GEO (Generative Engine Optimization)

Why it matters

Search is splitting into two behaviours, and the second one barely existed a short while ago. People still type queries and click links, but a growing share now asks an assistant a question and reads one synthesized answer, never seeing a results page. If your company is not among the sources that answer draws on, you are invisible in that moment, however well you rank in traditional search. That is the shift GEO responds to. It also changes what winning looks like: not a click to your site, but a mention or citation inside an answer someone else’s tool generated, which you cannot buy outright and cannot fully control.

The scale is not a guess. In July 2025 Google said the AI Overviews that appear above its search results reach more than two billion users a month. Pew Research Center studied the browsing histories of around 900 US adults in March 2025. When a results page carried an AI summary, people clicked a traditional result on 8% of visits, against 15% when no summary appeared. The sources the summary cited were clicked on about 1% of visits. Visibility is moving inside the answer, and getting there takes different work than reaching the top of a results page.

Does GEO replace search engine optimisation?

No. Search engine optimisation gets a page within reach of the machine, and GEO makes it quotable. Generative engines lean largely on the same indexes and quality signals as traditional search, so a page still has to be readable by the bot, indexed and credible. Content that ranks well in search results also ends up as a source for AI answers more often than content that doesn’t. The differences are in the goal and the level of detail. SEO aims for a place on a list of links; GEO aims for a citation in an answer the machine writes. SEO optimises the page as a whole, whereas in GEO the unit is a passage of text the model can lift into its answer as it stands. Success is measured in mentions and citations rather than clicks and rankings. The practical work still overlaps heavily, and the same clearly structured page serves both.

How do AI search engines choose their sources?

When a buyer asks an assistant for, say, a suitable e-commerce platform or a subcontractor in Finland, the assistant does not run a single search. Google describes its AI Mode as breaking the question into a set of parallel sub-queries (query fan-out), pulling candidate pages for each and assembling an answer from them. The model also reads a page in fragments rather than as a whole. The paragraphs that make it into an answer are the ones that answer a single question clearly and stand on their own, apart from the surrounding text. Structure is what decides it. A short, fact-heavy paragraph under a subheading is more useful to the machine than the same information scattered across the page. Engines also weigh different things, such as freshness and source types, and showing up in one does not guarantee showing up in another.

Techniques shown to work

The term GEO was introduced by a group of researchers from Princeton University, Georgia Tech and IIT Delhi, whose study (Aggarwal et al.), published at the KDD 2024 conference, tested nine optimisation techniques on a set of 10,000 search queries. The clearest gain in visibility within generative answers came from adding statistics, quotations and source citations to the text, by around 40% at best. Keyword stuffing, the oldest trick in traditional search engine optimisation, did nothing. The findings back a few practical ground rules:

• Answer first: the definition or core answer right under the heading, reasons and background after it.
• One thing per section, and subheadings that answer real questions.
• Named sources, figures and direct quotations instead of vague generalisations.
• Schema.org markup (such as DefinedTerm and FAQPage) that tells the machine how the page is structured.
• A visible date and regular updates, because some engines weigh freshness.

Trust and recognisability

Beyond the text, the engine judges the source. Content is easier to surface in an answer when the web makes it unambiguous who the company is, what it does and for whom. Consistency is a technical requirement here. When a company and its topics appear under the same names on the site and off it, the machine can tie scattered mentions to a single entity. Trust is built by the same things search engines have rewarded for a long time: named authors, visible contact details, real customer examples and claims that have a source behind them.

Bot access and llms.txt

It all starts with access. A generative engine’s bot has to reach the page, and that access is governed by the robots.txt file. There are three kinds of bot. Training bots collect material for training models (OpenAI’s GPTBot, Anthropic’s ClaudeBot, Google’s Google-Extended). Search bots build the index for AI search (OAI-SearchBot, Claude-SearchBot, PerplexityBot). User bots fetch a page when an assistant needs it mid-conversation (ChatGPT-User, Claude-User). Blocking a training bot is a matter of principle that does not drop your site from AI search. Blocking a search bot does. Google-Extended covers the training of Gemini models and the use of the page as a source for Gemini apps, not visibility in Google Search. AI Overviews uses the same Googlebot as the rest of Google Search. A separate llms.txt file, the content map for language models that Jeremy Howard proposed in September 2024, is so far mostly a talking point. No major AI provider has confirmed using it, and Google has said it does not support it. The file does no harm, but there is no proven benefit to it.

Measurement

Measurement in GEO is sampling, because there is no ranking list to check. The same questions that matter to buyers are run regularly through different AI search tools, and the answers are logged for whether the company is mentioned or cited and how its share compares to competitors. Tools have grown up around this, both as add-ons to SEO platforms and as dedicated monitoring services. A second measure is your own analytics, where traffic from AI services and bot fetches show up as their own sources. Reading the results takes patience. The same question can produce a different answer and different sources on different days, because the answer is generated afresh each time. A single measurement says little; the trend says a lot.

Criticism and open questions

The field is young, and it shows. There is not even agreement on the name. The same idea also goes by AEO or LLMO. How the engines pick their sources is not public and changes without notice, so individual technical tricks go stale fast. The research evidence is still thin and mostly empirical, and the percentages from any single study should not be read as promises about your own visibility. There is an open business question too. A citation without a click brings no visitor, so the benefit is at first recognition and trust, whose value only materialises as the buyer moves along. And as more and more players optimise in the same ways, the edge levels out. The most durable footing is the same as it has always been in search: content worth citing.

In practice

A firm publishes a clear, well-structured explainer on a topic it wants to be known for, written so an assistant can lift a plain definition and attribute it. Later, when buyers ask an AI tool about that topic, the tool summarises the firm’s page and names it as the source. The direct traffic matters less than being the reference the machine trusts enough to quote.

Frequently asked questions

Is GEO worth it for a company?

It depends on the business and its specific situation, but in many cases, yes. GEO can improve a company’s visibility in AI-generated answers, especially when potential customers are looking for information, comparing options, or seeking expert insights. The key is to create clear, trustworthy content that answers users’ questions directly.

How quickly does GEO deliver results?

More slowly than paid advertising. Once a page is readable by the bot and indexed, it can start to show up in AI answers within weeks, but a reliable position builds over months. Because answers are generated afresh each time, visibility also varies day to day, so it is worth tracking progress as a trend rather than as individual hits.

Do you need a dedicated tool for GEO?

Not at the start. GEO begins with content and the technical readability of the site, not with buying software. For monitoring there are tools, both as add-ons to search engine optimisation platforms and as dedicated services, and they help measure mentions and citations systematically. A small company can get a long way by running its own key questions through different AI search tools by hand and logging the results.

Where should you start with GEO?

Pick one topic you want to be known for and write a page that answers the buyer’s question in the very first paragraph. Make sure the bots can reach the page and that it is indexed. Add named sources, figures and clear subheadings so the machine can lift a passage from the page as it stands. The same work serves traditional search too.

Is GEO the same as AEO?

In practice, yes. AEO, or answer engine optimization, and GEO both mean shaping content so that AI surfaces it in an answer and names it as the source. The same phenomenon also goes by LLMO. The range of names shows the field is young and the term has not yet settled.

Sources

• Aggarwal et al.: GEO: Generative Engine Optimization (KDD 2024) – the study that introduced the term, testing nine techniques on 10,000 queries.
• Pew Research Center (2025): Google users are less likely to click on links when an AI summary appears – March 2025 browsing data on click-through.
• Alphabet, Google’s Q2 2025 earnings – AI Overviews at more than two billion monthly users.
• Google Search Central: Google’s crawlers – Googlebot, Google-Extended and their roles.
• OpenAI: bots documentation – GPTBot, OAI-SearchBot and ChatGPT-User.
• Anthropic: web crawler documentation – ClaudeBot, Claude-SearchBot and Claude-User.
• Perplexity: bots documentation – PerplexityBot and Perplexity-User.
• The llms.txt proposal (Jeremy Howard) – a content map for language models.
• Wikipedia: Generative engine optimization – overview and sources.