field note
first published 2026-10-09
Generative engine optimisation (GEO): what it is, what it isn't, and what to actually do
GEO is the work of becoming the source an AI answer engine cites. Most of it is engineering, not copywriting. A plain definition, the four layers that decide citation, the things sold as GEO that aren't, and a sequence you can run this month.
1,006 words · 5 min read · ai search
Generative engine optimisation is the work of becoming the source an AI answer engine cites. That is the whole definition. Everything else people attach to the acronym is either ordinary SEO wearing a new badge or a dashboard with a subscription.
I have been a technical SEO since 2007, and I run the instruments for this on my own site before I run them for anyone else. Here is the discipline as I practise it: what decides citation, what is sold as GEO but isn't, and what to do first.
Four layers decide whether you get cited
Access. The engine's crawler has to be allowed in. OpenAI uses OAI-SearchBot for search and GPTBot for training; Anthropic has ClaudeBot and Claude-SearchBot; Perplexity has PerplexityBot; Google has Google-Extended for its AI products on top of Googlebot. Each of those is a decision, and most sites made none of them: they inherited whatever the security product or the robots file did by default. The commonest expensive mistake is blocking the search bots while meaning to block the training ones.
The guestbook on this site's lab page counts every one of those crawlers by product. On a typical day they are all here, and together the AI readers now rival Googlebot in volume. They are not a future audience.
Extractability. The crawler has to find the answer in the HTML it receives. Most of the engines do not render JavaScript the way Google eventually does. A page whose body copy arrives via a client-side framework is, to them, a title and a navigation menu. This is a server-rendering decision, and it is the item I fix most often.
Entity clarity. The engine has to know who is speaking and whether to trust them. That is a structured-data graph that names the Organization or Person, links the pages to it with @id, and agrees with the visible copy. It is also the profiles elsewhere that say the same thing, linked from the graph with sameAs. Plugins that emit a FAQPage here and an Article there do not build a graph; they scatter islands.
Consistency. Engines increasingly read more than one page before they quote one. A site whose about page, footer and structured data disagree about a founding year, a location or a service list gives the model a reason to prefer a source that doesn't contradict itself. I wrote at length about this in the piece on sites being deposed; the short version is that the page is now a witness, and witnesses are cross-examined.
What is sold as GEO but isn't
Prompt tracking. Dashboards that run a list of prompts against several engines and report whether you were mentioned. The prompt is not a query; nobody types the same one twice, the answers vary run to run, and the metric cannot be acted on. I have written about why the prompt is the wrong string to optimise. Buy it if you need a chart for a meeting; do not mistake it for measurement.
"Writing for AI." A special register, extra FAQ blocks, paragraphs that open "In the context of…". The engines reward clear, direct prose that answers the question in the first sentence, which is what human readers reward. There is no second style guide.
Schema for its own sake. Marking up things that are not on the page, or stacking every type a plugin offers. The validators will pass it and the engines will ignore it or hold it against you. One connected graph that describes the real entities beats forty blocks.
Guaranteed citations. Nobody can guarantee what a model will quote. What can be guaranteed is the set of inputs: access, extractability, a valid graph, consistent facts, measured over time. If someone guarantees the output, ask them to show you the instrument.
What to do, in order
- Decide the crawlers. Write a one-page policy: which AI bots are allowed, which are blocked, and why. Implement it in
robots.txtand in the firewall, and test each user agent against the origin. Record the decision. - Fetch your ten most important pages without JavaScript. If the answer is not in the response, fix the rendering. Nothing else on this list matters until it is.
- Rebuild the structured data as a graph. Organization or Person at the root, pages linked by
@id, profiles linked bysameAs, zero validator errors. Three small businesses on three platforms did exactly this and went to zero errors; the platform did not matter. - Make the facts agree. One founding year, one address, one service list, everywhere the site states them.
- Publish the machine surfaces. An
llms.txt, clean feeds, and markdown mirrors of the pages that matter, advertised with aLink rel="alternate"header. Cheap, and it makes you easy to read. - Instrument the inputs. Bot hits by product from your logs, citation referrers where the engines send them, structured-data validity on a schedule, and a grade over time. The Agent-Ready Grader on this site is free and ungated and checks the three gates that decide whether an engine can read you at all.
The honest state of measurement
You will be able to prove the inputs moved long before you can prove the citations did. The engines do not report to you; Google folds AI Overview clicks into normal Search Console data; the others send referrers inconsistently. That is uncomfortable and it is the situation. The response is not to buy a dashboard that pretends otherwise. It is to measure what you control, keep a manual log of the queries that matter, and read your server logs like they are the only truth you have, because for this they are.
The definition, again
Generative engine optimisation is becoming the source the answer engines cite. It is mostly engineering: access, extractability, entity clarity, consistency. It is measured by instruments you build. And it is done on the same HTML Google reads, which is why a site that is genuinely legible to machines tends to do well on every engine at once.
asked straight — answered straight
01What does GEO stand for?
Generative engine optimisation: making a website the source that ChatGPT, Perplexity, Claude, Gemini and Google's AI Overviews draw on when they write an answer. You will also see AEO (answer engine optimisation) and AI SEO used for the same work.
02Is GEO replacing SEO?
No. The answer engines fetch and read the same HTML Google does, and most of them lean on conventional search indexes to decide what to fetch. GEO is SEO with the weighting moved toward access, extractability and entity clarity, and with a weaker measurement layer that you have to build yourself.
03How long does GEO take to show results?
The technical inputs change in days: crawler access, server-rendered answers, a valid graph, an llms.txt. Citations follow when the engines re-fetch, which for active sites is frequent. The honest answer is that you will see the inputs move before you can prove the citations did, which is why you instrument the inputs.
04Do AI engines use llms.txt?
There is no confirmation from any major engine that llms.txt affects citation, and I would not buy a service that claims otherwise. It costs ten minutes, it helps agents and crawlers find your machine-readable surfaces, and this site's grader awards a small frontier bonus for it. Treat it as hygiene, not strategy.
next step — ai search optimisation
AI search optimisation as engineering.
Make your site legible to ChatGPT, Perplexity and AI Overviews — then measure it. Send a brief — a few lines about the site, the problem and the evidence you have — and the reply is a straight read on fit and shape, from the person who does the work.
no forms · no funnels · jamie@jamiemckaye.com
written by
Technical SEO consultant and full-stack developer in Hersham, Surrey, in practice since 2007. One person, no handoffs: the audits, the code and the writing come from the same pair of hands. This site is the working proof — it grades itself on the same instrument it runs for clients.