What is GEO (generative engine optimization)?

GEO, short for generative engine optimization, is the work of getting a business named and cited when someone asks an AI assistant such as ChatGPT, Gemini or Google AI Overviews for a recommendation. It matters because most businesses are barely there yet: across the 90 European brands in our 2026 study of 187,810 AI answers, the median brand appears in 5% of the answers about its own category.

SEO gets you onto the list; GEO gets you into the answer. I'll say it upfront: a large part of the work that gets you named happens on other people's pages, and the part on your own site starts with letting the bots in.

Nobody publishes how often people ask an AI assistant a given question, so GEO is measured on a fixed list of buyer questions you choose, and every result describes that list.

Below I explain how it works, how it is measured and how to start. In this guide I cover:

• What is GEO?

• How AI assistants pick companies

• How GEO works

• GEO compared with SEO and AEO

• How GEO is measured

• What the research says about GEO

• Where to start

What is GEO?

GEO is the work that makes your business easy for an AI assistant to find, read and repeat when a buyer asks it which company to choose. Behind every AI assistant runs a generative engine: a search engine joined to a language model.

The engine runs searches, reads the pages it finds and writes one answer with links to the sources behind it. ChatGPT, Gemini, Google AI Overviews, Perplexity and Copilot all work this way, so a buyer gets a short list of named companies.

GEO works on two results of that answer:

• Mention: the answer names your brand in its text.

• Citation: the answer links one of your pages as a source.

Sidenote: GEO is also spelled "generative engine optimisation" in the UK, and parts of the market call it AEO or AI SEO. The work is the same; the names table further down shows which step each name covers.

The spread across our 90 tracked brands shows how low the base is:

Half of the 90 brands appear in fewer than 5% of AI answers about their own category, and even the brand at the 90th percentile reached only 29.6%. That gap is what GEO works on.

How AI assistants pick companies

An AI assistant names the companies whose pages its searches return and whose pages answer the question plainly. It works in five steps:

1. You type a question in your own words.

2. The assistant rewrites your question into several searches.

3. It runs those searches and collects the pages they return.

4. It reads each page and keeps what is useful.

5. It writes one answer and puts source links next to the parts it used.

The five steps take a few seconds:

Because the assistant rewrites your question into several searches, a page has to match one of those searches before it can be cited. Google's documentation on AI features calls this a "query fan-out": AI Overviews and AI Mode may issue "multiple related searches across subtopics and data sources" to build one response.

Take an invented example that runs through this guide. A factory manager types "who makes food-grade stainless steel tanks for dairies in Europe". The assistant answers with three suppliers: one named as the dairy specialist, one as the option with the shortest lead time and one as the cheapest for small volumes, each with a source link beside its name. The pages that win say plainly what each company makes, for whom, in which countries and to which standards. A page that only describes the market gets read and then left out.

Our own tracking shows the same steps on a real question. A business owner asked ChatGPT, in Lithuanian, for a marketing agency that could get them recommended in ChatGPT answers:

The question became 3 searches before the model wrote a word, and of the 29 pages it read, it cited 4. Being read by the engine is the easy part; being picked as a source is where most pages drop out.

The answer itself named three agencies, and each name came with its source:

Each named company has a source link beside it, and each linked page says plainly what the company does, for whom and where. That is the same shape as the tank example.

Passage shape matters at the reading step: when Metehan Yesilyurt of Peec AI scored two passages with 12 open rerankers, a direct shortlist of tools beat a product description on almost every model.

Further reading: how AI engines decide which brands to recommend.

How GEO works

GEO work starts with access and ends with counting, and each of the four steps changes something the tank supplier can check:

The first two steps are the work and the last two are the measurement. In order:

1. Open the door. Let the AI companies' bots reach and read the site, so the tank supplier's product pages can be collected at all. That means a robots.txt file that allows them and pages that show their text without JavaScript, because Vercel measured that the crawlers behind ChatGPT and Claude fetch JavaScript files and do not run them.

2. Give the engines something worth citing. On the supplier's own site, that means key pages whose first two lines answer a buyer question, such as which tank sizes it makes and for which dairies, backed by numbers and standards. Off the site, it means being listed and described correctly on the review sites, forums, directories and comparison articles the engines already read.

3. Watch what gets cited. Ask the same buyer questions every day and record which of the supplier's pages the engines use. That routine is prompt tracking, explained in our guide.

4. Count the recommendations. Record how often the engines name the supplier, on which questions and in what words, and what that turns into in enquiries.

Step 1 hides one trap, because OpenAI alone sends three different bots:

OpenAI's crawler documentation says each robots.txt setting is independent, so a site can refuse training through GPTBot and still appear in ChatGPT search through OAI-SearchBot.

Sidenote: a robots.txt line that blocks GPTBot stops training use only. Check which bot your file blocks before you change it.

Step 2 matters off your own site as much as on it, since engines often name a brand while citing other people's pages:

Of the 87 brands the engines named in our study, 34 (39%) never had their own website among the 30 sites cited most in their market, which is why step 2 reaches past your own site.

GEO compared with SEO and AEO

SEO and GEO share one website, and AEO, AI SEO, LLMO and AIO are other names for GEO or for one of its steps.

GEO and SEO

SEO and GEO differ on four things a marketing team decides:

• What you try to win: SEO wins a place in a list of links, while GEO wins your name inside one written answer.

• What decides it: keywords, links, site authority and clicks decide SEO; for GEO it is whether a passage answers the question with facts and named sources.

• How you measure it: SEO is read in positions, clicks and impressions, GEO in the share of answers that name you and the citations of your pages.

• What backfires: thin pages written to hit a keyword fail in SEO, and a page that talks around the question gets skipped in GEO.

Here are the first three side by side:

The two feed each other. AirOps found that 43.2% of pages ranking first in Google were cited by ChatGPT, 3.5 times the rate for pages beyond Google's top 20. Google's own guide to generative AI search goes further: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." I read that as agreement more than conflict: the foundations are SEO, and GEO adds the off-site work and the measurement in AI answers.

AEO, AI SEO, LLMO and AIO

Your agency or your CEO will use one of these names, so here is what each usually means, which step it covers and what we call it:

Five of the six names describe GEO or one of its steps, and AIO is too vague to use, which is why we call the whole job GEO. Each name has its own guide: our guide to AEO, our guide to AI SEO, our guide to LLM SEO and our guide to AIO. Wikipedia's article on GEO notes that usage varies "with no consensus definition", so ask a vendor which of the four steps they will run on your account and who measures the result.

Buyers search for some of these names far more than others:

"AI SEO" gets 6,600 searches a month in the US and 1,600 in the UK, ahead of "generative engine optimization" at 4,400 and 880, so that is the name most buyers will use with you. The short forms "geo" and "aeo" are left out, since both also mean other things.

How GEO is measured

GEO is measured on a fixed list of buyer questions, asked in the same tool over time, and two numbers carry most of the weight:

• Visibility: the share of tracked answers where an engine names your brand, as Peec AI defines it.

• Share of voice: your brand's share of all brand mentions on the same questions.

Back to the tank supplier: ask twenty buyer questions once each, get named in five of the answers, and visibility is 25%.

We show the two numbers on our own brand, because it is the one project we can publish without a client's permission:

On 88 Lithuanian buyer questions with no brand name in them, Agenzy reached 25.9% visibility and 21.8% share of voice from 17 September to 16 October 2026, first of the 66 tracked brands on both. We read both daily in Peec AI; the free route is to ask your ten questions in a logged-out ChatGPT window every week and write down who gets named.

A client's numbers move over weeks, as one software company's did:

Visibility across 100 tracked buyer questions went from 0.25% to 11.9% in 14 weeks, after the bot block was lifted and the pages were rewritten. The client is Outcraft, which sells payment recovery software to online shops: in May 2026 its robots.txt blocked the AI bots, and ninety days later it held 98% share of voice on its core category question, as our B2B SaaS case study sets out.

Sidenote: the two figures measure different things. The 98% is share of voice on one category question; the 11.9% is visibility across all 100 tracked questions.

Each engine needs its own reading, because they cite different sources:

Across 109 business projects in our study of the sources AI engines cite, ChatGPT, Google AI Overviews and Perplexity shared a median of 2 of their top 10 cited sources, so one engine's result says little about the others.

There is no universal good score, because a category with four suppliers and a category with forty do not compare. Read your own trend on the same questions in the same tool over a month or more, since the same question can produce a different answer an hour later.

Further reading: the GEO metrics we report and AI share of voice, explained.

What the research says about GEO

The term GEO comes from a paper by researchers at Princeton University and IIT Delhi, presented at KDD 2024 and published on arXiv. They built a test set of 10,000 questions across 25 subject areas, took the top five Google results for each, rewrote those pages nine different ways and measured how much of each page reached the answer.

The paper's two main results fit in one picture:

Adding quotations from credible sources lifted a page's share of the answer by 41%, statistics by about 30% and citing sources by 27%. Keyword stuffing, the old SEO trick, scored below the untouched page, and about 10% worse when the team tested it on Perplexity.

The finding that matters most for a smaller company is who gained: citing sources raised the fifth-ranked page's share by 115%, while the first-ranked page lost 30%.

Sidenote: the paper ran on GPT-3.5 and a test set. Its authors say the methods will need to change as the engines change, so read 41% as a best case.

Later work is more cautious. C-SEO Bench, a 2025 benchmark across six domains, found most such rewriting methods largely ineffective, and their gains shrank as more sites adopted them.

My reading of both: write pages that answer the question with sourced facts, because that helps readers on any engine, and judge the result on your own buyer questions over time. Our own published data is on our research page.

Where to start

First check whether GEO fits you this year. I tell a business to spend the money elsewhere for now when:

• Its buyers never ask an assistant, as with emergency call-outs booked by phone, walk-in trade or supply bought through a tender portal.

• Nobody can change the site this quarter, so pages that need JavaScript or a security rule that blocks the bots stay as they are.

• The plan runs shorter than half a year, which leaves no time for content to be read and counted.

If none of those applies, three steps are worth doing this month, in this order:

1. Check the door. Open your robots.txt file, confirm that the AI search bots are allowed, OAI-SearchBot for ChatGPT among them, then check that your firewall does not block them anyway.

2. Write down ten buyer questions. Use your buyers' words and no brand name, ask each one in ChatGPT and Gemini and note who gets named, using how we build prompt sets as the method, or ask us for a free AI visibility check of your category.

3. Rewrite the three pages that matter. Take the three questions worth the most to you and put the answer in the first two lines of the page that should win, followed by the facts behind it, as in how to rank on ChatGPT in 2026.

When we run it, month one covers the audit, the technical pass and the baseline, content starts in month two and the before and after report comes in month three (our GEO service page).

Further reading: our longer guide to where to start with GEO, and more explainers in our guides.

Questions people ask about GEO

Do I need GEO for a small business?

You need GEO when buyers in your category ask an assistant before they shortlist. Size matters less than on Google: the GEO paper found that lower-ranked pages gained most from citing their sources, so one page that answers one buyer question is a real start.

Which AI engines matter in Europe?

ChatGPT carries most AI chatbot traffic in Europe, with Gemini second, as the September 2026 shares show:

ChatGPT holds 79.0% in Statcounter's European reading, 80.0% in the UK and 75.2% in the US. Statcounter counts web visits, so Gemini's 12.7% in Europe likely misses use inside Google's own app.

Can I do GEO myself?

You can do parts of GEO yourself: open robots.txt, allow the AI search bots and rewrite your main pages so each answers a buyer question in its first two lines. The hard parts are the daily measurement and the mentions on outside pages the engines trust, and our guide to prompt tracking shows where most teams stop.

What does GEO cost?

Agenzy's GEO work starts from 5,000 EUR per month ex VAT, as our GEO service page states. In our August 2026 scan of about 100 GEO providers, the most common band for mid-market work was 3,000 to 8,000 EUR a month, moving with the number of markets and languages. We compare published prices in what a GEO agency costs.

Is GEO a fad?

GEO is a new name for a change in buyer behaviour that is already measurable: people ask assistants which company to choose and act on the answer. Nikhil Lai, Principal Analyst at Forrester, wrote in November 2025 that AEO is "significantly, but not fundamentally, different from SEO". I agree: the same site serves both, and what changes is where you get named.

Final thoughts

GEO is the work of getting your business named and cited when a buyer asks an AI assistant for a recommendation. It starts with letting the bots in and ends with counting how often the engines name you, on the questions your buyers really ask.

This week, write down ten questions your buyers ask, ask each one in ChatGPT and Gemini, and note who gets named. That list is your first measurement.

If you have a question about GEO this page doesn't answer, ask me on LinkedIn.