Generative engine optimisation explained

Generative engine optimisation is the label that has stuck for the work of getting your content surfaced by large language model search. ChatGPT search, Google AI Mode, Perplexity, Claude with web access. The acronym is GEO. You will also see AEO (answer engine optimisation) and LLMO (LLM optimisation). They all describe the same shift, which is that a growing share of decisions now happen inside an AI answer instead of on a results page.

The honest read in 2026 is that GEO is real, it is early and most of what is sold as a GEO service is recycled SEO with a new front page. There are useful things to do. There are also tools and promises that should make you suspicious. Here is how I think about it for clients.

From clicks to citations

Traditional SEO optimises for a click. A user types a query, Google returns ten links, you fight for one of those slots and a fraction of users come through. GEO optimises for a citation. A user asks a question, an AI assembles an answer, and your page is either named as a source or absorbed silently into the answer with no attribution.

That is the structural change. Even when you do get cited, the click-through is lower because the user already has their answer. The benchmark studies I trust suggest AI answers cite three to four sources from roughly ten pages evaluated per query. That is a different funnel from blue links.

There is a softer version of the same effect that is easy to miss. Even when the user does not click your citation, your brand has been named in the answer. For some buying journeys that is the entire job. A buyer doing research at the awareness stage may never visit your site, but they walk away from the AI answer with a shortlist that has your name on it. In other journeys, the citation is more like a credit line at the bottom of a Wikipedia article. Useful for trust signals, not for traffic.

How LLM search actually works

Two mechanisms decide whether your site shows up in an AI answer. The first is training. Whatever was in the training corpus when the model was built influences which entities and facts the model knows about. You cannot game training. You can only be the kind of site that consistently gets included.

The second mechanism is retrieval. When the user asks something current or specific, the system runs a web search, fetches a handful of pages, summarises them, then cites. This is where GEO work pays off, because the retrieval step is closer to a traditional search engine. Pages still need to be indexable, the content still needs to answer the question and the source still needs to look authoritative.

Different engines weight different things. Perplexity leans on journalism and earned media. ChatGPT search pulls from directories and third-party listings more than people expect. Google AI Mode favours content the brand controls. You cannot optimise for one without watching the others, but you can stop trying to game any single one.

What survives from SEO

Most fundamentals carry over. The work I have been doing for SEO clients for years is the same work that helps GEO.

Domain authority still matters. Models and retrieval systems both lean on signals that approximate trust, so older domains with real backlink profiles get cited more often. Structured data still matters. Schema.org markup gives both classical search and AI retrieval a clean way to extract entities, dates, prices, authors and ratings. Freshness still matters. AI engines weight recency. A 2023 page on a fast-moving topic loses to a 2026 page that has been updated.

On-page clarity also matters more than people give it credit for. Direct answers near the top of a page. Clean headings. Short paragraphs that contain a single claim. AI systems extract these aggressively because they are the easiest passages to lift cleanly into an answer.

One thing has shifted in how on-page clarity gets rewarded. Classical SEO tolerated long preambles. You could open a recipe page with 500 words of childhood memory before the ingredients showed up, and Google would still rank it. AI retrieval has no patience for that. The system grabs the cleanest passage near the top and moves on. If the cleanest passage is filler, your page is invisible. If it is the actual answer to the question, you are in.

What is genuinely new

A few things are new, and they are the parts of GEO worth paying attention to.

Entity coherence. AI systems are better than classical search at understanding what a page is about as a whole. If your homepage, services pages, blog and structured data all consistently describe what you do and who you serve, the model has a much higher chance of associating your domain with that entity. Inconsistency hurts you in a way it never did on Google.

Citation-worthiness. AI answers prefer sources that contain specific, verifiable claims. Numbers, dates, named cases, original research, first-hand experience. Bland marketing copy with no specifics never gets cited because there is nothing to lift cleanly. If your service page reads like every other service page, you are invisible to the citation layer.

Distinctive viewpoint. The systems are tuned to prefer sources with a stance and a perspective rather than aggregated summaries. A page that takes a position, then defends it with specifics, performs better than one that lists every possible angle. This is closer to how a journalist would pick a source than how a search engine ranks one.

Machine-readable signals. There is a slow convergence on making content easier for machines to consume cleanly. Schema.org continues to expand, OpenGraph still matters and a small file called llms.txt has been proposed as a way to give AI agents a curated map of your site. I will come back to that one.

Author identity is part of this too. Sites that name their authors clearly, link them to credible profiles and use Person schema get cited more often than anonymous content. The Princeton research that has been doing the rounds suggests citation counts, author bylines and inline statistics together can lift AI visibility by 30 to 40 percent against an unoptimised baseline. The numbers are rough, but the direction is consistent across every study I have read.

The llms.txt question

The llms.txt proposal is a markdown file at the root of your domain that summarises your site for AI agents. The pitch is that it gives a clean, machine-readable surface that an AI can route on without crawling your full HTML. Adoption has grown slowly. The most recent figure I have seen put it on roughly 10% of crawled domains.

Major AI crawlers, as far as anyone can verify in 2026, mostly ignore llms.txt and crawl your HTML directly. There is no enforcement, no formal standard at W3C and no public statement from OpenAI, Google or Anthropic that they use it. Where it does seem to help is with IDE-based agents, MCP servers pointed at documentation and machine-to-machine workflows. So treat it as a small, low-cost bet rather than a search visibility lever. I add one to client sites that have a real documentation surface. I do not add one to a five-page brochure site, because there is nothing for it to point at.

On GEO overlay tools and ranking services

A wave of products has appeared that promise to optimise your site for AI search. Some are useful diagnostic tools that scan your content for clarity, structured data and entity coherence. A handful go further and promise to rank you on ChatGPT, Perplexity or Google AI Mode within a fixed window. Treat that second category with the same scepticism you would treat any other guaranteed-ranking pitch.

Nobody controls which pages an LLM cites. The retrieval models change, the source weighting changes and the underlying language model gets swapped out every few months. A pitch that promises a position on ChatGPT is the same pitch that promised top three on Google twenty years ago, with the same likely outcome.

The diagnostic tools have a place. If a scan tells you your service pages have weak entity coherence or no schema, that is a real finding. Act on the finding. Do not pay anyone for a "ranking guarantee" in AI search.

The same is true of the broader category of "AI SEO tools". A scanner that lints your content against good GEO practice is fine. A dashboard that shows you which AI engines mention your brand and what they say about it is genuinely useful as a measurement layer, in the same way Google Search Console is useful. The line to watch is anywhere the marketing promises a rank, a position or a guarantee.

What a UK business should actually do

If you are running a UK business and you want to be cited more often in AI search, the practical work is short.

Audit your core pages for direct answers in the first 60 words. If a user asks the question your page is meant to answer, can an AI lift a clean two-sentence response from the top of the page? If not, rewrite the opening.

Add structured data. Organisation, Service, FAQPage, Article, Person schema for authors. Schema.org is mature and well-supported by retrieval systems. The payoff is large for the effort involved and it is not new work.

Make your entity coherent. Pick the two or three things you want to be known for, then make sure your homepage, services pages, About and case studies all describe those things in compatible language. Drop the marketing speak that fits any business in any sector.

Publish things worth citing. First-hand experience, named clients with permission, real numbers, opinions you can defend. Bland thought leadership rounds out a content calendar but it almost never gets cited.

Update older pages. AI engines weight freshness. A 2024 guide that is still accurate becomes more cite-worthy the moment you revisit it and stamp it with a 2026 review date.

Do not chase llms.txt unless you have a real documentation surface that benefits from a machine-readable map. Do not pay for AI ranking guarantees.

The longer view

GEO will look different in two years. The specific engines will change. Citation patterns will change. The underlying principle is unlikely to. Authoritative sources, clear writing, specific claims, coherent entities. That is what classical search rewarded too, just less strictly.

The businesses that win in AI search will be the ones that were already doing the boring fundamentals well. They will publish things worth citing, mark them up cleanly and trust that the retrieval systems get better at finding them over time. Everything else is noise.

If you want a second pair of eyes on your site from a GEO angle, get in touch. I also write about how customers find products with AI and what web MCP means for your website if you want the wider picture. For implementation, see AI integration services.