What Is Generative Engine Optimisation (GEO) and AI Search About?

Search still works the way it always has. People ask a question online, and “something” goes and finds the answer. What’s changed is that a growing chunk of that “something” is now ChatGPT, Gemini or Claude, not a results page.

Generative Engine Optimisation (GEO) is what it takes to actually show up when that happens. Plenty of explainers cover the basics, but having spent the last two years genuinely deep in this, and having built Ebb, our own LLM measurement tool along the way, we wanted to go a level further, digging into the actual mechanics of how an AI decides what to cite, using real examples from our own client work.

GEO isn’t a replacement for SEO. It’s SEO plus two extra jobs

Here’s the bit I think gets missed the most, which is that GEO doesn’t start from a blank page. When we rebuilt PureGym’s content strategy around AI Overviews, the starting insight was simple – that the content Google chooses to feature in AI answers is overwhelmingly content that already ranks well organically. AI visibility isn’t some separate discipline bolted onto SEO. It’s the same fundamentals, just sharpened, with two extra jobs stacked on top – winning the fan-out queries behind a prompt and actually getting selected once you rank for them.

What that thinking did for PureGym means that their blog now appears in over 7,000 Google AI Overviews, roughly 10 times more than their closest competitor.

How an AI actually decides what to search for fan-out queries

This is the part most explainers skip entirely but something I’ve been curious to really get under the skin of. When someone prompts an LLM, it doesn’t just search that exact phrase. It breaks the prompt down into several sub-queries known as fan-out queries and runs those behind the scenes before it ever generates an answer.

Some of these read exactly like something a human would type. Others really don’t. Ask an LLM “what are good gyms in Cardiff?” and it might fire off fan-out queries like “good gyms in Cardiff UK,” “best fitness centres gyms Cardiff reviews,” and “best gyms Cardiff.” Two of those sound normal. One sounds like nobody, ever, has typed that into Google. All three still shape what you get back.

Ranking for the phrase someone might actually ask isn’t really the target. Ranking for the fan-out queries behind it is, which is a much more interesting problem.

Ranking isn’t enough. You also have to get selected

This is the bit I think most GEO content seems to consistently miss. When an LLM does a web search, it isn’t reading full pages. It gets snippets, a title and a meta description, and then decides which of those are actually worth pulling into its answer.

Which means, infuriatingly, a page can rank page one for a fan-out query and still not get cited, if the snippet doesn’t make the relevance obvious enough. We’ve seen this happen to genuinely well-ranking pages from big brands, they show up in the search results happening behind the scenes but never make it into the final answer, because the title and description didn’t connect the dots clearly enough.

A generic, brand-first title like “Skyscanner | Compare Cheap Flights” might rank perfectly fine and still underperform on selection next to something more intent-explicit, like “Find Cheap Flights to Paris | Compare 100+ Airlines.” Clever or vague copy might charm a human. It tends to just confuse a retrieval system trying to match a specific need.

So if you’re ranking for the right fan-out queries but not showing up in the actual AI answer, this is very likely why.

Two routes to citation, and they need completely different tactics

There’s a second distinction worth knowing about, and it’s one I keep coming back to, which is whether an AI is answering from its knowledge bank (what it learned in training) or from a live web search it’s just run. Both matter. Interestingly, they don’t respond to the same tactics.

Remitly’s Immigration Index is a knowledge bank example. What mattered here was consistency, not volume. “Immigration Index” got name-checked in 55% of nearly 300 pieces of coverage across 64 countries. That repetition, the same term surfacing again and again across independent sources, seems to be exactly what builds enough confidence for an LLM to recall it directly. The campaign page now holds the number one AI Overview slot for high-intent terms like “immigration index” and “best countries for immigrants.”

PureGym leans on the other route entirely, web search retrieval. Freshness and structure did the heavy lifting including direct-answer formatting, comparison tables, question-led titles, all the stuff that plays well when an AI is retrieving in real time rather than recalling from memory.

Working out which route actually matters most for your priority prompts is a genuine strategic call. Chasing knowledge bank influence means prioritising consistent naming and Digital PR coverage that ends up feeding training data. Chasing web search visibility means prioritising the structure and freshness that wins in real-time retrieval. Pick wrong and you’re optimising for the wrong problem entirely.

What actually earns a citation once you’re in the running

Once something’s ranking and reasonably well structured, a few things still separate what gets cited from what gets skipped.

Clear signals of expertise and trust seem to matter more here than in traditional SEO, even with E-E-A-T being a major ranking factor. Concrete numbers, named experts, visible proof points, award wins, verified results, credentials, give an LLM something to actually hang its confidence on, rather than just judging relevance.

Explainer content also tends to win out over sales content and I don’t think that’s a coincidence. LLMs are trying to answer a question, not respond to a pitch, so content that genuinely walks someone through a concept has a much better shot at being the thing an AI reaches for than content that’s obviously trying to sell them something.

Seeing whether any of this is actually working

All of the above is only useful if you can see it happening, and this is where it gets genuinely hard. Citation rates, fan-out query rankings, whether you’re being pulled from knowledge bank or live search…all of it shifts week to week and none of it is visible unless you’re deliberately tracking it.

It’s a big part of why we built Ebb, tracking exactly this across ChatGPT, Gemini and Claude, benchmarked against named competitors, so everything above turns into something you can actually measure, rather than something we all just nod along to at conferences.

If you’re figuring out how to show up in AI answers, get in touch, or check out more of our Insights below. You can also read more about our approach to AI here.