Generative Engine Optimisation is the work of making a website something an AI system can find, understand, trust and quote. It matters because a growing share of buying decisions now happens inside an answer rather than on a results page, and the mechanics of getting into that answer are not the mechanics of ranking.

Traditional search returns a list and you compete for a position on it. AI search returns a synthesised answer and you compete to be one of the passages it was built from. Those are different jobs. Some of the work overlaps, quite a lot of it does not.

The short version

If you only read one paragraph: AI engines retrieve passages, not pages. They need to reach your site, extract a self-contained answer from it, and find a reason to trust the source. Most websites fail at the second and third of those while passing the first, which is why so many businesses are technically crawlable and practically unquotable.

How bad is it in practice

I benchmarked this properly rather than guessing. In the SearchScore AI Visibility Index for UK Accountants 2026, 1,038 firms were checked across more than 250 signals:

  • The average score was 53 out of 100.
  • One firm in 1,038 reached the AI-Ready band, which starts at 80.
  • Zero firms laid their answers out in the format AI actually quotes from.
  • 82% of firms that had written a questions-and-answers section never added the markup that lets an AI read it.

That last number is the one I keep coming back to. The hard part, working out what clients ask and writing good answers, was done. The thirty seconds of markup that makes it machine-legible was not. I wrote up what that pattern means in zero out of 1,038 accountancy firms write for AI.

What GEO actually consists of

1. Access

Can AI crawlers reach you at all. This is robots.txt, crawler permissions, rendering, and whether your content exists in HTML rather than only after JavaScript runs. This one is not theoretical: across 9,487 UK law firms, the 964 with JavaScript-built sites were 76 times less likely to be visible. Most sites pass this. It is necessary and nowhere near sufficient. See what llms.txt is and what it does not do.

2. Extraction

Can a machine lift a usable answer out of your page. This means a self-contained passage of roughly 100 to 200 words that still makes sense quoted on its own, placed directly under the heading that asks the question. If your answer only makes sense after three paragraphs of preamble, it will not be quoted. Practical version: how to appear in ChatGPT answers.

3. Structure

Can a machine tell what it is looking at. Organisation and Article schema, author markup, FAQ markup where you genuinely have questions and answers. Not for rich results, which Google has largely retired, but to remove ambiguity about who published what. Full detail: schema markup for AI search.

4. Trust

Is there a reason to quote you rather than someone else. A named author with stated credentials, evidence, dates, and corroboration from places that are not your own website. This is where most sites are weakest and it is the least automatable part. See E-E-A-T for AI search.

5. Entity clarity

Does the machine know who you are, consistently, across every surface it can see. One name, one canonical page, one set of profiles that resolve. Contradictory signals are worse than thin ones, because they make you unresolvable rather than merely unknown.

GEO and SEO are not the same thing, and not opposites either

The overlap is roughly half. Crawlability, site speed, sensible information architecture and genuine subject depth serve both. What differs is the target: SEO optimises for a ranked position, GEO optimises for being the passage that gets used.

Google’s own guidance says optimising for AI search is essentially just SEO. That is accurate within Google’s scope and misleading outside it, which I have argued at length in reading Google’s AI search guidance properly. The short form is in GEO vs SEO.

Why nobody notices they are losing

When you slip down Google, something reports it. Traffic dips, a dashboard turns red, somebody asks why.

When an assistant recommends a competitor instead of you, nothing anywhere records that it happened. There is no line item for the client who never contacted you because your name never came up. The loss is real and the signal is absent, which is exactly why entire professions can sit at an average of 53 without anyone raising it.

Related reading: the agent already chose your competitor and AI content blindness is real.

Where to start

  1. Check you are not blocking the crawlers. Ten minutes. Usually fine, occasionally catastrophic.
  2. Tag the FAQ you have already written. If you are in the 82%, this is the highest-return half hour available.
  3. Add Organisation and Article schema. On the accountancy benchmark this was worth close to 20 points, the biggest single controllable gain.
  4. Put a named author with credentials on the pages that matter.
  5. Rewrite your key pages answer-first. One self-contained passage under each heading.

If you would rather see where you stand before doing any of it, the audit is free at SearchScore, which is the tool I built for exactly this. Disclosure noted so you can weigh it accordingly.

Everything on this site about AI search