How ChatGPT, Perplexity and Gemini choose which brands to mention

Nobody publishes the ranking function

No one outside OpenAI, Perplexity, Google, xAI or Anthropic can show you the code that decides which brands appear, and none of those companies publishes it. What exists instead is observable behaviour: which sources show up, which brands show up beside them, and what changes when the sources change.

Plenty of articles offer a clean formula, typically five factors with confident percentages. Some hold up when you check them against real answers. None of them is the system itself: their authors measured outputs, not the code. Treat each claim as a hypothesis, run it against your own prompt set for two weeks, and keep the ones that predict what you see.

To reason about it at all, separate three parts of an answer. Retrieval decides which pages are in play. Generation writes the answer from those pages. Memory is what the model absorbed during training. A brand gets named when the retrieved passages support it, or when the model reaches for it from memory because the passages say little.

The four inputs you can observe

What the retrieved sources say about you, whether you fit the question as asked, whether sources agree about you, and how prominent you already are outside these answers.

• Source content. An answer can only include what the retrieved passages contain. If no page in the shortlist compares you with the alternatives, the comparison never happens.
• Query fit. Fit the question as asked: category, segment, budget, use case. A prompt about the best tool for enterprise procurement will not surface a self-serve product priced for freelancers, no matter how good its reviews are.
• Consistency across sources. When your site, review pages and directories agree on the product name, category and pricing, engines get one story. When they disagree, something has to give, and it will not be the engine that guesses wrong in your favour.
• Prominence in training-visible material. How often your brand appears in what a model read during training. When retrieval is thin, familiarity acts as the tiebreaker, which is how established brands keep getting named first.

How ChatGPT, Perplexity, Gemini and Claude differ

Perplexity answers from live web search and shows the pages it used. ChatGPT searches when it decides to and otherwise writes from training. Claude answers from training data unless web search is enabled. Gemini blends Google's index with what the model learned.

How strong is the search link for ChatGPT? A February 2025 study by Seer Interactive matched 500 ChatGPT search citations against results for the same 100 queries: 87% matched Bing's top 20 organic results, against 56% for Google. The authors called their own sample directional. The fair reading: being findable by the search index behind the answer still matters, and one 100-query sample did not close the question. Perplexity's own documentation describes web-grounded answers with citations in one call (Perplexity docs), which is why its answers expose sources more consistently than a chat window with search switched off.

What a brand can actually influence

You control your pages, your profiles and what you correct on the sites you do not own. Model weights, sampling and an engine's source shortlist are out of reach, and that is where most GEO anxiety comes from.

• Pages that answer the question completely. The direct answer in the first 40 to 60 words, detail after, headings that state the question, tables for comparisons. Engines extract passages and passages that stand alone get used.
• Consistent facts about your brand. Name, category, pricing and key claims written the same way across your site, documentation and listings, so corroboration is possible.
• Independent sources. Reviews, comparisons, forum threads and news, written by other people. When your own page and an outside review agree, an answer has something it can repeat.
• Crawl access. If an AI crawler cannot fetch a page, retrieval cannot shortlist it. Check which bots your robots.txt allows before blaming the algorithm.
• Corrections on third-party pages. Outdated pricing on a directory, a stale review, a forum answer with the wrong acquisition story. Slow to fix and often the highest return work there is.

For the measurement half of this: Competitor gaps in advisible shows the answers that name a rival instead of you, and Sources lists which domains fed those answers, with CSV export when you want to hand the list to whoever fixes the pages.

How to check a claim about the ranking algorithm

Ask for the sample, the engine and the date, then reproduce the claim against your own prompt set before you act on it.

Studies in this area use different engines, different query sets and different months, and the products underneath them ship changes without telling you. A result reported in March can describe a system that no longer exists by September. Prefer primary sources with a stated sample over a summary of a summary, and date every claim you keep.

The practical version: run the same 15 to 25 prompts daily for two or three weeks, read the answers and their citations yourself, and note which sources recur across engines. Patterns that survive that routine are worth acting on. Start with which sources AI engines cite for what the citation data looks like, and why AI answers change for why one day of checking tells you nothing.