What is generative engine optimization (and what it is not)

Generative engine optimization, or GEO, is the work of getting a source used and named inside an answer that a model writes. The term came out of a 2023 research paper that asked a simple question: when a generative engine composes an answer from several sources, what makes one of those sources survive into the visible text, with a name attached.

That is the whole idea. It is narrower than most of what gets sold under the label, and knowing the boundary is what keeps the work honest.

What is actually different from SEO

Classic search returns a list. Ten links, each with a title and a snippet, and the ranking is the product. You can see your position, you can see who is above you, and the click is yours to win.

A generative answer returns a paragraph. Somewhere behind it, the model retrieved a set of documents, decided which ones to lean on, wrote a text, and, depending on the surface, attached a handful of links. Three things move that were not moving before.

Retrieval replaces ranking as the first gate. Before anything can be cited, it has to be retrieved. On the assistants with web access this is usually a search query the model writes for itself, and it rarely matches the phrase a person typed. Someone asking "which company should I use to book award flights" produces retrieval queries that no keyword tool would have handed you.

Citation is not proportional to use. A model can lean heavily on a document and cite something else, or cite a page it barely used. The link list under an answer is a partial record of the reasoning, not a ledger.

The answer changes between runs. Same question, same model, one hour later, different wording and sometimes a different set of names. This is the single fact that breaks the habit of "I asked ChatGPT and we were not there".

What is not different

Almost everything else. The page still has to exist, be reachable, load, and say clearly what it is about. Every assistant with web access is reading pages through the same plumbing a crawler uses: a URL, a status code, HTML, headings, text.

The basics decide whether the page is available to be used at all. A page blocked in robots.txt is not retrieved. A page whose content only exists after JavaScript runs is a coin flip. A page with no clear claim in the first two hundred words gets summarised into someone else's paragraph.

So GEO does not replace technical work. It sits on top of it, and it adds one thing that classic SEO never had to care about: whether your name survives the rewrite.

What can be measured today, and what cannot

This is where most of the confusion lives, so it is worth being blunt about the two halves.

Measurable, on your own property. Whether an AI Overview block appears for a given query, and which domains it cites, because that block comes back in the search results payload like any other block, which is the subject of a separate piece on measuring the click. Whether a named assistant mentions your brand in reply to a fixed question, because you can ask it, on a schedule, and store the answer. Which sources it cited alongside you. Where in the answer your first mention falls. Clicks and impressions on your pages, from Search Console, which is free and yours.

Not measurable by anyone right now. How many people asked that question inside an assistant. How many of them saw your name. How many clicked. There is no impressions report inside ChatGPT or Gemini, and any vendor quoting you a volume for a prompt is modelling, not measuring. If a number like that appears in a deck, ask where it came from.

We say this in public because the distinction survives contact with a client meeting. A measurement you can reproduce beats an estimate you have to defend.

Three checks you can run this week

None of these need a tool. They need a spreadsheet and an hour.

One. Write the ten questions your buyer actually asks. Not keywords. Questions, in the words a person uses out loud. "Is it worth paying for X." "What is the difference between X and Y." "Who does X for a small team." These are the shape of the input a generative engine gets.

Two. Ask each question three times, in a clean session, and save every answer as text. Three runs, not one. Then count: how many of the three named you, at what point in the text, and which other companies appeared every single time. The names that show up in all three runs are the incumbents in that answer, and they are the ones you are actually competing with.

Three. Take the sources the answers cited and open them. Not your pages. Theirs. Comparison pages, forum threads, directories, one review site nobody in your team reads. That list is the retrieval surface for your category, and it is usually shorter and stranger than the list of sites you have been trying to rank against.

By the third check most teams find the same thing: the pages being cited are not the pages they optimised.

What we measure in Crownler

We built the instrument around the same boundary. Once a week, for each question a client wants watched, the tool asks ChatGPT, Claude and Gemini with web search on, and stores the complete answer. Not a score, the text. From the stored text it extracts, deterministically, which watched brands were named, how many times, the character offset of the first mention, a short snippet around it, and every source domain the answer cited.

Two details are deliberate and worth stating.

The models we query are the accessible ones, not the top tier. The person asking your category question in the wild is on the free or cheap plan, and measuring on a model most of your buyers never touch would give a prettier number and a worse read.

The result is cached for forty hours. Asking the same question twice on the same day costs twice and returns nearly the same answer, so the second click is served from what we already paid for. Anyone can force a fresh run, and then they pay for it, on purpose.

Alongside that, the daily watch checks Google and Gemini for the fronts a client marked, records the position, and raises a flag when it drops three places or more, or when it leaves or enters the top ten. Every check is a row, never an overwrite, so the thirty day line is a record and not a snapshot.

Where this leaves the term

Generative engine optimization is a real problem with a small, measurable core: get retrieved, get used, get named. Around that core there is a lot of language that promises what nobody can currently see.

If you take one thing from this, take the habit: before you accept any number about AI visibility, ask what was stored, when, and whether you can read it back. Every claim in this piece is one you can check on your own site, this week, for free.

Crownler reads your site as a map, measures each intention on Google and in AI answers, and buys external data only when you ask for it. What the tool does · How the method works · How to see if ChatGPT and Gemini cite your brand · AI Overviews, measured