GEO: making a catalog readable and citable by AI agents.

What is GEO?

GEO, for Generative Engine Optimization, covers the practices that make content visible and citable in the answers produced by generative engines: ChatGPT, Perplexity, Google AI Mode, Gemini. The term was formalized in 2024 by a Princeton team in a paper published at the KDD conference, which also introduced a dedicated benchmark (GEO-bench) and demonstrated that optimizing content for these engines can boost its visibility by up to 40% in that benchmark. This experimental measure does not predict a site's citations or traffic in current engines. Established

The distinction with SEO is sharp, and the practitioner's line captures it well: SEO gets you ranked, GEO gets you quoted. A classic engine returns a list of links the user clicks. A generative engine synthesizes an answer from several sources and cites only a few. Being absent from that synthesis means not existing for that shopper. This video lays out the definition and the differences in two minutes.

Why machine readability becomes decisive

An agent can use HTML, an API or a visual representation of the page. Computer-use agents can therefore account for layout. For a catalog, the verifiable concern is access to prices, variants and policies on the surface actually used. The UCP framework separates this readability from the ability to complete a transaction.

“What Is Your AI Agent Buying?” by Allouah and coauthors compares agents in the experimental ACES environment. It observes choices that depend on the model, its version, product position and attributes, across visual and text interfaces. These findings motivate tests for each shopping journey; they do not establish a universal selection penalty for missing information.

What an agent actually reads on a product page

Making a page readable by an agent means exposing clean data in the surfaces it consumes. Five layers of code matter, and each maps to a concrete audit point.

• Initial HTML. Exposing price, availability and description here helps tools that do not execute JavaScript.
• schema.org structured data. Check Product, Offer (price, currency, availability) and brand. An AggregateRating must describe real, visible reviews when they exist.
• Reviews in plain text. Test access on the surface used and consistency with any review markup.
• Policy pages. Shipping, returns, warranties: accessible and explicit, not buried in a script. What is ambiguous for a human is blocking for an agent.
• Discovery files. A complete sitemap, a robots.txt open to AI crawlers, the llms.txt file, and agent surfaces such as agent-card.json.

These five points are the operational translation of the semantic layer of the UCP framework: making catalog data machine-addressable. It is the foundation without which the Agent and Transaction layers have nothing to work with.

Check consistency between visible content and data

Google's structured data guidelines require markup to reflect visible content. This is a useful editorial check; it does not show that every agent verifies every promise or that markup causes a recommendation.

Concretely: a displayed price must match the price declared in the Offer; a visible rating must match the AggregateRating; a shipping promise must point to an accessible policy page. This comparison helps detect inconsistencies; its effect on an agent must be tested separately.

Methods tested on GEO-bench

Aggarwal and coauthors compare several methods on GEO-bench. Adding statistics and quotations yields gains in their experiments. The figure distinguishes two benchmark measures from one measure in their 2024 Perplexity experiment. It measures neither traffic nor sales nor the current behavior of search engines.

For a catalog, this translates into simple moves: factual, self-sufficient descriptions (composition, dimensions, compatibility), quantified attributes rather than superlatives, reviews exposed in plain text, and promises backed by proof. These checks aim to improve accuracy and usefulness. Their effect on citations must be measured by engine and query.

Measuring a catalog's GEO readiness

You only improve what you measure. GEO readiness is audited on the surfaces an engine or agent actually reads: AI crawler accessibility, schema.org completeness, reviews in HTML, claim/proof coherence, sitemap and discovery file health. It is diagnostic work, page by page, before any acquisition spend.

To compare audit tools, ask for the list of surfaces analyzed, the scoring method and examples of reproducible findings. You can also consult the VerityScore presentation and check that the proposed scope matches the pages and uses you need to assess.

GEO and UCP: two sides of the same shift

GEO and the Universal Commerce Protocol address two distinct moments of the same shift, the move to commerce that is addressable by agents. GEO is about discovery: being read, understood and cited by engines and agents at the moment of recommendation. UCP is about the transaction: being queryable and purchasable by a delegated agent once the recommendation is made.

Both depend on the same foundation, data quality. As the Implement UCP guide notes, the protocol carries your offer, it does not improve it: a poor catalog stays poor, whether exposed via UCP or read by a generative engine. GEO is the discipline that strengthens that foundation on the readability and citation side. That is why, in the framework, it belongs to the same semantic layer as merchant readiness.

Frequently asked questions