When someone asks ChatGPT "what's the best project management tool for a 10-person agency," something has to decide which two or three tools get named. It isn't a ranking algorithm in the search-engine sense, and it isn't random. It's a retrieval-and-synthesis process with its own logic, and that logic has been studied enough now to describe with some confidence.
The five factors that keep showing up in the research
Analysis of how large language models select brands to recommend converges on five recurring factors: mention frequency across the sources the model draws from, the authority of those sources, the sentiment attached to the brand where it's mentioned, how well the brand fits the specific intent behind the question, and whether the surrounding content is structured clearly enough to extract (IndexLab's 2026 analysis of LLM brand-selection behavior).
A separate breakdown frames it slightly differently but lands on similar ground: training data frequency, contextual relevance to the specific prompt, authority signals, recency, and how sensitive the answer is to prompt wording (GeoVector's research on LLM brand-recommendation logic). The overlap between two independently produced breakdowns is itself useful signal — mention frequency and source authority aren't a guess, they're the consistent finding.
It's about problems, not brand names
One detail matters more than it might seem: LLMs are answering problem-driven questions, not producing "top vendors" lists from memory. The realistic query isn't "best CRM" — it's "best CRM for a five-person real estate team that doesn't want to pay for seats they won't use." The model is matching the specific shape of that question against the content and mentions it has access to, not surfacing whoever has the loudest brand campaign (industry commentary on LLM brand-selection behavior).
That has a direct implication: generic "why choose us" content rarely gets cited, because it isn't written to answer a specific, narrow version of a buyer's question. Content built around the actual situations buyers describe — by size, industry, budget, or constraint — has a much better shot at being the passage an LLM lifts into an answer.
Retrieval and citation are two separate steps
It's worth separating two things that get conflated: being retrieved by the model when it searches the web, and being cited in the final answer. Research on ChatGPT's citation patterns found the model cites only around 15% of what it retrieves — meaning a page can be read and considered without ever being named (Kime.ai's ChatGPT citation research). The same research found citations concentrate heavily in the first 30% of a page, and that a relatively small set of roughly 30 domains account for about two-thirds of citations within any given topic, according to analysis reported by Search Engine Land.
That's a meaningfully different game than ranking. It means the goal isn't just "get retrieved" — it's earning a spot among the narrow set of sources that get cited once retrieval happens, which usually means being clear, well-structured, and mentioned favorably across multiple independent sources rather than just your own site.
Third-party mentions outweigh brand-owned content
This is probably the single biggest mindset shift for marketers coming from an SEO background. A brand's own website matters, but earned mentions — reviews, comparison articles, independent write-ups, forum discussions — carry more weight in whether an LLM recommends that brand, because the model is trying to synthesize what people who aren't the brand itself are saying. Broader research on generative engine optimization backs this up: one large-scale analysis found earned media supplies roughly 84% of the links AI engines cite, while paid content supplies a fraction of a percent (Wellows' 2026 GEO statistics roundup).
Practical implication: If nobody outside your company is writing about you — no reviews, no comparisons, no third-party mentions — an LLM has very little independent material to cite, no matter how good your own site is. Earning that third-party footprint is now part of the visibility problem, not a separate PR concern.
What this means for how content gets planned
If mention frequency, source authority, and sentiment are the levers, the practical response isn't "write more blog posts." It's: get genuinely useful, specific content in front of the kinds of sources LLMs already trust — review sites, comparison content, trade publications — and make your own site's content structured and specific enough that it's a strong retrieval candidate for the narrow, real questions your buyers actually ask. That's a different content calendar than one built around keyword volume alone, and it connects directly to what's covered in how AI visibility gets measured and generative engine optimization as a discipline.
Frequently asked questions
Do LLMs rank brands the way Google ranks web pages?
No. There's no visible position list. An LLM either mentions a brand in its answer or doesn't, based on what it retrieves and judges relevant and well-supported for that specific question.
Does paid advertising influence which brands ChatGPT recommends?
Based on current research into how citations form, paid content supplies a very small share of what AI engines cite — earned, independent mentions carry far more weight than brand-placed or sponsored content.
Can a smaller brand outrank a bigger competitor in an AI answer?
Yes, if it has stronger, more specific third-party coverage for the exact question being asked. Being the most-searched brand generally doesn't matter as much as being the best-matched, well-documented answer to a narrow, specific version of the buyer's question.
Does the same question always return the same brand recommendations?
Not necessarily. Because retrieval and synthesis aren't a fixed, cached process the way a search ranking is, asking the same or a slightly reworded question can surface a different set of brands.
What role does review sentiment play?
A meaningful one. If a brand is mentioned frequently but with negative sentiment in the sources an LLM draws on, that sentiment appears to factor into whether and how it's recommended, not just whether it's mentioned at all.
Is keyword optimization still useful for AI visibility?
It isn't irrelevant, but it isn't the primary lever either. Structuring content to clearly and specifically answer real buyer questions matters more than keyword density for whether an LLM extracts and cites it.
How can a company find out what LLMs are already saying about it?
Manually running a representative set of real buyer questions through ChatGPT, Gemini, and Perplexity, and noting what's said and which sources are cited, is a practical starting point before investing in dedicated monitoring tools.