Generative engine optimisation (GEO): what it is and how to do it

Generative engine optimisation (GEO) is the work of getting your business named and cited in the answers AI assistants write. This guide sets out what it is, how those answers are put together and how to start.

By AI Search Ranking. Published 7 October 2026. Last updated 7 October 2026. In Generative engine optimisation.

On this page

What generative engine optimisation is

Generative engine optimisation is everything you do so that an AI assistant mentions your business and links to your pages when a buyer asks it a question. In the United States it is spelt generative engine optimization. Most people shorten it to GEO.

A generative engine is a search tool that writes a reply instead of handing back a list. The main ones today are Google's AI Overviews and AI Mode, ChatGPT, Gemini, Perplexity and Claude.

GEO has two targets, and they are not the same:

A mention. The assistant names your company or product in its reply.
A citation. The assistant lists one of your pages as a source for what it said.

You can have either without the other. If you want the comparison with search engine optimisation, our guide to GEO vs SEO covers it. This page is about GEO itself.

Where the term comes from

The name was coined in an academic paper released in November 2023 and presented at a data science conference in 20242. Its authors wanted to know whether a website owner could change a page so that a generative engine used it more.

They tried nine ways of rewriting a page and tested them on a set of 10,000 questions2. Three findings are worth knowing:

Evidence helped most. Adding quotations and statistics to a page raised its visibility in the generated answers by up to 40%2.
Keyword stuffing backfired. Repeating the search terms scored lower than leaving the page alone2.
Results varied by subject. A method that worked in one field did less in another2.

The team also ran its methods on Perplexity, a live product, and reported gains of up to 37%2.

Read those numbers with care. The study measured how much of an answer came from a page. It did not measure clicks, enquiries or sales. It also dates from 2023, and the products it tested have been replaced by newer versions since.

Is it a real discipline?

The work is real. Whether it deserves its own name is disputed.

Google's position is that preparing for its AI features is "still SEO"1. It says those features are rooted in its core ranking and quality systems1. On that view, GEO is a new label for good practice.

We think the label earns its place for one reason. Google is one engine among several, and the others run search crawlers of their own3. A business can rank well in Google and still be missing from ChatGPT or Claude. Somebody has to check, and the checking and the remedies are different enough to need a word.

How an AI answer is put together

No company publishes its full recipe. The outline below comes from what the companies do say.

The question is widened. Google says its AI features can turn one question into a set of related searches, run together, to gather more material1. So the engine may look at pages on subtopics the buyer never typed.

Pages are fetched. Each engine has its own crawler for search. Google uses Googlebot. ChatGPT uses OAI-SearchBot, Perplexity uses PerplexityBot and Claude uses Claude-SearchBot3. A page those crawlers cannot reach is out of the running.

An answer is written and some sources are shown. This is the part nobody outside can see. No platform explains how it picks between two pages that both qualify3.

That gives GEO two jobs. The first is eligibility: make sure each engine can read you. The second is preference: give it a reason to choose you.

What the platforms ask for

Less than most advice suggests.

Platform	What its owner says
Google's AI Overviews and AI Mode	The page must be indexed and able to show in Search with a snippet. There are no extra requirements4
ChatGPT	Allow OAI-SearchBot. It is separate from GPTBot, which gathers training data3
Perplexity	Allow PerplexityBot. It is not used for training3
Claude	Allow Claude-SearchBot. Blocking it may reduce your visibility3

Of these, only Google describes at any length the content it prefers. It asks for pages that are helpful, reliable and written for people, with a point of view of their own1. It adds that you need no special structured data, no particular writing style and no pages cut into small pieces for AI1.

Our AI search overview quotes each company's document in full, with links.

What gets cited in practice

The documents say who is eligible. They do not say who is chosen. So we tested it.

On 4 October 2026 we asked ChatGPT, Gemini, Perplexity and Claude nine questions a UK buyer might ask, and noted every website each one cited3. Three questions were about business services, three about software and three about products.

Between them the assistants cited 209 websites3. Where a website was cited for a question, 85% of the time only one assistant cited it3. Each assistant has its own reading list.

The kind of website that gets cited also changes with what is being bought.

Type of website cited3	Business services	Software	Products
The supplier's or brand's own website	47%	40%	21%
Review and comparison sites	14%	22%	27%
Newspapers, magazines and trade publishers	6%	8%	36%
Blogs by agencies, consultants and other vendors	13%	30%	1%
Business directories	13%	0%	0%
Retailers and marketplaces	0%	0%	8%
Other	6%	1%	6%

For a firm selling services, its own website carries the most weight. For a product, publishers and review sites matter more than the brand's own pages. That is why the plan for a business supplier, a software company and an online shop should not look alike.

We assigned the website types ourselves, and some are a judgement. The figures come from one run of nine questions. They show a pattern. They are not a law.

The assistants do not behave alike

Being named and being cited are separate results, and the gap differs from one assistant to the next. In the same test we counted how often each assistant named a brand, and how often it also cited that brand's own website.

Assistant3	Brands it named	Times it cited the brand's own website
ChatGPT	77	15
Gemini	58	17
Perplexity	55	18
Claude	73	6

Claude named almost as many brands as ChatGPT and linked to their websites far less often. Perplexity cited the most sources in each answer, about 14 on average3.

So with some assistants the wording of the mention is most of what a buyer sees. That makes accuracy matter as much as presence.

How to do generative engine optimisation, step by step

First, check that the engines can read you

Open your robots.txt file and look for rules that block the four search crawlers named above. Then ask whoever runs your hosting whether a firewall blocks them, because a firewall can refuse a crawler that robots.txt allows.

Confirm that your important pages are indexed in Google. Our technical SEO work starts here, because nothing later matters if this fails.

Second, record where you stand

Write down the questions your buyers ask at each stage: the problem, the category, the comparison and the check on a supplier. Put every question to every assistant, more than once.

For each answer, note four things. Were you mentioned? Was your website cited? Was the description of you correct? Who appeared in your place? Our method explains why we fix the question set and keep it unchanged.

Third, put your facts in order

Assistants repeat what they find. If your site gives two different prices, or your directory profiles describe services you dropped, the answer will carry the confusion.

Make the basic facts identical everywhere you control: what you sell, whom it suits, what it costs, where you work. Where an assistant states something false, trace it to the page it came from and correct it there.

Fourth, make your pages worth citing

Answer the question plainly, near the top. Then back the answer with evidence a reader can check: your own data, named sources and direct quotations. That matches what the original research found2.

Write for the buyer. Google says outright that there is no need to write in a special way for AI1. A page that helps a person decide is a page an assistant can use.

Fifth, earn a place where the assistants look

Your baseline shows which websites the assistants cite in your category. Those are the review sites, directories, trade titles and comparison articles that matter for you.

Work on being listed, reviewed and written about on them, honestly. Google warns that chasing inauthentic mentions is less helpful than it seems1.

Sixth, measure every month

Run the same questions again and compare. In Search Console, the generative AI performance report shows how people reach you through Google's AI features1. Bing Webmaster Tools has an AI Performance report that counts citations in Copilot5.

Change one thing at a time where you can, so that you know what made the difference. Our GEO service does this work for you.

What does not work

Keyword stuffing. It did worse than doing nothing in the research2.
Shortcut files and tags. Google says no new machine-readable files or special markup are needed for its AI features1. None of the other platforms we checked asks for one either3.
Paid or invented mentions. They put your name beside claims you cannot stand behind, and Google says they help less than people think1.
Pages produced in bulk. Google asks for content that is not a commodity1. A hundred thin pages give an assistant nothing to quote.
A screenshot as proof. One answer on one day shows very little, because the next run may differ.

What nobody outside these companies knows

Three things are unknown, and anyone selling GEO should say so.

Nobody outside can say how an engine picks one qualifying page over another3. Nobody can say how much of a recommendation comes from a live search and how much from what the model learned in training3. And nobody can say why one question gets two different replies an hour apart.

So a guaranteed citation is a guess with a price on it. What can be promised is the work: remove what blocks you, give the engines something accurate to cite, and measure on fixed questions over time.

Common questions

Can we do this ourselves? Yes. The first three steps need care more than specialist skill. Most teams want help with the fifth, because earning coverage takes contacts and time.

Which assistant matters most? The one your buyers use. Ask a few customers. If you cannot find out, check all of them, because being cited by one says little about the rest3.

Does it suit a small business? Yes, and the baseline is cheaper for a small one. A local firm may need only a handful of questions to see where it stands.

How soon will we see a change? We do not promise a date. Access problems are fixed in days. Coverage on other websites takes months. We plan a first programme of ninety days because that is long enough to fix the foundations and measure a first change.

To see how the assistants describe your business now, book a discovery call.