What is generative engine optimisation? The honest definition.

Where the term came from

The acronym GEO was used in a November 2023 academic paper titled "GEO: Generative Engine Optimization" (Aggarwal et al., 2023) - read on arXiv, by researchers from Princeton, IIT Delhi and others (the first version also listed Georgia Tech and the Allen Institute for AI). It was accepted to the KDD 2024 conference. The paper proposed nine ways of rewriting a page, gave each rewritten page to a generative engine alongside four other search results, and measured how much of the answer it won compared with the unedited page. Adding quotations, statistics and citations did best. That paper is the citable origin. The research code and benchmark dataset are on GitHub if you want to reproduce the lift findings. The industry adopted the term through 2024 alongside parallel acronyms (AEO, LLMO) that describe roughly the same discipline.

How generative engine optimisation works

An AI engine that searches the web does not send the reader to your page. It runs its own searches, pulls a handful of pages into the model, and writes an answer from them. Google's guide to its generative AI features (last updated 10 July 2026) names the two steps. Retrieval-augmented generation, which Google also calls grounding, fetches pages from its search index and writes the answer from them. Query fan-out runs several related searches at once to cover the question from more angles.

So GEO has two jobs. First, get your page retrieved for the searches the engine actually runs. Second, make the retrieved page easy to quote. The first job is the bigger one: if your page is never retrieved, nothing on it can be cited.

What later research found

The 2023 paper is still quoted as proof that a few rewrites lift AI visibility by 40%. The studies since then are more careful.

• The lift only applies once you are already in the sources. A July 2026 review of 45 GEO studies says the original gains are "conditional on a source already being present in a fixed context" and prove "neither organic discoverability nor durable traffic effects".
• Rewriting tricks fade as more sites use them. C-SEO Bench (accepted at NeurIPS Datasets and Benchmarks 2025) found most of these methods "largely ineffective", found traditional SEO aimed at ranking the page higher in the engine's sources more effective, and saw gains shrink as more sites adopted the same methods.
• Being retrieved matters more than page polish. An October 2026 study of 158,099 answers and 468,037 citations found that which pages ChatGPT's own searches retrieved explained 32% of the variation in citations between pages of the same website, against 2.8% for the pages' own properties. Pages whose title and paragraphs matched the questions people ask were cited more. FAQ blocks, structured data, summary boxes and page speed only looked effective when comparing different websites. It also found ChatGPT writes brands it already knows into its searches.
• A second 2026 study points the same way. Across about 2 million citations from 19 B2B SaaS workspaces, matching page content to the prompts people use was the strongest page-level predictor. The FAQ, structured data and Core Web Vitals effects reversed or fell to zero once each website was compared with itself.
• Third-party coverage carries weight. A September 2025 study found AI search showed "a systematic and overwhelming bias towards Earned media (third-party, authoritative sources)" over brand-owned and social content.

Most of these are preprints, and several are observational. Each is one study, not a rule. Read together, they point the same way: be the page that answers the question people actually ask, and get named by sources the engines already trust.

Why GEO matters in 2026

The answer surfaces are already in front of buyers. Google now shows an AI Overview on a large share of informational queries - including, when I checked Google UK in October 2026, on the query "generative engine optimisation" itself. LLM referral traffic to websites grew 72.7% year on year (Duda / Search Engine Journal, April 2026). A larger dataset from Similarweb (September 2026) counted 770.7 million AI referral visits a month worldwide between June 2025 and May 2026, up 117.4% on the year before, while still calling AI referrals "a low single-digit percentage of most sites' total traffic". Traditional Google search still drives more decisions overall, but the consideration step increasingly happens inside a generated answer before anyone clicks a blue link.

In the UK, ChatGPT sent 80.02% of the AI chatbot referrals StatCounter tracked in September 2026, followed by Gemini (10.14%), Copilot (5.61%), Perplexity (2.7%) and Claude (1.51%). That measures traffic sent to websites, not how many people use each tool.

The practical consequence: a brand can hold position 1 in classic Google and still be absent from the answer a buyer actually reads. That absence is invisible in normal analytics. Nothing dropped, no alert fired - the buyer simply never saw you. GEO is the discipline of closing that gap.

What GEO actually involves

• Schema and structured data. Article, Product, Person, Organization. FAQPage and HowTo are still valid schema.org markup, but Google stopped showing FAQ rich results on 7 May 2026 and retired How-to rich results in September 2023. Structured data tells engines who you are and what the page is, but it is not a shortcut to citations. Google's July 2026 guide says structured data "isn't required for generative AI search" and that there is no special markup to add, while still recommending it for rich results. The two 2026 studies above found its citation effect disappeared once each website was compared with itself.
• Content depth. AI engines compose answers from substantive sources. Thin pages get skipped.
• Named author signals. Person schema, credentials, links to professional profiles. Engines like Perplexity weight named authorship heavily.
• Bot access management. OpenAI and Anthropic run separate crawlers for search and for model training. OAI-SearchBot, Claude-SearchBot and PerplexityBot fetch pages for search answers; OpenAI says sites that block OAI-SearchBot "will not be shown in ChatGPT search answers". GPTBot and ClaudeBot are training controls. Google-Extended covers Gemini training and grounding in the Gemini app, but does not affect Google Search. For AI Overviews and AI Mode, the crawler that matters is Googlebot, and Search Console now has a setting, rolled out to all sites by 31 August 2026, to keep your site out of those features.
• Community presence. Reddit was the most-mentioned domain in ChatGPT answers in Ahrefs' September 2026 US data, at 16.8%. In a separate Ahrefs dataset, also US, ChatGPT's Reddit citations fell 88% between March and late September 2026.
• Citation tracking. Tools that sample AI engine answers on a schedule and report what shows up. Profound, AIClicks, Otterly, plus the free tools on this site.

What GEO is not

• Not a replacement for SEO. Traditional Google search still drives more decisions than AI search in 2026 for most UK B2B verticals.
• Not a guaranteed channel. No AI engine has published a confirmed ranking formula. The signals correlate; they are not deterministic.
• Not a separate department. Most teams should treat GEO as 20% extra work on top of an SEO practice that has matured.
• Not something you can fully outsource. The named author, the community presence, the genuine expertise - these need a real person at the brand, not just an agency.

One thing to settle before any of this. If the site already carries a lot of content published in bulk, start there instead. Google's scaled content abuse policy targets volume without value regardless of who or what wrote it, and no amount of GEO work outruns that. It is an AI content cleanup job first, and a GEO job second.

How GEO differs from traditional SEO

Same underlying signals, different scoring weights. SEO optimises for ranking position in a results list. GEO optimises for citation inside a generated answer that may or may not show a list at all. The differences worth knowing:

• Surface vs page. SEO targets the SERP. GEO targets the AI answer surface that often sits above or replaces the SERP entirely.
• Click vs mention. SEO success is a click. GEO success can be a brand mention in the AI answer without a click - influenced pipeline, not direct traffic.
• Rank vs share-of-voice. SEO measures rank position. GEO measures citation frequency, citation share-of-voice against named competitors, and query coverage.
• One engine vs many. SEO is mostly Google + Bing. GEO covers ChatGPT, Perplexity, Claude, Gemini, AI Overviews, AI Mode, Copilot at minimum.
• Crawl + rank vs retrieve + generate. SEO assumes the engine fetches, indexes, ranks and serves your URL. GEO works through retrieval-augmented generation - your content gets pulled into the model's context window and summarised, not served directly.

The work overlaps but the decisions diverge. A page that ranks position 1 in Google can still fail to be cited in AI Overviews if it lacks the extractability signals (answer-first structure, named author, citable statements).

What GEO looks like in the UK right now: my own data

In May 2026 I audited 231 UK professional-services websites across three sectors - 73 law firms, 56 accountancy firms and 102 estate agents - measuring structured data, AI crawler access and captured citations. The headline numbers:

• Roughly a third of firms do not say what they are in structured data. 29% of law firms and 32% of accountancy firms carried no structured data that identified what kind of business they were.
• Silent crawler blocks are common and accidental. 27% of accountancy firms - including PwC and BDO - returned errors to GPTBot or ClaudeBot at the hosting layer while serving browsers normally. Nobody at those firms chose that.
• Zero of 83 readable estate agent homepages used FAQ schema. Google stopped showing FAQ rich results on 7 May 2026, and says structured data is not required for its AI features.
• Citation is rarer than being named. I captured 10 city answers, 9 from ChatGPT and 1 from a Google AI Overview. Together they named around half the estate agents I audited, but only 5 agent websites appeared as a cited source: 4 in the single AI Overview and 1 in ChatGPT. The rest were map pins, not authorities.

One pattern is worth singling out. The 5 cited agent websites all had one thing in common when I read them: an explicit awards or authority claim - an "award-winning" page, a regional award announcement - that the engine could quote directly. Not better homepages. A quotable claim about their standing. Five pages, four of them from one answer, is a pattern I noticed rather than a proven mechanism.

How do you measure whether GEO is working?

Not with rank positions. The working metrics are citation frequency (how often engines cite you across a fixed query set), citation share-of-voice against named competitors, and query coverage (what proportion of your buyer questions mention you at all). Sample the same queries across engines on a schedule and track the trend. For Google's own AI features you now get first-party numbers too: Search Console's generative AI performance report, rolled out to all sites worldwide by 31 August 2026, shows your AI Overviews and AI Mode impressions by page, country and device. I keep a full measurement walkthrough and a list of tracking tools with UK pricing if you want the detail.

The five-principle framework for AI visibility

The discipline shakes down to five working principles. Apply them in this order:

1. Bot access - verify the search crawlers can reach you: Googlebot, OAI-SearchBot, Claude-SearchBot and PerplexityBot. Silent blocks here can keep you out of that engine's search answers. GPTBot and ClaudeBot govern AI training, and Google-Extended governs Gemini training and Gemini app grounding, so decide on those separately.
2. Extractable structure - 30-second answer blocks at the top of each section, short declarative sentences, clean H2/H3 hierarchy. AI engines lift quotable passages, not paragraphs of brand voice.
3. Named-author E-E-A-T - byline + Person schema + credentials. Anonymous brand content loses to named-expert content for the same query.
4. Topical depth - 1,500-2,500 words on priority queries, covering the head plus 3-5 sub-questions. Pages bundling sub-topics win multiple citation rounds in query fan-out.
5. Off-site authority - Wikipedia entity, Wikidata, Reddit presence, citations from credible publications. Slowest layer; compounds over quarters.

That order matters. Bot access first (one-day fix, biggest impact if blocked). Then structure, then named author, then depth, then authority. Spending six months on Wikipedia entity work while OAI-SearchBot is blocked is the textbook failure mode. To appear in ChatGPT search, allow OAI-SearchBot; GPTBot only controls whether OpenAI trains on your content.

The honest read on the discipline

GEO became a marketing category faster than it became a discipline. By 2026 the genuine practitioners and the repackagers are starting to separate, but the noise is still loud. A simple test: if a "GEO service" does not include schema work, named author setup and bot access verification, it is not GEO. It is SEO with a price uplift. If you want the end-to-end walkthrough rather than the definition, I wrote a deeper GEO field guide covering the whole process step by step.