Generative engine optimisation (GEO) is work on how AI assistants such as ChatGPT, Perplexity, Gemini, Claude and Microsoft Copilot find, describe and recommend a business. When a buyer asks one of them which accountants in Bristol understand landlords, or whether a particular software tool is worth the money, the assistant names a handful of options and often links the pages it used. GEO is about whether you are one of those options, and whether what the assistant says about you is true.
This page covers the assistant side of AI search optimisation. If your concern is the AI summary at the top of Google’s results, Google AI Overviews optimisation fits better. If it is about shaping pages so one direct answer can be quoted, look at answer engine optimisation.
What GEO changes for a business
An assistant that answers from the live web works in two steps. It runs one or more searches for the question, then writes its answer from the pages that came back. That process is called retrieval-augmented generation, and it means an assistant can only mention you if three things hold: its crawler or search provider can reach your pages, the pages it retrieves say plainly what you do and for whom, and enough independent sources agree with them.
GEO works on all three. When your buyers ask an assistant the questions they ask before contacting a supplier, your business should appear for the needs you genuinely serve and be described correctly: the right services, the right area, a realistic idea of price and the kind of client you suit. Being left out is one problem. Being described wrongly, with a closed office or a service you stopped offering years ago, is often the more urgent one, because it puts off people who were already looking for you.
Who GEO suits, and who it does not
It suits UK businesses whose buyers compare options before getting in touch and are likely to ask an assistant somewhere along the way:
- software and B2B suppliers, where buyers ask an assistant to draw up or compare a shortlist;
- accountants, solicitors, consultancies and other professional practices, where people ask who specialises in their exact situation;
- private clinics and specialist services, where people ask what a treatment or service costs and who offers it;
- online shops in categories where people ask for the best option under a set budget.
I would not start here for a business that Google barely indexes, or whose website does not yet say clearly what it sells. Assistants retrieve ordinary web pages, so technical SEO fixes come first and make every later step cheaper. A trade or salon whose work comes from people nearby is usually better served by local search and its Google Business Profile. Assistants are asked for local recommendations too, but as far as I can tell those answers lean on the same maps listings, directories and reviews.
What has evidence, and what does not
GEO is young, and much of what is sold under the name is guesswork presented as method. This is how I sort the claims.
Documented by the platforms themselves
- Assistants that search need access. OpenAI’s documentation of its crawlers says sites that block OAI-SearchBot will not be shown in ChatGPT search answers, apart from plain navigational links. Perplexity documents PerplexityBot as the crawler that surfaces sites in its answers, and Anthropic lists separate crawlers for Claude’s training and for its search.
- Training and answering are separate switches. OpenAI’s GPTBot collects training data, and blocking it does not block ChatGPT’s search crawler. Google describes its Google-Extended control as covering Gemini training and grounding in the Gemini app, so a site that blocks it to opt out of training may also be stepping out of Gemini’s live answers.
- Copilot answers from Bing. A site that Bing indexes poorly starts at a disadvantage in Microsoft’s assistant, whatever its Google rankings.
Supported by research, with caveats
The term itself comes from the research paper that introduced GEO, by Aggarwal and colleagues, first published on arXiv in November 2023. On a test engine the authors built, adding quotations, statistics and cited sources to a page raised its visibility in generated answers by up to 40%. That supports writing pages full of checkable evidence, but it was measured on a lab system, not on ChatGPT or Perplexity as they run today.
Unproven, or contradicted
- That one specific change will make a named assistant start recommending you. Nobody outside these companies, me included, can show a controlled test of that on the live products.
- That an llms.txt file gets you cited. No major assistant has confirmed relying on one.
- That a single “AI visibility score” is precise. Answers change with the wording, the user’s location and history, and the day, so any score is a sample. Ask how it was sampled.
What the GEO work covers
Access for assistant crawlers
I check robots.txt line by line against each assistant’s documented crawlers, then check the firewall or CDN in front of the site, because bot-protection settings can block AI crawlers without anyone choosing to. I also look at the raw HTML of key pages: several AI crawlers do not appear to run JavaScript, so content that only loads in the browser may be invisible to them. And I check how Bing indexes the site in Bing Webmaster Tools, which many businesses have never set up.
One clear, consistent description of the business
Assistants have to work out that the business on your website, your Companies House record, your Google Business Profile, your LinkedIn page and a trade directory are the same entity. I agree a short fact sheet with you (trading name, what you do, who for, where, how you price) and make every source say the same thing. On your own site that includes structured data naming the organisation and linking its official profiles.
Pages for the questions buyers actually ask
Recommendation and comparison questions need pages that answer them: who you serve and who you do not, where you work, how pricing works with real ranges in pounds where you can publish them, and an honest account of how you differ from the alternatives. A page that only says “bespoke solutions, get in touch for a quote” gives an assistant nothing to repeat. I rewrite or brief these pages so the facts are specific and supported.
The independent sources assistants draw on
When an assistant names suppliers, it usually cites comparison articles, review platforms, trade press, professional body directories and forum threads alongside company websites. I map which of those sources appear for your questions and then work on being accurately present in them, through listings, corrections, contributions to trade publications and the kind of link building and digital PR that earns genuine coverage.
Correcting wrong answers
Some errors are a hallucination, where the model fills a gap with something plausible. More often, in my reading, the assistant is faithfully repeating an out-of-date source: an old directory entry, a news story about a previous address, a former partner’s profile. I trace the wrong statement back to where it is likely coming from and fix it there, because assistants offer businesses no way to edit their answers directly.
How I run a GEO project
- Write the prompt set with you. Questions in your buyers’ own words at each stage: describing the problem, asking for options, comparing two suppliers, and checking you by name.
- Sample each assistant. I run every prompt in ChatGPT, Perplexity, Gemini, Claude and Copilot, in fresh sessions and more than once, and record the date, whether you appear, how you are described and which pages are cited.
- Map the sources. The cited pages show which sites carry weight for your questions and where competitors appear and you do not. That decides where the effort goes.
- Fix access and facts first. Crawler access, Bing indexing, the fact sheet and the structured data, since everything later depends on them.
- Improve pages and earn coverage. Page rewrites and new pages for unanswered questions, then work on the third-party sources.
- Re-sample on a fixed schedule. The same prompts, the same way, so each round compares like with like, reported alongside GA4 visits referred by the assistants themselves.
Common problems on UK sites
- An About page with no named people, so nothing connects the business to the expertise behind it.
- A trading name that differs from the Companies House name, with nothing on the site connecting the two.
- Service pages built around “contact us for a quote”, with no price guidance, no named client type and no location.
- Directory and review listings still showing a previous address, phone number or business name.
- No independent coverage at all, so the only source describing the business is the business itself.
What you receive
- The agreed prompt set and a dated baseline for each assistant, in a spreadsheet you keep.
- A source map showing which sites are cited for your questions.
- A crawler access report with the exact robots.txt and firewall changes needed.
- The business fact sheet, ready to copy into profiles and listings.
- A prioritised action plan covering pages, structured data, listings to correct and coverage to pursue.
- Re-sample reports on the agreed schedule, with GA4 referral figures from the assistants.
Scope depends on your site and how many questions and assistants matter to you, so I agree it in writing before any work starts. My guide to how AI answers choose their sources is a useful read beforehand if you want the background.
Next step
Send me your website and three or four questions your buyers might ask ChatGPT or Perplexity before choosing a supplier like you. Book a consultation through my contact page and our first conversation can start from what the assistants say about your business today.