Generative engine optimization, without the hand-waving

Generative engine optimization is the work of getting your brand named and cited inside AI-written answers — in ChatGPT, Google's AI Overviews, Perplexity and the rest — rather than ranked on a list of links. Most of the underlying work is familiar; what changes is what counts as winning, and how badly it can be measured.

The term arrived with a 2023 research paper and was adopted by the industry faster than the evidence behind it accumulated. That gap is worth being honest about: some of what gets sold as GEO is ordinary content work with a new label, some of it is genuinely different, and a fair amount is invented. This guide separates the three.

What GEO actually means

A generative engine answers a question by writing prose, grounded in documents it retrieves. Three consequences follow, and they are the entire discipline:

• There is no position to hold. You are either named in the sentence or you are not. There is no second place that still gets a click.
• The answer changes between runs. Nothing is cached and served; it is composed. So a single observation tells you almost nothing about your standing.
• Other people's pages decide your fate. A model writing a balanced answer reaches for comparisons and discussions, not vendor pages. The fastest route into an answer is frequently a page you do not own.

GEO, SEO and AEO, stated plainly

These three get used interchangeably in marketing copy and they are not the same thing. The distinction that matters is what each one is trying to win.

AEO — answer engine optimization — is the older term, from the era of featured snippets and voice assistants, where the engine picked one existing passage rather than writing a new one. In practice the two now overlap so heavily that the distinction is mostly vintage. Where it still earns its keep is the reminder that being the answer and being in the answer are different goals.

The six surfaces, and why the number matters

“AI search” is not one place. It is several products with different retrieval behaviour, and a brand can be durable on one and invisible on another. Tracking a single engine and calling the result your AI visibility is the most common error in this whole category.

The split that matters most is between the engines that browse the live web and the ones answering largely from training. A page published last month can be cited by the first group within days and go unmentioned by the second for a year — and a tool averaging both into one score hides exactly that.

What the research actually shows

One finding has held up better than the rest, and it reframes the whole exercise. A SISTRIX study of 82,619 prompts across 17 weeks found that 86.5% of prompts keep a stable core of one to five domains that persist week after week — while roughly 89% of the remaining names rotate in and out.

Read that twice, because the second half is the important half. Most of the names you see in an AI answer are not holding a position. They are passing through. If you check once, see yourself, and conclude you are winning, there is a good chance you have observed rotation rather than standing.

Core vs Carousel

That is the distinction we built the product around, and it is measurable from data you already have to collect. Run the same question every week and count the proportion of runs in which a brand appears:

• Core — named in at least 70% of runs. A position. Getting here is a genuine win and losing it is a genuine problem.
• Emerging — 40% to 70%. Contested: on the way in, or on the way out, and worth watching closely.
• Carousel — under 40%. Present sometimes. Pleasant to see and not yet worth reporting as a result.

The classification needs a minimum of 4 observed cycles before it says anything — below that the rate is noise wearing a percentage sign. That floor is a deliberate constraint, not a limitation we are hiding: a verdict delivered on two data points is worse than no verdict, because someone will act on it.

This costs nothing extra to compute — it falls out of tracking the same prompts over time. The reason competitors do not report it is not cost. It is that “mentioned 3 times in 12” is a less flattering number to put on a dashboard than “mentioned”.

The practical checklist

In rough order of return. The first three are dull and account for most of the effect; the interesting ones are further down and matter less.

1. Be reachable by the crawlers that feed the engines

The AI user-agents are separate from Googlebot, and a robots.txt written before they existed will not mention them — so sites block them without knowing. This is the cheapest possible failure to have and the easiest to check: the free crawler checker takes a domain and tells you which ones can reach you.

2. Answer the question outright, near the top

A self-contained passage that states the answer is easier to ground a generated sentence in than a page that builds towards it. This is the most consistent pattern in the evidence, and it is also just clear writing.

3. Be specific enough to be worth quoting

Numbers, dates, named limits and stated constraints survive summarisation; adjectives do not. “Fast and affordable” carries nothing attributable. “Returns in about ten seconds, capped at 500 pages” does.

4. Get named on the sources engines already read

When a comparison article gets cited for your category every week and names four rivals and not you, being added to it beats publishing a fifth page on your own domain. This is the highest-leverage item on the list and the one least like traditional SEO.

5. Make the page machine-legible

Question-shaped headings, a visible date, structured data, an llms.txt. None of these is a lever on its own; together they are the difference between a page a model can parse and one it has to guess at. Our own 31-check audit scores exactly this, for free, on any URL.

How to measure it

1. Track questions, not keywords. “Best CRM for a ten-person agency” is a prompt someone types. “CRM software” tells you nothing about whether you would be recommended.
2. Run them on a fixed cadence. Weekly is enough. The point is comparable runs, not frequent ones.
3. Report a rate, never an event. Named in 9 of the last 12 is a fact. “Mentioned” is a headline.
4. Record who was named instead. Every absence names somebody else, and that list tells you which pages the engine is actually reading.
5. Keep the answers. A number you cannot trace back to the sentence that produced it is a number nobody can check, including you.

What nobody can measure

Every honest account of this discipline has to include this section, and most do not.

• Traffic from AI answers. Search Console folds AI Overview clicks into ordinary organic and analytics receives them as plain Google traffic. No tool can separate them — see the AI Overviews guide for the detail.
• Why you were chosen. The output does not expose the selection. Any “citation score” explaining your standing is a model of a mechanism nobody outside the engine has seen.
• How many people saw it. There is no impression count for an AI answer.
• Revenue attribution. Follows from the first three. Any figure here was assumed somewhere upstream.

What remains measurable is the thing that matters most: whether you are named, how often, on which surfaces, and who is named when you are not. That is enough to work with — and it is the whole of what the free visibility checker reports, with the stored answer behind every number.

Questions people ask