What is Generative Engine Optimisation?

Key takeaways

• GEO is the practice of making a brand retrievable, parseable and verifiable enough that AI answer engines cite it. The unit of success is a mention inside an answer, not a position in a list.
• Four factors raise the odds that an engine names a brand: the page is reachable, parseable in isolation, corroborated elsewhere, and attributable to a recognised entity. None guarantees a citation.
• Check the technical layer first, because it is the quickest to fix: a blocked crawler or client-side-only body copy keeps a site out of an engine however good the writing is. Relevance, expertise and outside recognition still decide the rest.
• GEO is measured with AI visibility percentage, share of voice, citation count, sentiment, and AI referral traffic isolated in analytics. All five need a held-constant prompt set to mean anything.
• GEO does not replace SEO. The technical foundations are shared, and organic still sends most qualified traffic for most brands.

Someone researching your category today may never see a list of ten links. They ask a question, read one synthesised paragraph, and act on it. Whether your company appears in that paragraph is now a distinct discipline with its own inputs, its own failure modes, and its own numbers. This guide covers what it is and how it is done.

What is Generative Engine Optimisation?

Generative Engine Optimisation, usually shortened to GEO, is the practice of making a brand retrievable, parseable and verifiable enough that AI answer engines cite it by name. Those engines include ChatGPT, Perplexity, Google Gemini, Claude, Microsoft Copilot, and Google's AI Overviews and AI Mode. Each of them answers a question directly rather than returning a list of documents, and each decides, per answer, which sources to draw on and which brands to name.

The discipline is sometimes called AEO, answer engine optimisation, or AI search optimisation. The labels differ; the object is the same. You are trying to be the thing the machine quotes.

What makes it a separate practice rather than a subset of SEO is the mechanism. A ranking system orders documents and lets a person choose. A generative system retrieves passages, synthesises them into new prose, and attributes selectively. Optimising for the first is about relevance and authority at the level of the page. Optimising for the second is about extractability and corroboration at the level of the passage and the entity.

How is GEO different from SEO?

SEO optimises a page to rank in a list of results a person then chooses from. GEO optimises a page to be extracted, quoted and attributed inside a single answer the person never leaves. The technical hygiene overlaps almost completely, which is why the two should never be run as separate programmes. Google says as much: from its point of view, optimising for generative AI search is still SEO. What differs is what counts as a win, and where the work happens.

The practical consequence: a site with weak organic foundations will not do well at GEO either, because the same crawl and structure problems limit both. Fix the foundations once and both benefit.

Why do AI engines cite some brands and not others?

No engine publishes its rules, but four factors consistently raise the odds that a brand gets named, and they reinforce each other. Most GEO advice addresses one of them in isolation, which is why so much of it produces no measurable change.

1. Reachable. The engine's crawler is permitted to fetch the page, and the substance of the page exists in the raw HTML rather than appearing only after JavaScript executes.
2. Parseable. The relevant passage answers one question and holds up when read on its own, separated from the rest of the page. Retrieval systems often work on fragments rather than whole documents.
3. Corroborated. The claim, or something close to it, appears somewhere the engine already treats as reliable: a review platform, a trade publication, a well-regarded community thread.
4. Attributable. Structured data and consistent entity information make it unambiguous which organisation the passage belongs to. An engine that cannot resolve the entity confidently will describe the category instead of naming a company.

Read as a sequence, these are also a diagnostic order. If a brand is invisible in AI answers, start with the first, because it is the quickest to rule out, then work down. A site can pass all four and still lose to a source the engine finds more relevant.

How do you find out what AI engines say about you today?

You ask them, systematically, and you write down what comes back. There is no dashboard that substitutes for this, because the answer to a given question varies by engine, by phrasing, and over time. The method matters more than the tooling.

Build a prompt set and hold it constant

Split the set between branded queries, the ones where a person already knows your name, and category queries, the ones where they do not. Branded queries tell you whether the engine understands who you are and whether it has anything wrong. Category queries tell you whether you exist at all in the consideration set.

• Branded: what is [brand], what does [brand] do and for whom, is [brand] recommended for [category], what do people say about [brand], who competes with [brand].
• Category: best [service] for [segment], who offers [specific capability], [option A] versus [option B], how do I choose a [service] provider.

Run each in a clean session with live retrieval enabled, so personalisation and chat history do not contaminate the result. Then record five things per answer: whether the brand appears, where in the answer it sits, the sentiment attached to it, which competitors are named, and every source the engine cited.

Hold the set constant month to month, and run the key prompts more than once: answers vary between runs, and engines change their behaviour over time. Treat the result as an indicator on a defined sample, not a census of every answer about you. The value is in the trend, and a set that keeps changing produces a number that cannot be compared to itself.

One caveat worth knowing. Querying an engine through its API with web search enabled gives an approximation of what a person sees in the consumer app, not a replica: models, settings and personalisation can differ. Monitoring tools work this way. It is useful for reading direction and comparing periods. It is not a guarantee of what any individual user saw.

Trace what the engine gets wrong

Inaccurate descriptions, discontinued services and wrong team names are common. Some are fixable, but not by arguing with the output. Search the exact phrasing the engine used, look for the page it came from, check when that page was last updated, and pursue the correction there. Often there is one; sometimes the engine blends several sources or produces an error with no single origin, and then the remedy is clearer, more consistent information in the places it already reads.

What stops a site from being cited?

The technical layer is not the only reason a site goes uncited, but it is the one to rule out first: it is quick to check, and a failure there cancels everything else. These are the checks worth running monthly, in roughly the order that they matter.

The robots.txt case deserves emphasis because it is both the most damaging and the most common. Blocking rules are often added years earlier by someone managing bandwidth or acting on a general policy about AI crawlers, and nobody revisits the decision once the business wants AI visibility. It is one line in one file, and it is worth checking before anything else.

How should a page be structured so an engine can quote it?

None of this is required: Google states that its AI features need no special optimisation beyond good SEO. These are practices that tend to help, because retrieval systems often work on passages rather than whole pages, and they make a page clearer for people too. Write in self-contained sections that answer one question each, and put the answer in the first sentence.

• Question-shaped headings. A heading that matches the phrasing a person would use gives the retrieval system something to match against. Stylish headings read better and retrieve worse. Both have their place, and the balance depends on whether a page exists to convert or to be cited.
• Answer first, context after. Open each section with a direct answer, then explain. Our working rule of thumb is a few sentences, roughly 30 to 80 words; Google says there is no ideal length.
• Sections that stay focused. Long enough to be substantive, short enough to be read whole. We aim for roughly 100 to 300 words, as a habit rather than a rule.
• A takeaways block under the H1. Frequently lifted more or less verbatim into synthesised answers, so it deserves more care than a summary usually gets.
• Real HTML tables. Structured comparisons are exactly what engines like to quote, and exactly what an image destroys.
• A named author with credentials. A real byline linked to a real author page makes the expertise visible to readers and engines, and anonymous content gives neither anything to check.
• Honest dates. Update the modified date when the content genuinely changes. Touching the date without touching the content buys nothing and erodes trust.

A page that answers one question well tends to beat a page that covers a topic comprehensively but answers nothing directly.

What is llms.txt, and do you need one?

llms.txt is a plain-text file at the root of a domain that lists the pages a site considers most worth reading, each with a one-line description. It sits at yoursite.com/llms.txt and is written in simple Markdown. It is a curated index, not a sitemap: the point is selection, so it should stay short and omit tag pages, thin content and legal boilerplate.

Adoption across engines is uneven, and Google has said it does not use llms.txt for its AI features, so treat it as optional: it never replaces crawlable, server-rendered pages and useful content. The case for having one anyway is that it takes an afternoon to build, a few minutes a month to maintain, and it forces a genuinely useful exercise: deciding which twenty pages on your site actually represent you. Order the list by service pages, then pillar guides, then case studies, then team and about pages.

How do you earn citations off your own site?

Start from the sources the engines already quote for your category, rather than from a generic outreach list. Monitoring gives you this for free: every answer names its sources, and across thirty answers a pattern emerges. Usually a review platform, one or two trade publications, LinkedIn long-form, and a community thread or two.

Those specific domains are the target list, because they are already inside the retrieval set. A placement on a publication an engine has never cited for your category is worth considerably less than one on a publication it cites every time.

• Editorial guest placements on publications already appearing in answers for your target queries. The author bio matters as much as the article: full name, role, company, which is what feeds the co-citation signal.
• Review platforms, approached with a structured request. Asking three specific questions about the problem, the alternative considered, and the measurable outcome produces reviews with substance. Reviews containing concrete outcomes are far more citable than generic praise.
• Community answers where the question genuinely matches your expertise. Substantive, complete, and useful whether or not anyone clicks through. Promotional answers are obvious to readers and to engines.
• Long-form on LinkedIn, published from a personal profile rather than a company page. In our experience, personal authority counts for a lot in professional and B2B queries.
• Entity consistency across all of the above. The same company name, description and founding details everywhere. Contradictions between your site, your LinkedIn page and a directory listing give an engine a reason to describe you vaguely.

Unlinked mentions still count. Co-citation, the pattern of your brand appearing alongside a topic across independent sources, is what builds the association, and a dofollow link is a bonus rather than the objective.

How is GEO measured?

Five numbers, all of which depend on running the same prompt set every month. Without a held-constant set, none of them can be compared to their own previous value, which makes them decoration.

• AI visibility. The share of monitored prompts where the brand is named, expressed as a percentage. The headline number.
• Share of voice. Across category prompts, your mentions as a proportion of all brand mentions including competitors. This is the one that reveals whether you are gaining or merely present.
• Citations. The count of distinct pages from your domain that engines used as sources, and which pages are new this month.
• Sentiment. The positive, neutral and negative split across all mentions, with any negative mention traced back to its source.
• AI referral traffic. Sessions, engagement and conversions from answer engines, isolated in analytics.

That last one needs a check. GA4 now groups known AI assistants into an AI Assistant channel by default; other analytics platforms may not, and newer assistants can still land in Referral or Direct. Completing that list takes an hour and often changes the conversation, because the traffic was already arriving. For Google's own surfaces, the generative AI performance report in Search Console shows how AI Overviews and AI Mode use your pages. It is part of the measurement infrastructure we set up anyway.

How long does GEO take?

Technical fixes can change what an engine is able to read within days. Citation gains take a quarter to read reliably, and compound over a year. The phases are distinct enough to plan around.

Month one is measurement and removal of blockers: establish the baseline, fix crawler access, validate schema, set up the analytics channel. It is unglamorous and it is where the fastest movement usually happens, because sites that were invisible were often invisible for a mechanical reason.

Months two and three restructure existing pages and start publishing answer-first content against the queries the baseline showed you losing. Months four to six is where off-site corroboration begins to register, which is the slowest signal because it depends on other people's publishing schedules. From there it compounds, provided the prompt set stays constant and the work does not restart every quarter with a new strategy.

If you would rather have this done for your brand, see how we run GEO.

Common questions