01Definition
What GEO is
People now ask AI assistants what to buy: ChatGPT, Claude, Gemini, Perplexity and Grok, and the AI answers Google shows above its search results. The assistant writes one answer and names a few brands and products in it. Generative engine optimisation, or GEO, is the work of making sure your brand and products are among them, described accurately, when that happens.
That last part matters most. Ask an assistant about a brand by name and it will usually say something about it. Ask it for “the best merino walking socks for blisters” and it names three or four brands. If you aren’t one of them, that person may never hear of you.
AI agents take this a step further. They are given a buying task, look through stores, shortlist products and pick one. A product an agent can’t find or can’t read doesn’t get picked.
02Comparison
GEO and SEO: what changes
GEO and SEO share a lot: clear, accurate pages, facts other sites back up, and markup machines can read. What changes is the result. A search page shows a list of links and the person searching picks one. An assistant writes one answer and recommends a few names in it.
Scroll the table sideways to see both columns.
| SEO | GEO | |
|---|---|---|
| What the person asking sees | A list of links to choose from | One written answer naming a few brands |
| What you’re after | A high position in the list | Being named, and described correctly |
| What it rests on | Links, page authority, on-page signals | What the assistant learned in training, plus the pages it reads while answering |
| The main number | Rankings and organic traffic | How often answers name you, with its likely range, against your competitors |
| How you measure it | Search Console, Semrush, Ahrefs | Asking each assistant the questions people ask, several times, and reading the answers |
| How steady it is | Rankings move gradually | The same prompt can get a different answer each time it’s asked |
So GEO doesn’t replace SEO. Good SEO helps, because assistants with live web search read the pages search engines rank well. But a page can rank well and still not be quoted, and a brand can rank modestly and still be named often because review sites and retailers talk about it.
03Mechanism
How assistants decide what to say
An answer draws on two things: what the model learned before it was released, and what it reads on the web while it writes the answer.
What the model learned
Models are trained on a large snapshot of the web and other text. Brands that appear often in sources the model treats as reliable, such as review sites, retailers, magazines and reference sites, become part of what it “knows” about a category. That knowledge only changes when a new model is trained, so it moves slowly and you can’t edit it directly.
What it reads while answering
Most assistants can now search the web while they answer. ChatGPT, Claude, Gemini, Perplexity and Grok all can, and Google’s AI Overviews and AI Mode are built on Google search. The pages they find, such as your product pages, a retailer’s category page, a review site’s list or a forum thread, shape what they say. These pages change all the time, which is why answers change too.
What tends to help
No one outside the AI companies knows exactly how each assistant chooses. These are the things answers visibly lean on, and you can check them for yourself by reading the answers and the sources they cite:
- Facts written down in textAn assistant can only repeat what it can read. Materials, sizing, durability, care and who a product suits belong in the page text. A size guide inside an image, or a claim with nothing behind it, gives an answer nothing to quote.
- Accurate prices and stock everywhereAnswers quote prices and stock from your site, your product feeds and retailers’ pages. If an old retailer page still shows last season’s price, an answer may quote it.
- Other sites that talk about youReview lists, retailer category pages, gift guides and forum threads are cited constantly. If the list an answer leans on doesn’t include you, the answer probably won’t either.
- A clear description of what you sellSay plainly what each product is and who it’s for, in the words people use, and say it the same way on your site, your feeds and your marketplace listings.
- Pages machines can readStructured product data, a readable feed, and robots rules that let AI crawlers and AI agents in.
04Framework
The questions people ask
People don’t all ask the same question. Someone with no brand in mind asks differently from someone comparing two, and a parent asks differently from a student. A brand can be named readily for one kind of question and be missing from another, so it’s worth measuring them separately.
Discovery and situation questions don’t name you, and that’s how people find brands they haven’t heard of. Count your mention rate on those prompts alone. Prompts that name you, such as comparisons and brand prompts, are worth reading for what they say about you, but they will nearly always mention you, so mixing them in makes the rate look better than it is.
05The commercial case
What being missed costs
It’s tempting to put one big number on this. We don’t, because no published figure we could trace to its source holds across categories, and your own numbers are better than anyone’s average.
What you can work out is this. Your analytics show how many visits arrive from AI assistants, how often those visits buy and what they spend. A measurement of how often answers name you shows how much of the category’s AI conversation you’re part of. Put those together and you can estimate what being named more often might be worth, as long as you keep the assumptions in view.
The calculator does this sum with your own numbers and shows each assumption it makes.
06Measurement
How to measure it honestly
Measuring AI visibility means asking the assistants the questions people ask about you and reading the answers. It’s easy to do badly, because answers vary from one ask to the next and a small test can look far more certain than it is.
What to count
- Mention rateThe share of answers that name you, on prompts that don’t name you. Worked out for each assistant separately, because they behave differently. Keep chat assistants and Google’s AI answers apart rather than pooling them into one number.
- How many answers arrivedAssistants time out, refuse and come back blank. A rate from 60 answers out of 81 planned isn’t the same as one from 81 out of 81, so report coverage next to the rate.
- Whether a match is really youA brand called “Summit” or “Fellside” can be matched by answers about something else. Read the answers before trusting the count.
- Who is named insteadYour rate next to your competitors’ on the same prompts. 40% means something different when the leader is at 80% than when it’s at 15%.
- What the answers rely onThe sources they cite, and the prices and stock they quote against your catalogue.
How many prompts, and how many asks
Ask each prompt several times, because one ask can’t tell you a rate. But what decides how certain a result is, is mostly how many different prompts you ask. Answers to the same prompt tend to look alike, so asking one prompt nine times still tells you about one prompt.
That’s why every rate needs a likely range worked out across prompts. The explorer below uses a 95% Wilson interval, the same kind of range Persipica puts on every rate.
Telling a real change from noise
When two results’ ranges overlap, the difference between them isn’t established: the two measurements can’t show it either way. That doesn’t mean nothing changed, only that the test was too small to tell. With 27 prompts and a mention rate of 25%, two tests will usually show a change only if the rate rises by about 49 points. Smaller real changes are there, but this size of test won’t reliably catch them; more prompts will.
Persipica does this for you: it sends the questions people ask about you as prompts to five assistants and Google’s AI answers, three times each on the chat assistants, keeps every answer and source, and puts a likely range and a prompt count on every rate. How Persipica measures sets out the method in full.
07Implementation
Where to start
Start by measuring, so you know where you’re missing and what the answers lean on instead. Then work through the gaps, most specific first.
1. Put the facts in text
Read the answers that name a competitor instead of you. What do they say about the competitor that they can’t say about you? Often it’s a fact the competitor publishes and you don’t: wash-test results, a size guide for growing feet, what the product is made of. Publish your own facts, only the ones you can stand behind, in plain text on the product or category page.
2. Fix wrong prices and stock
Check the prices and stock answers quote against your catalogue. Where they’re wrong, find the page they came from. It may be your own feed, your structured data or an old retailer page. Fix your own; ask retailers to fix theirs.
3. Reach the pages answers lean on
Look at which third-party pages the answers cite most: a review site’s list, a retailer’s category page, a magazine’s gift guide. If you’re not on them, that’s where to direct outreach, sampling or retail partnerships. You can’t edit those pages, so this is the slowest track.
4. Let agents and crawlers in
Check that your robots rules let AI crawlers read the site, that product data and prices are in structured markup and feeds, and that key information isn’t locked in images or behind scripts a crawler won’t run.
5. Measure again
Ask the same prompts after each change and compare the ranges. If they overlap, the change isn’t established yet; keep measuring rather than declaring a win. And even a clear change doesn’t prove your fix caused it, because the assistants and the web change too.
08FAQ
Frequently asked questions
What’s the difference between GEO and AEO?
Answer engine optimisation (AEO) is an older term from the days of voice assistants and Google’s featured snippets. GEO is about generative AI that writes a whole answer: assistants such as ChatGPT, Claude, Gemini, Perplexity and Grok, and Google’s AI Overviews and AI Mode. People use the two terms loosely and often mean the same thing.
How quickly do changes show up in AI answers?
There’s no fixed timeline. Assistants that search the web while answering can pick up a new or changed page once they find it. What a model learned in training only changes when a new model is released. The honest way to know is to measure before and after, with enough prompts to tell a change from noise.
Do different assistants give different answers?
Yes. They learn from different data, search the web differently and cite different sources, so a brand can be named often by one and rarely by another. Measure each assistant separately, and keep Google’s AI answers apart from the chat assistants rather than averaging them together.
Does GEO matter for smaller brands?
Often more than for big ones. Well-known brands are already written about widely. A smaller brand with a good product but little coverage may not be on the pages answers lean on, and that’s something it can work on.
How many prompts should a test include?
As many different prompts as you can manage, asked several times each. Asking more prompts narrows the range on every result; asking the same few more often barely does. With 27 prompts, two tests will usually show a change only if the rate moves by about 49 points from 25%, so small tests can only catch large changes.
Is a mention the same as a sale?
No. A mention rate shows how often answers name you. A mention can be neutral or even negative, and it doesn’t show what people buy. Visits and sales from AI assistants come from your own analytics.