A plain guide to GEO (Generative Engine Optimisation)

Where the name comes from

As search shifted from links to generated answers, a new discipline appeared: optimising to be in the answer rather than near it. That is GEO. For twenty years the destination of a search was a page of ten links, and the whole trade of optimisation was built around competing for those links. Assistants moved the destination. Somebody asks a question, gets a written paragraph back, and two or three businesses are named inside it. There is no list to scroll and no page two. Nobody needed a word for optimising that until it started producing real enquiries — and then several words arrived at once, which is why you will also see AEO and AI SEO used for overlapping work.

How it differs from SEO

SEO earns rankings. GEO earns recommendations. The tactics overlap and the technical foundations are shared, but the priorities are not the same. Ranking rewards relevance, links and page-level signals. Being named rewards something closer to reputation: can this system identify you unambiguously, does the public record agree about you, and is there enough corroboration to make a recommendation feel safe. That last word matters more than people expect. A model that is uncertain about a business does not gamble — it recommends somebody it is confident about, or it avoids naming anyone at all.

How an assistant actually builds an answer

Three sources feed most answers, in varying mixes. Training data: what the model absorbed when it was built — broad, dated, and impossible to edit directly. Live retrieval: the system searches the web at the moment you ask and reads what it finds, which is why some assistants show clickable citations and why a change you made last week can show up quickly there. And connected structured sources, such as a mapping or business-listing service the platform has licensed. Knowing which mix a platform leans on tells you which lever moves it. A business that is strong in live retrieval but weak in the wider record will be named by some assistants and ignored by others — and that split is diagnostic, not random.

The building blocks

Five, in rough order of how quickly they pay. A clear identity: one canonical name, address, phone number and description, repeated identically everywhere it appears. Structured data: machine-readable markup that states what a reader would otherwise have to infer, and that matches the visible page exactly. Answer-shaped content: pages that state their answer in the opening lines instead of burying it beneath a story about your journey. Credible third-party references: reviews, accreditations, registers, trade bodies, local coverage — sources that corroborate you independently. And crawler access: making sure the AI user agents are not blocked before any of the rest of it can matter.

Entity clarity, since it decides most cases

An entity is the machine-readable idea of your business. It is what lets a system say "this company" rather than "some firms in that area". Entities break in mundane ways: a trading name that appears nowhere in the official record, a limited company name that appears nowhere on the website, an address updated in three places out of seven, a Google Business Profile in a category that no longer describes what you sell. A human reads past every one of those. A machine treats them as conflicting evidence and hedges. Fixing them is patient list-work — unglamorous, cheap, and consistently the thing that moves answers first.

Where llms.txt fits, honestly

llms.txt is a plain text file at your site root that lists your important pages and describes what your business is, aimed at AI systems rather than readers. It is a sensible courtesy, it takes minutes, and it costs nothing. It is not a ranking signal. No major platform has committed to reading or obeying one, and we have never seen a case where adding it alone changed an answer. Treat it as housekeeping — do it, then move on to the work that actually decides whether you are named.

What does not work

Hidden text aimed at models rather than readers — the modern white-on-white keyword, with the same eventual outcome. Buying links. Mass-produced location pages with the town name swapped, which readers and machines both recognise instantly. A hundred AI-written articles with nothing new in them, which mostly add noise. Manufactured community mentions. And any product sold on the claim that it will force an assistant to recommend you — there is no submission form, no paid inclusion, and no tag that compels a model. Anyone selling that is selling nothing.

How to measure it without fooling yourself

Ask the assistants, and write down what they say. Use the questions a customer would genuinely type, not your brand name — brand-name prompts almost always return something and tell you nothing. Keep the wording identical between runs, or you are measuring your own phrasing rather than your visibility. Record the date, because these systems change without notice and an undated screenshot is close to worthless six months later. Run it monthly; weekly mostly measures the platforms’ natural variability and tempts you into reacting to noise. Traffic and rankings are proxies that move for unrelated reasons.

What to do first, if you are doing it yourself

Four steps, in this order. Check robots.txt for blocks on GPTBot, ClaudeBot, PerplexityBot and Google-Extended — occasionally this alone explains everything. Rewrite your homepage opening so it states plainly who you are, what you do and where you operate, in the words a stranger would use. Make your name, address and phone number identical across your website, your Google Business Profile and every directory you appear in. Add Organization or LocalBusiness structured data that matches your visible text. None of this requires an agency, and together it removes most of the ordinary reasons a business gets skipped.

Why assistants hedge instead of naming anyone

You will often get an answer that describes a category rather than a company — "look for a local firm with good reviews and relevant accreditation" — with no names in it at all. That hedge is information. It usually means the system found candidates but could not corroborate any of them well enough to commit, or found conflicting details and chose safety over a specific recommendation. Where a whole sector is being hedged around in your area, the opportunity is larger than in a sector where competitors are already named confidently, because it means nobody local has done the work yet. It is worth testing for deliberately.

The six platforms, briefly

ChatGPT and Claude both answer from a mix of what the model holds and what they retrieve when asked, and both will name businesses when the evidence supports it. Perplexity is retrieval-first and shows its sources, which makes it the easiest place to diagnose why you were or were not named — you can read the pages that persuaded it. Gemini and Google AI Overviews lean heavily on Google’s own understanding of a business, which puts unusual weight on the Google Business Profile. Copilot sits alongside Microsoft’s search index. Because the mix differs, so does the fix, and a report that averages all six into a single score hides the part you needed to know.

Local queries versus national ones

These are different battles and worth separating before spending anything. A local query — a service plus a town — returns a short list, and a well-organised small firm can realistically get onto it, because the competition for those few slots is thinner than people assume. A national query returns the businesses with the deepest corroboration, and displacing them head-on is rarely a sensible plan for a small company. The way in nationally is specificity: the sub-sector, the unusual application, the customer type nobody else names explicitly. Being the obvious answer to a narrow question beats being the fortieth plausible answer to a broad one.

Products versus services

If you sell a service, the question behind the query is usually "who should I use", and the answer turns on identity, location, credibility and reviews. If you sell products, the question is more often "what should I buy", and the answer turns on the product data — accurate specifications, availability, price, structured product markup, and third-party corroboration such as reviews and retailer listings. Both are GEO, but the order of work is different, and an approach that treats a shop like a solicitor will waste months on the wrong signals.

Reviews, and why they carry more weight now

Reviews were always useful for persuading humans. For assistants they do something extra: they are independent, dated, and hard to fabricate at scale without leaving a pattern. Volume matters less than steadiness — a handful arriving every month reads as an active business, while forty in one week reads as a campaign. Replying matters too, because it is public evidence that somebody is home. And distribution matters: reviews across two or three relevant platforms corroborate each other in a way that a single profile cannot.

What to look for in whoever you hire

Ask them to show you the assistants’ raw answers, dated, before and after. Ask what they will not promise. Ask how they measure, and be sceptical of any precise claim about traffic from AI assistants, because that data does not reliably exist yet. Ask whether the work stays yours if you leave. And ask what they would do first for your specific business — if the answer is a package rather than a diagnosis, they have not looked at you. Good GEO work is unglamorous and specific, and it should be explainable to you in plain English.

How long it takes

Corrections to inconsistent information can show up within weeks, particularly on the platforms that retrieve live from the web. Structured data lands as soon as it is crawled. Content restructuring takes as long as it takes to write properly. Corroboration and authority build over months, and anything involving reviews moves at the speed your customers move. Anyone quoting you a fortnight is guessing, and anyone quoting you a guaranteed position is guessing about something that cannot be guaranteed at all.