What is GEO?

GEO is short for generative engine optimization. It has nothing to do with geography or geomarketing: it is about the answers AI systems write, and about who gets named in them.

You can end up in such an answer in two ways, and they rarely coincide. If your name is in the text, you are mentioned. If a page of your site is in the source list below it, you are cited. A GEO report that does not say which of the two it counts cannot be compared with any other report.

An assistant with web search turns your question into one or more searches, retrieves the pages that come with them and writes an answer from that text. This is called retrieval-augmented generation. Without web search the model draws on what it read during training: then it names names, but no sources.

Why that matters: whoever reads the answer often has no reason to click any further. If you are not in that answer, you do not exist in that search, however well your page ranks in Google.

The term comes from the paper GEO: Generative Engine Optimization by Pranjal Aggarwal and co-authors from Princeton University, IIT Delhi and other institutions. It went online in November 2023 and was presented at KDD 2024, a data mining conference. The English Wikipedia attributes the term to the same paper.

The authors built a benchmark of 10,000 search queries from 9 datasets and 25 domains: 80% informational, 10% transactional and 10% navigational. They applied 9 methods to one source and measured how much of that source ended up in the answer: in words, weighted by position in the answer, and as an impression, rated by GPT-3.5 on seven aspects.

Two caveats belong with this. The benchmark was 80% informational queries; queries with a purchase in sight were 10%. And the biggest gain, +115.1%, only holds if all competing sources make the same change: then the source lower in the list rises and the top one drops.

The market has not settled on a name yet. In the Netherlands the full term, generative engine optimization, is searched less often than the informal variant AI SEO. And the most searched term of all is AI Overviews, the name of Google’s answer block above the search results. Dutch searchers recognize the phenomenon before the discipline it belongs to.

Below that sits a second row of names for the same work. The English Wikipedia lists answer engine optimization (AEO), LLMO, AIO and AI SEO as other names for generative engine optimization, without naming a difference. AEO dates from the era of answer boxes and voice assistants, and the debate about which name is right is still going.

For anyone writing about this, the fragmentation has a practical consequence. A page that only uses the full term misses most of the people searching for it. This page sticks with GEO and names the variants where they belong.

Four things decide whether a business ends up in the answer, here in the order in which you tackle them.

Across all our measurements, a Dutch buying question draws on average 34.2 different sources. Weighted by how often they are actually used, 19.0 effectively remain; the narrowest question comes out at 3.3. A ChatGPT answer with sources shows an average of 3.4 of them. Own research + client data

So for each question there is a set of about 19 candidates, from which ChatGPT fills a few slots each time. The more of them name you, the more often the answer names you.

You can do this work yourself, have an SEO agency do it, or outsource it to a party that measures and executes. Which kinds of parties there are, what they do and what they cost is in who does GEO optimization (in Dutch). What you hand over to AIGeschikt is on outsourcing GEO, the plans on AIGeschikt’s pricing page.

Whether blogs work for this, we measure on our own site with a control group: GEO optimization with blogs (in Dutch).

By asking the assistants a fixed list of buying questions again every week, and recording for each answer whether your name is in the text and which sources are listed. Typing your own name into ChatGPT once is not a measurement.

Those five do not add up to one number: each has its own source and says something different. The concepts around them, from citation to zero-click, are explained in AI visibility in plain language (in Dutch).

Measuring yourself is possible: freeze a list of buying questions, ask them to the same assistants every week, record your name and the sources for each answer, and only look at the trend after four weeks. Then divide the number of returning sources by the total: that number tells you how stable your market is.

For each client, AIGeschikt runs 1,200 measurements per week: 100 fixed buying questions on ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. How that works is described under measuring AI visibility. That is the measurement of the service; the research figures on this page come from a separate measurement.