How ChatGPT Decides Which Brands to Recommend

Two different ChatGPTs

The first thing to understand is that "ChatGPT" is two products depending on the question. For many queries, the model answers from training alone. For others, especially anything involving current products, prices, or "best of," it runs a web search and reads pages before answering, citing them. The two modes favor different things.

Training-mode answers reward long-standing, widely repeated presence: the brands that appeared across the web for years. Search-mode answers reward what is on the pages it opens today: your current site, the review platforms, the lists, the forums. You can influence the second quickly and the first slowly, and the second feeds the first over time.

What the search mode seems to weigh

Watching which sources get cited for commercial questions, a few things recur:

Pages that answer the question directly. Comparison pages, buyer's guides, and "X for Y" pages with a clear answer near the top are cited far more than homepages or feature pages.

Independent sources over self-description. For "best" questions, review platforms, community threads, and publication lists are cited heavily. A brand's own site is cited for facts about that brand, not usually for the verdict.

Agreement across sources. When several cited sources name the same product, it tends to lead the answer. When sources disagree, the answer hedges or drops the contested name.

Recency signals. Dated content, "updated" notes, and current pricing appear to matter for commercial queries, where stale information is a liability.

What it does not seem to weigh much

Keyword density, exact-match domain names, and the volume of pages on your site. A single excellent comparison page outperforms fifty thin blog posts. Social follower counts on their own. And, reassuringly, paid placements: there is no evidence that advertising anywhere changes what the model says.

The mistakes that keep brands off the list

• No page that matches the question. If buyers ask "best X for small teams" and your site only has "Features" and "Pricing," there is nothing to cite.
• Facts that disagree across the web. A directory listing from 2023 with old pricing can be the source the model reads.
• Blocking the crawler. Some sites block OpenAI's search crawler in robots.txt without realizing it, usually from a blanket rule meant for training bots.
• JavaScript-only content. If the key facts render client-side, the crawler may see an empty shell.
• Absence from the sources it reads. If ChatGPT cites G2 and a subreddit for your category and you are on neither, you are competing with one hand.

What this means in practice

AEO GEO Labs approaches ChatGPT visibility as two parallel tracks. On your site: a page for every real buyer question, written to be quoted, with Product, Organization, and FAQ schema and an llms.txt that states what you do in plain words. Off your site: a log of every source ChatGPT cites in your category over a month, then the slow work of getting into those sources with real listings, real replies, and corrected facts.

Then measure. The same buyer questions, every week, with a record of whether ChatGPT named you, cited you, and what it read. That is the only honest way to know whether anything you did worked.

Try it today

Ask ChatGPT your single most valuable buyer question and look at the citations. If your competitor's comparison page is there and yours is not, that is the first page to write. To see the pattern across hundreds of questions instead of one, AEO GEO Labs runs the audit live on a short call.