What is GEO (generative engine optimization)?

The acronyms all mean the same thing

GEO (generative engine optimisation), AEO (answer engine optimisation), LLMO, AI SEO, AI visibility, LLM visibility. Different vendors picked different words for the same activity, and some of them picked a word specifically so they could claim to have invented a category. If someone tells you AEO and GEO are fundamentally different disciplines, they are selling you the difference.

We use "GEO" because it's the most common, and "AI visibility" when we're describing the thing being measured rather than the work.

Why it isn't just SEO

Search gives you ten links and lets you choose. An assistant gives you an answer and a short list of names. There is no page two, and for most questions there is no fourth or fifth option either.

That changes what "ranking" means. Being the 8th best result on Google still earns clicks. Being the 8th name an assistant would have mentioned, if it had kept going, earns nothing at all. Visibility is much closer to binary than it used to be, which is why the first thing worth measuring is not a position but a yes or no: does it say your name?

The other change is where the answer comes from. A model may name you because it read your site, or because it read a directory, a Reddit thread, a local news piece or a competitor's comparison page that happened to mention you. Your own website is one input among many, and often not the loudest one.

Branded and unbranded questions — and why only one of them tells you anything

This is the single most important idea on this page, and it's the thing most free tools get wrong.

Ask an assistant "tell me about Acme Dental" and the answer will contain the words "Acme Dental". Of course it will — you put them there. A tool that counts that as a mention is measuring its own question. We call that an echo, and we score it separately from a real mention for exactly that reason.

The question that carries information is the one that never says your name: "best family dentist in Leeds". If you come up there, an assistant reached for you unprompted, which is the entire thing you're trying to find out. That's why 25 of the 30 questions in our report never mention your brand.

Branded questions are still worth asking, but for a different purpose: they test whether the assistant has its facts right about you. Wrong opening hours, services you dropped two years ago, or a confident description of a similarly-named business in another city. That's a different problem from invisibility, and it's often a more urgent one.

Why you get a different answer every time you ask

Ask the same assistant the same question twice and you can get two different lists. This is not a bug you can configure away — the model picks a fresh answer from a range of likely ones every time, and grounded ones also depend on whatever the live web returned that second.

The consequence is uncomfortable for everyone selling in this space, including us: a single run is an anecdote, not a measurement. If a tool asks once and reports "you have 40% visibility", that number has an error bar around it wide enough to drive a truck through, and the tool almost certainly won't tell you how wide.

There are only two honest responses. Ask more times, or say plainly how uncertain the number is. We do a limited version of the first: your 3 most important questions get asked 3 times each, so we can tell you which findings held up and which appeared once. And we do the second throughout, because 30 questions is a sample and we say so on the method page.

The surfaces disagree with each other, and that's the useful part

"AI search" is not one thing. We look at 5, and they behave differently enough that a single score across all of them would hide the finding:

• ChatGPT — with web search. The assistant most of your customers actually use.
• Gemini — with Google Search. The default assistant on Android and across Google's apps.
• Claude — from memory, no web access. Heavily used by business buyers.
• Google AI Overviews — live search results. The AI answer your customers see without asking for it.
• Google AI Mode — live search results. Google's dedicated AI answer tab.

The pairing that matters most is an assistant answering from memory against the same assistant answering with web search on. If it names you only when it can search, the model doesn't know you — it found you. If it names you from memory, you're actually in there. Those two situations need completely different work, and a tool that reports one blended number cannot tell them apart.

For the numbers behind this: our 34-business benchmark and pricing are both public.

A memory-only version of Gemini is also on the roadmap — the grounded Gemini above is live today. See the method page for why, and what we're doing about it.

What actually moves it

Here is the part that costs us money to say. Google's own documentation on its AI features states there are "no additional requirements … nor other special optimizations necessary" beyond being indexable and eligible for a snippet.

There is no schema tag that makes an assistant like you. There is no "GEO keyword". Anyone selling a secret technique is selling a story, and the ones selling volume (a thousand blog posts, mass directory submissions) are selling the thing most likely to get you filtered.

What does appear to matter is unglamorous:

• Being machine-readable at all. A noindex or nosnippet directive, or a site that renders only in JavaScript, can remove you from AI answers entirely. This is the most common finding we see that is both severe and trivially fixable.
• Saying concretely what you do and where. Models reach for specifics. "Award-winning solutions for the modern enterprise" is invisible in a way that "sports physiotherapy in central Bristol, evenings and Saturdays" is not.
• Being described consistently somewhere other than your own site. Directories, review sites, local press, industry lists. Your own claims about yourself carry the least weight of anything a model reads.
• For local businesses, your Google Business Profile. Category, hours, and name/address/phone consistency across the web feed the map results that AI answers lean on heavily.

Notice that all four of those are things you'd want to be true anyway. That is not a coincidence, and it's the honest summary of the field as it stands.

What GEO is not

• Not a ranking you can buy. There is no paid placement in an assistant's recommendation list. If there ever is, it will be labelled as an ad.
• Not attributable to revenue. AI referrals mostly arrive as direct or unassigned traffic. Nobody (us included) can currently tell you that an AI mention produced a sale. Anyone who claims a revenue figure is modelling, not measuring.
• Not a keyword tool. Search volume doesn't transfer. A question nobody types into Google may be extremely common in a conversation with an assistant, and the reverse is just as true.
• Not permanent. Model updates reshuffle answers. Anything measured today is true today; the useful move is to re-measure later and compare, not to treat one number as a standing fact.
• Not urgent for everyone. If nobody would ever ask an assistant for a business like yours without already knowing your name, this whole field is currently irrelevant to you, and you should spend the money elsewhere.

So should you do anything about it?

Honest answer: you can't tell without measuring, and the measurement is cheap relative to guessing. The failure mode we see most often isn't a business that's invisible — it's a business that assumed it was fine because it ranks well on Google, and had never checked whether that transferred.

Start with a free checker. Ahrefs, Semrush and HubSpot all have one, and if one of them answers your question, you're done. When you want to know which competitors keep getting named instead of you, on which kinds of question, and whether it held up when asked again — that's when this becomes worth paying for.