How GEO Works: How Generative AI Learns About, Retrieves and Recommends B2B Brands

GEO works on the two paths by which a generative model knows your brand. Training data shapes what the model believes before anyone asks. Live retrieval, through query fan-out and grounding, decides which current sources it reads when a buyer asks. A brand gets recommended when both paths return consistent, specific, corroborated facts that match the buyer's constraints, across enough independent sources.

This is part three of our series on how search and AI visibility work. How AEO works covered the passage: how an engine extracts and cites a piece of text. This part covers the brand: how models form a view of a company and decide whether to recommend it. For definitions and history, see what is generative engine optimization and GEO vs AEO. For a breakdown of trust signals, see how LLMs decide which brands to recommend.

The short version

1. Learn During training, a model absorbs what public text says about your company and category.
2. Resolve When a buyer asks, the model works out which real-world entities the question refers to.
3. Fan out It splits the question into sub-questions and runs several searches in parallel.
4. Ground It reads current pages from a search index and ties its answer to them.
5. Shortlist It assembles a set of candidate vendors from those sources and its own knowledge.
6. Recommend It names the vendors whose facts best fit the buyer's constraints and describes them.

Two paths into an answer: training data and live retrieval

Every generative answer about a company draws on some mix of what the model learned and what it just read. The two paths use different crawlers, run on different clocks and respond to different work.

The controls are separate by design. OpenAI's crawler documentation says GPTBot collects content that may be used to train its foundation models, OAI-SearchBot surfaces sites in ChatGPT search, and each setting is independent. Anthropic's crawler page draws the same line between ClaudeBot for training and Claude-SearchBot for search quality. Google says Google-Extended governs use of content for Gemini training and grounding in Gemini Apps and Vertex AI, and does not affect inclusion or ranking in Google Search. Common Crawl's CCBot builds an open repository of web crawl data that anyone can use. The full list is in how AI crawlers work.

From prompt to brand recommendation, step by step

Step 1: What the model already believes

A model's parameters hold a compressed picture of its training text. It has no list of facts about your company. It has patterns: which names appear near which categories, which capabilities are attributed to whom, which companies appear together. If most of the public text about you predates a rebrand or a pivot, the model's default description will too.

This path matters more than teams expect, because engines do not always search. Anthropic's web search documentation says Claude answers directly for stable knowledge and searches for current or changing information. A buyer asking "what does Acme do" in a chat without search gets the training-data answer. Our guide to getting your brand into an LLM's knowledge covers this path in depth.

Step 2: Entity resolution

Before a model can say anything useful about "Acme", it has to decide which Acme. Search systems solve this with knowledge graphs; Google's Knowledge Graph holds millions of entries describing people, places and things, typed with schema.org vocabulary. Language models resolve entities from context: the category, location, product names and people mentioned alongside the name.

Ambiguity is expensive. A company that shares a name with a consumer brand, an older firm or a place competes for the same tokens, and the model may blend facts from both. The fixes are covered in how entities and schema work and brand name disambiguation.

Step 3: Query fan-out

When the model searches, it expands the question. Google described AI Mode's query fan-out technique at launch in May 2025 as breaking a question into subtopics and issuing many queries at once, with a Deep Search mode that can run hundreds of searches for one report. OpenAI's web search guide describes reasoning models that search, analyze results and decide whether to keep searching.

A B2B prompt such as "which contract management platforms work for a regulated fintech on Salesforce" might fan out into sub-questions about Salesforce-native contract tools, fintech compliance requirements, vendor comparisons, pricing models and reviews. Each sub-query returns its own sources. A vendor that has a page answering only the head term competes in one of those searches, not all of them.

Step 4: Retrieval and grounding

Grounding means tying the answer to retrieved sources. Google's guide to generative AI features in Search defines it as using core Search ranking systems to retrieve relevant, current pages from the Search index to improve quality, accuracy and freshness. The Gemini API grounding documentation lays out the same loop for developers: the model decides whether search would improve the answer, generates and runs queries, processes the results and returns a response with inline citations to source URLs.

This is retrieval-augmented generation as described in the 2020 RAG paper, applied to the open web. Two consequences follow. Classic search strength carries over, because the retrieval step uses a search index. And what the model reads can override what it learned, which is how a current, well-sourced page corrects an outdated training-data description.

Step 5: Assembling a consideration set

For recommendation questions, the model gathers candidate vendors from everything it read plus its own knowledge. Third-party pages carry a lot of weight here, because a single roundup or review category page lists many vendors at once. Each engine leans on different source types. Orbit Media's 2026 study of B2B prompts found ChatGPT favored primary sources such as documentation, research and vendor pricing and help pages; Perplexity's most cited domains were LinkedIn, YouTube and Reddit; Claude leaned on review sites and listicles, with Clutch its top third-party domain; and Gemini made heavy use of listicle roundups.

If you are absent from the pages engines use to build lists, you are absent from the list. Our guide to off-site assets that get B2B brands cited covers which ones to build.

Step 6: Recommending and describing

The model then matches candidates against the buyer's constraints and writes. A vendor gets named for "a 300-person fintech on Salesforce" when the sources the model read state, explicitly, that it serves companies that size, works with regulated financial firms and integrates with Salesforce. Implied fit does not survive this step. When sources conflict, answers tend to hedge or drop the vendor. And because generation is probabilistic, the same prompt can name different vendors on different runs, which is why measurement needs repeat runs; see LLM answer volatility.

What the GEO research actually established

The term comes from "GEO: Generative Engine Optimization" by Aggarwal and colleagues, published at KDD 2024. The authors built GEO-bench, a benchmark of diverse queries across domains with relevant web sources, and reported that content optimization can boost visibility in generative engine responses by up to 40 percent. They also found that the efficacy of strategies varies across domains.

Read that carefully. It shows that content changes can move visibility in generative answers under test conditions. It does not promise a figure for your category, and the domain variation is the more useful finding for B2B teams: what works for consumer health questions may not work for procurement software. Test against your own prompt set.

The four conditions for being recommended

This is our working model, drawn from the mechanism above. A brand appears in a recommendation when it meets all four:

1. Known. The model can resolve your name to one entity, and its trained description is roughly right.
2. Retrieved. Your pages and the third-party pages about you are crawlable and rank for the sub-questions fan-out produces.
3. Corroborated. Independent sources repeat the same core facts: what you do, for whom, with which systems, with what proof.
4. Fitted. Those facts explicitly match the constraints buyers put in prompts: segment, industry, integration, region, compliance.

Corroboration has to be earned. Google's guide says pursuing inauthentic mentions is not effective, because its core systems focus on high-quality content and other systems block spam.

How to tell which path an answer came from

When an engine describes your company wrongly, the fix depends on the path that produced the error. Run the same prompt two ways and compare.

1. Search off. In an assistant that lets you turn web search off, or through an API call with no search tool, ask "What does [Company] do, and who are its main competitors?" The answer reflects the model's own knowledge. Note the category, products, customers and competitors it names.
2. Search on. Ask the same question with search enabled. Note which sources are cited and whether the description changes.
3. Compare. If the search-off answer is wrong and the search-on answer is right, retrieval is correcting stale training data, and consistent public facts plus future model releases will fix the default. If both are wrong, the error lives in the pages engines retrieve. Open the cited sources and find the outdated claim.
4. Check the entity. If either answer mixes in facts about a different organization, you have an entity resolution problem, not a content problem.
5. Repeat. Run each version at least three times. A single run can mislead, because generation varies.

Developer tools make the retrieval path easier to see. OpenAI's web search returns a sources field that lists every URL the model consulted, not only the ones cited inline, and Anthropic's web search results include each page's URL, title and age. Those lists show whether your pages were read and passed over, or never retrieved at all. The first is a content and corroboration problem. The second is an access or ranking problem.

What you control: the GEO checklist

• Entity basics. One company name, one one-sentence description and one category label, used identically on your site, LinkedIn, review platforms and partner directories.
• Crawler policy. Search crawlers allowed; training crawlers allowed or blocked as a deliberate decision recorded with legal, not a CDN default.
• Fit pages. Pages that state segment, industry, integration and region fit in plain sentences, one per major buyer constraint.
• Sub-question coverage. Pages for pricing model, implementation, security, alternatives and comparisons, so fan-out searches find you.
• Third-party presence. Complete review profiles, inclusion in credible category roundups, partner marketplace listings, and executive commentary in industry publications.
• Proof that travels. Named case studies with specific outcomes that others can reference.
• Structured identity. Organization markup with sameAs links to your official profiles.
• Correction loop. A monthly check of how engines describe you, with and without search, and a named owner for fixing the sources behind errors.

Illustrative example: a pharma cold-chain vendor left off the shortlist

Illustrative example. The company, prompts and findings below are invented to show the mechanism. They are not a client result.

A European company sells temperature monitoring for pharmaceutical logistics. Its sales team keeps hearing that buyers "asked ChatGPT for options" before the first call, and the company is never on those lists. The marketing lead runs twenty buyer prompts across four engines, three times each, and works through the four conditions:

1. Known. Without search, two engines describe the company as a general IoT hardware maker, its positioning before a pivot three years ago. One confuses it with a similarly named logistics firm.
2. Retrieved. With search, engines fetch the homepage but none of the product pages. The pages exist, but core copy loads through client-side JavaScript that AI crawlers do not run.
3. Corroborated. The two category roundups engines cite most do not mention the company. Its review profile still carries the old category.
4. Fitted. Prompts ask about EU Good Distribution Practice compliance. The site mentions "regulatory readiness" but never names the guideline, validation documents or supported carriers.

The fix sequence follows the conditions in reverse order of cost: server-render product pages, publish a GDP compliance page with specifics, update review profiles and directory listings to the current category, add Organization markup with sameAs links, then pitch the roundup authors with a documented customer example. Live answers are re-measured monthly. Training-data answers are tracked separately, with the expectation that they change only when new model versions ship.

Glossary

• Consideration set: the candidate vendors a model gathers before choosing which to name.
• Entity: a distinct real-world thing, such as a company or person, that systems can identify unambiguously.
• GEO-bench: the query benchmark introduced in the 2023 GEO paper to test visibility in generative answers.
• Grounding: connecting a model's answer to retrieved sources to improve accuracy and freshness.
• Knowledge cutoff: the date after which a model's training data contains no information.
• Parametric knowledge: what a model learned in training and stores in its weights.
• Query fan-out: splitting one question into many related searches run in parallel.
• Search crawler: a bot that collects pages so an engine can retrieve and cite them in answers.
• Training crawler: a bot that collects pages that may be used to train future models.

Next in the series

The next explainer narrows to one engine family: how Google AI Overviews and AI Mode select and cite sources, using only Google's own documentation. To see where your brand stands across all four conditions, request a free audit or work through the B2B AEO and GEO playbook.