no prices.

What was actually done. Nothing on the homepage. The listings were corrected and completed on all four sources, with the same service names and the same price ranges used on the business’s own pages. One page was rewritten so that the price question was answered in the first two sentences rather than in a table three screens down.

What happened. The general recommendation share moved four points, which we reported as no change, because it is. Price answers went from 0% to being named in roughly a third — the kind of movement that is worth reporting, because it started from nothing and was concentrated in the questions the work targeted.

What we could not say. That the listings caused it. We can say what was changed, when, and what moved after — and that the movement was in the question group the work aimed at rather than spread evenly, which is weak evidence but better than none.

How long does this take?

Slower than it sounds, and the reason is structural rather than technical.

The half of the work that is yours — pages that answer the buying question — can be done in days. The half that confirms it cannot: a directory correction sits in their moderation queue, a roundup sits in a publisher’s editorial calendar, and neither responds to urgency.

A reasonable expectation: first measurement in a week, the work over four to eight weeks depending on how many sources need correcting, and a second measurement after that. Anything faster is either a very small source list or a supplier reporting drift.

What does it cost you in effort?

An hour at the start, mostly confirming that the question set sounds like your buyers — you will spot two or three questions nobody would ask, and correcting them matters more than anything else you do that week.

After that, whatever time your team needs to supply accurate information about services and prices. The bottleneck is almost never the measurement; it is getting a straight answer from inside the business about what it actually offers and for how much.

Frequently confused: GEO and AI Overviews

These are different surfaces and they move independently.

Being cited in an Overview says little about whether ChatGPT names you. We measure and report them separately for the same reason we never average across assistants: a combined figure describes neither. The Overview side is covered in Google AI Overviews optimization.

A short glossary

Generative engine — a system that answers by writing text rather than returning links. ChatGPT, Claude, Gemini, Perplexity, and Google’s generated surfaces.

Share of answers — the percentage of answers to a question type that named a given business, across repeats. The metric that replaces rank.

Citation — a source shown beside or under an answer. Visible on some surfaces, absent on others.

Drift — movement in a share caused by the system’s own variability rather than by anything anyone did. Under about five points, assume drift.

Where to start

Measure once. Everything before that is guesswork, and everything after it is checkable.

The free check runs a short measurement on your business and emails you what the assistants said, with no call attached. If you would rather see the shape of a full measurement first, the sample report is open on the site with no form in front of it, and pricing is published.