How does ChatGPT choose brands? The sources behind its recommendations

When ChatGPT recommends a brand, it draws on two things: what the model learned about your brand in its training data, and the sources it finds in a live web search for that question. For commercial questions the second route carries more and more weight, because users expect a current answer with sources. This post covers both routes, the factors that shape which sources get picked, and the steps to make your brand one of them.

Where does ChatGPT build its answer from?

ChatGPT's answer combines general knowledge from its training data with the pages it reads during a search. Which one dominates depends on the question: for a general question like "what is a CRM", the model often answers from its own knowledge; for a current, commercial question like "agencies offering GEO services in Istanbul", it goes to web search and shows source links.

Training data

The model has learned what was written about your brand on the web and in licensed sources up to its training date. What matters here is how often and how consistently your brand is mentioned. You cannot change training data directly after the fact, but you can leave an accurate, consistent trail on the web for the next model version.

Live web search

OpenAI says ChatGPT search uses third-party search providers and content supplied by partners. For a site to appear in this search, it has to be open to OAI-SearchBot, OpenAI's search crawler. GPTBot, used for model training, is a separate crawler; the two are managed separately in robots.txt.

Which sources get picked for commercial questions?

For commercial questions, the engines look at third-party sources first: directories, comparison lists, trade publications and user reviews. According to public 2026 studies, every page ChatGPT cited for "best X" commercial queries was list-structured, and three quarters had a year in the title (Position Digital, 2026). For these queries, directories and comparison lists outrank agency websites.

In practice this means your own website is not enough on its own. For category questions, the model knows your brand from what others write about you.

Four factors that shape source choice

Four factors decide source choice: consistency, structure, proof and accessibility. None of them guarantees a recommendation on its own; together they make your brand a candidate.

1. Consistency

Your brand name, one-sentence definition and service list should be the same on your site, on LinkedIn, in directory profiles and in the press. If a brand is called a "digital agency" in one source and a "software company" in another, the model cannot form a clear definition.

2. Structure

More than half of citations come from the first 30 percent of a page (Otterly, 2026). A 40–60 word definition at the top of the page that answers the question on its own, real HTML tables and visible FAQs create passages the engine can quote.

3. Proof

Pages with named sources in the body are cited more often (Otterly, 2026). Instead of adjectives like "leading" or "best", use figures tied to a case study, with the method and period stated.

4. Accessibility

If a page cannot be crawled, it cannot be cited. Allow OAI-SearchBot in robots.txt, do not load content behind tabs or accordions on click, and make sure the full text is in the page's HTML.

Steps to get your brand recommended

The path to being recommended starts with measurement, continues with off-site sources and finishes with on-page structure. The order matters: if you make changes without knowing where you stand, you cannot see their effect.

1. Measure. Build a fixed set of 20–40 questions related to your brand and run them in ChatGPT. Is your brand mentioned, in what position, citing which source? For the method, see how to measure AI visibility.
2. Find the sources. List the pages cited in the answers. Which directories and lists are your competitors in, and which are you missing from?
3. Fix off-site. Complete your directory profiles, make your definition sentence the same everywhere, and make collecting reviews a step in your sales process.
4. Rework your pages. Add a definition block, a sourced data point, a comparison table and a visible FAQ to your service and guide pages.
5. Measure again every month. Same question set, same table. Change only shows up in the series.

The first step of this work is Roicool's AI visibility analysis service; the full engagement is described on our GEO agency page.

What is different for B2B and healthcare brands?

In B2B, purchase decisions follow long research, and part of that research now happens in AI assistants. Buyers build a shortlist with questions like "which CRM for X" or "agencies working in sector Y". So for B2B brands, being mentioned in category and comparison questions comes before being recognized correctly in brand questions. Case studies and profiles in industry directories are the most useful sources for these questions.

For healthcare brands, accuracy comes first. If the model matches a clinic with the wrong treatment or the wrong city, that is not just a visibility problem but a trust problem. For healthcare brands the first step is usually brand questions: are the clinic name, specialist physicians' names, address and treatments offered the same in every source? Plain, unexaggerated language that follows advertising and promotion rules is safer both for regulation and for the model.

What doesn't work?

Some tactics look tempting in the short term but are either ineffective or harm the brand. Adding review markup to your own site that rates yourself is against Google's policy. Text written only for AI that reads as nonsense to people, and stuffing product pages with hidden instructions, cost you trust. Because AI answers are not deterministic, no engagement can guarantee a place in a specific answer; be wary of any offer that promises it.