Generative engine optimisation (GEO): what it is, what the research found and how to do it.

GEO in two sentences.

Generative engine optimisation is the practice of improving how often, and how accurately, AI answer engines use and credit your content when they generate a response. It covers whether the engine can retrieve your pages, whether your passages are worth quoting, and whether other sources support what you say.

In the United States, and in the original research, it is spelled “generative engine optimization”. In marketing, GEO has nothing to do with geography or geotargeting.

If you already know you want help rather than a definition, our generative engine optimisation services page explains how we scope the work.

Where the term came from.

The term comes from an academic paper, not from a marketing team. “GEO: Generative Engine Optimization” was written by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande: four authors at Princeton University and the Indian Institute of Technology Delhi, and two independent researchers. The first version went on arXiv on 16 November 2023; a revised version was accepted to KDD 2024, the ACM knowledge discovery and data mining conference, held in Barcelona on 25–29 August 2024.

The authors started from a worry about website owners. Their abstract says generative engines answer questions “by synthesizing information from multiple sources and summarizing them”, which leaves content creators with “little to no control over when and how their content is displayed.” They proposed GEO as a way “to aid content creators in improving their content visibility in generative engine responses.”

Two details are often lost second-hand. The paper defines a generative engine as two parts: a search engine that fetches sources and one or more language models that write from them. And it treats the engine as a black box: the authors changed the text of web pages and measured what happened to the answer.

What the research tested.

The researchers built a test set of 10,000 questions, fetched the top five Google results for each, asked a language model to write a cited answer from those five pages, and then rewrote one page at a time in nine different ways to see whether its share of the answer changed.

The test set, GEO-bench, draws questions from nine sources across 25 subject areas: 80% informational, 10% each transactional and navigational. Answers came from OpenAI’s GPT-3.5 Turbo model, sampled five times per question.

How “visibility” was measured

An AI answer has no ranked list, so the paper used two measures. Position-adjusted word count is the share of the answer’s words in sentences citing your page, weighted towards the top. Subjective impression is a language model’s score for seven qualities, such as how influential the citation is and how likely a reader is to click it.

The nine methods

Six of the nine only change how existing content is presented. The three that add something new, quotations, statistics and source citations, proved most useful.

What it found.

Evidence beat keywords. Pages rewritten to include quotations, statistics and cited sources took a larger share of the generated answer, and the paper reports that the best methods improved on the unchanged page by 41% on position-adjusted word count and 28% on subjective impression. Keyword stuffing scored below the unchanged page.

Lower-ranked pages gained the most

The result that matters most to smaller firms is in the paper’s Table 2. When every source page was optimised at once, the page Google ranked fifth gained far more than the page ranked first, which often lost ground.

Quotation addition and statistics addition showed the same pattern: −22.9% and −20.6% for the page ranked first, +99.7% and +97.9% for the page ranked fifth.

The authors read this as good news for smaller publishers: a generative engine works from what a page says, so link strength matters less once the page is retrieved. One condition applies. The page still had to be among the five retrieved.

Three more findings worth knowing

• ✓It depends on the subject. Cited sources helped most on factual and law-and-government questions; statistics on law, debate and opinion; quotations on people-and-society, explanation and history questions.
• ✓Combinations helped. On a 200-question sample, fluency plus statistics beat any single method by more than 5.5%.
• ✓It held up on Perplexity. On 200 test questions run through Perplexity, quotation addition improved position-adjusted word count by 22%, statistics addition improved subjective impression by 37%, and keyword stuffing did 10% worse than the unchanged page.

What the paper does not prove.

The paper shows that evidence-rich passages win a larger share of an answer once a page has been retrieved. It does not show how to get retrieved, it does not study company recommendations, and it says nothing about enquiries or revenue.

The authors list their own limits. They tested two engines; methods “may need to adapt over time as GEs evolve”. They did not measure effects on search rankings. Their questions were mostly informational, and the answering model, GPT-3.5 Turbo, dates from 2022. In the Perplexity test the pages were uploaded as files, so it did not test Perplexity’s live web retrieval. We cover how Perplexity finds and cites live sources separately.

There is also a gap between a research question and a buying question. “Which accounting firms in Leeds work with software companies?” is answered by naming firms, not by quoting a paragraph. Later research points beyond your own website: a paper posted to arXiv in September 2025 by Mahe Chen and colleagues reported that AI search engines show “a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content.”

Our own small study found a different mix, a warning against treating any single study as the rule. Across 36 answers to 12 B2B buying questions in November 2026, about 91% of the 203 cited sources were company or vendor websites. The method and the caveats are in our study of where AI answers get their sources. Question type, market and engine all change the result. We cover what changes for UK B2B firms in a separate guide.

How generative engines pick sources.

A generative engine does not answer from memory alone. It searches, retrieves a small number of pages, reads passages from them and writes an answer with links. To be cited you have to survive each of those steps. Our LLM SEO guide covers the two routes into an answer, training data and live retrieval.

What the engine makers document

Google’s guide to generative AI features says they are “rooted in our core Search ranking and quality systems”: retrieval relies on “our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index”, and query fan-out is “a set of concurrent, related queries generated by the model”. OpenAI says sites opted out of its OAI-SearchBot crawler “will not be shown in ChatGPT search answers, though can still appear as navigational links.” Perplexity’s help centre says it searches the internet “in real-time” and that “Each answer includes numbered citations linking to the original sources”. Bing’s AI Performance report makes the fan-out step visible: its “grounding queries” are “the key phrases the AI used when retrieving content that was referenced in AI-generated answers.”

For one engine in depth, see our guide to Google AI Mode for B2B sites and how ChatGPT decides which companies to recommend.

The lesson: retrieval comes first. A page that is blocked, not indexed or unreadable without JavaScript never reaches the step the GEO paper studied. Our guide to AI crawlers, robots.txt and schema covers that technical layer.

What the engine makers say.

Google accepts the term exists and declines to treat it as a separate discipline. Microsoft uses the term for its own tooling. Neither publishes a ranking formula for AI answers.

Google’s guide, last updated 10 July 2026, has a note headed “What about ‘AEO’ and ‘GEO’?” Its answer: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” It advises checking third-party AEO or GEO services against its guidance on evaluating SEO advice. Its AI features page says “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

Microsoft takes a warmer line on the name. When Bing launched AI Performance in Bing Webmaster Tools on 10 February 2026, it called the release “an early step toward Generative Engine Optimization (GEO) tooling”.

Together, the two positions give a sensible reading: GEO is a real change in where buyers see you, built on the same foundations as search. Be wary of anyone who sells it as a secret technique, or who says nothing has changed.

GEO next to SEO.

GEO does not replace SEO; it changes what success looks like. SEO earns a position in a list. GEO earns a sentence, a name or a link inside an answer the buyer may never click away from.

That last point is the commercial reason the subject matters. Pew Research Center’s analysis of the browsing of 900 US adults in March 2025 found that people who saw an AI summary on Google clicked a traditional result in 8% of visits, against 15% for those who did not. We chart that and the other studies in our article on zero-click searches and AI Overviews.

If fewer buyers click, being named in the answer carries more weight. Three things differ most: the unit of competition (a passage, not a page), the role of other websites, and the measurement (mentions and citations, not positions).

You will also see “AEO”, answer engine optimisation, used for much the same work. We compare the three labels in AEO, GEO and SEO: a leader’s guide.

A practical B2B method.

For a B2B company, GEO is six jobs done in order: choose the questions, record a baseline, clear technical access, strengthen the passages, fix what other sites say, and re-measure. The research supports step four most directly. For software companies, our guide to GEO for B2B SaaS works through it page type by page type.

Steps 1 to 3: the groundwork

Write 20–30 questions the way a buyer would ask a colleague, including comparison and “who should I talk to” questions. Ask each one in each engine your buyers use and note, with the date, whether you are named or cited, who else appears and which sources are linked. Our guide on how to check your AI search visibility gives a manual method. Then confirm the search crawlers are not blocked, key pages are indexed in Google and Bing, and the main text is in the page’s HTML.

Step 4: strengthen the passages

This is the step the GEO paper tested. Take the pages that should answer your buyer questions and, for each main claim, add what the research rewarded: a specific number with its date, the source it came from, or a quotation from a named, credible person. Open each section with a direct answer. Do not add keywords for their own sake. The detail is in our guide to page patterns AI engines quote.

Steps 5 and 6: other sites, then re-measure

Engines read review platforms, directories, partner listings, trade press and community threads as well as your site. Make the facts match everywhere, then earn genuine coverage where buyers’ questions are already answered; seeded praise is not the goal. See the off-site sources AI trusts for B2B firms for a ranked list. Then run the same questions monthly, with the same wording, and compare against the baseline.

Measuring GEO.

There is no single GEO score that every tool agrees on. Measure it with a fixed question panel you run yourself, the two free reports the search engines provide, and the enquiries that follow.

The panel is the step-2 baseline repeated: named, cited or absent, per question and engine. Google announced Search Generative AI performance reports for Search Console on 3 June 2026 and says they reached all websites by 31 August 2026; they show impressions: “How often URLs from your site appeared in generative AI features in Search and Discover.” Bing Webmaster Tools’ AI Performance report shows total citations, cited pages and a sample of grounding queries for Copilot and Bing’s AI summaries; Bing says it “does not indicate ranking, authority, or the role of any page within an individual answer.” Neither report shows enquiries, so keep your analytics and CRM alongside.

To see your position before building a panel, our free AI visibility report runs a sample of buyer questions for your company and shows who is named and cited. For leadership teams, how to report AI visibility to your board sets out six measures on one page.

For budgets, the scope, published market ranges and red flags are in what AI search optimisation costs.

The GEO readiness check.

Tick each statement that is true today. You get a score out of 100 and the three gaps to close first. Nothing you tick is sent anywhere.

The weights are our judgement of what blocks the most, not measured effects.

Questions buyers ask.

What is GEO in digital marketing?

GEO stands for generative engine optimisation. It is the work of making a company’s content and reputation easy for AI answer engines such as ChatGPT, Google’s AI Overviews and Perplexity to retrieve, quote and cite. It is unrelated to geographic targeting.

Does GEO replace SEO?

No. Google says optimising for generative AI search is “still SEO”, and its AI features retrieve pages from the normal search index. GEO adds attention to quotable passages, third-party sources and mention-based measurement on top of sound SEO.

Which GEO strategies does the research support?

In the original paper, adding quotations, statistics and cited sources performed best, and smoother, simpler writing also helped. Keyword stuffing scored below the unchanged page. The tests used a research benchmark and an older model, so treat the results as direction.

How long does GEO take to show results?

No engine publishes a timescale, and the research did not measure one. OpenAI notes that a robots.txt change can take about 24 hours to register for search. Content and reputation changes depend on recrawling, which is why a dated baseline matters.

Should we hire someone for GEO or do it in-house?

A team that knows its buyers can do the question list, baseline and passage edits in-house. Outside help earns its place on technical access, third-party sources and consistent measurement.

If you want that support, our GEO and AEO services page sets out how an engagement starts.

Sources: Aggarwal et al., “GEO: Generative Engine Optimization”, arXiv:2311.09735; full text, version 3, 28 Jun 2024; Chen et al., “Generative Engine Optimization: How to Dominate AI Search”, arXiv:2509.08919, 10 Sep 2025; Google Search Central, guide to generative AI features; Google Search Central, AI features and your website; OpenAI, overview of crawlers; Perplexity help centre, how Perplexity works; Bing Webmaster Blog, 10 Feb 2026; Pew Research Center, 22 Jul 2025; Google Search Central Blog, 3 Jun 2026. All accessed 12 November 2026.