How LLMs decide which brands to recommend
There's no ranking algorithm to reverse-engineer. Recommendations emerge from what the model read, what it retrieves, and how the question is phrased.
Training data: the long memory
A language model's default sense of "the best CRM" or "a great running shoe" comes from its training corpus — years of reviews, comparison articles, forum threads, documentation, and discussions. Brands that were consistently praised across many independent sources get woven into the model's associations; brands that were rarely written about barely register. This is why AI recommendations feel conservative: they reflect accumulated reputation, with a lag. A brand that got great press last quarter may not surface until models retrain.
Retrieval: the short memory
Search-enabled assistants (Perplexity by default, ChatGPT and Gemini in their browsing modes) blend that long memory with live web results. There, the shortlist is heavily shaped by which pages the assistant retrieves — often listicles, review roundups, and comparison posts. Two practical consequences: recency matters much more, and the sources assistants habitually cite become disproportionately important real estate.
Phrasing changes the answer
"Best CRM" and "best CRM for a five-person agency" produce different shortlists. Models are pattern-matchers: qualifiers like budget, team size, industry, and use case activate different neighborhoods of association. That's why serious measurement uses a broad panel of realistic question phrasings rather than one canonical query — and why our prompt panels are built from real search demand across many user situations.
And there's randomness — always
Even with identical wording, models sample their output: the same question asked twice can yield shortlists that differ at the margin. The core names stay stable; positions four through six churn. Any single answer is an anecdote. Trends only become real when they're larger than this sampling noise — which is why every score on helloranked ships with a confidence interval, and why we refuse to report movement smaller than it.
The same question, four different answers
Ask ChatGPT, Claude, Gemini, and Perplexity the same question and you routinely get four different shortlists. Each model has its own training mix, its own retrieval behavior, and its own safety and style preferences — so brand visibility is model-specific, not a single number. Some categories show strong cross-model consensus around one or two leaders; others split sharply, with a brand dominating one assistant and barely registering in another. The LLM Observatory tracks this agreement (and disagreement) per industry, and every industry page — from CRM to laptops — breaks scores out per model for exactly this reason.
What this means for brands
- Your AI visibility is mostly a trailing indicator of your reputation footprint — broad, independent, positive coverage.
- Search-mode visibility is more actionable in the short term: the pages assistants cite can be influenced this quarter.
- Different assistants disagree; measure them separately before optimizing for any of them.
The practical playbook — what plausibly moves these answers and what doesn't — is covered in how to improve your brand's AI visibility; the measurement mechanics live in how the helloranked score works.