The Shared Pipeline
Every assistant answer that names a brand is produced by the same five-stage process: the prompt is interpreted, evidence is retrieved, entities are selected, the selection is framed with justification, and sources are assembled into citations. Brands win or lose at specific stages, and different assistants weight those stages differently.
Selection is the decisive stage. In a typical category, the top three recommended brands capture roughly 55 to 65 percent of all recommendation weight; the long tail shares the rest. Concentration is highest in mature categories and lowest in emerging ones, where model families disagree far more.
How ChatGPT Recommends
The GPT family is the broadest recommender: it names more brands per answer than any chat-first competitor, with a measurable tilt toward well-known incumbents and a strong sensitivity to community consensus. Recommendations are frequently justified with community-sourced language ("users report…"), especially in consumer software and electronics.
For challengers, ChatGPT is often the most realistic first surface to win: its broad consideration set leaves room for newcomers with strong community corroboration — but incumbents defend their framing here more than anywhere else.
How Claude Recommends and Cites
The Claude family is the most conservative recommender: fewer named brands per answer, the narrowest consideration set, more hedging on advisory prompts — and the highest stability across repeated sampling. It leans hardest on official documentation and long-form canonical content when grounding claims, citing official sources at the highest rate of any family.
The highest-leverage Claude asset is documentation depth: precise entity facts, thorough product documentation, and long-form evidence. Once inside Claude's narrow set, presence there is the strongest trust signal — but entering it is the hardest.
How Gemini Selects Information
The Gemini family shows the strongest alignment with its surrounding knowledge ecosystem: knowledge-graph presence, entity panels and structured data correlate with recommendation more strongly than for any other family. Brands with inconsistent entity information see their retrieval silently capped here before content quality ever gets evaluated.
Gemini is where entity hygiene pays off most visibly: consistent naming, complete structured data and coherent third-party profiles function as prerequisites rather than bonuses.
How Perplexity Differs
Perplexity behaves like a citation engine: it grounds nearly every answer in explicit sources, cites external material far more frequently than the chat-first assistants, and weights freshness heavily — recently updated pages are cited at a multiple of stale but otherwise comparable pages. Its trust curve across source classes is the flattest, giving well-structured newer sources a fighting chance against incumbents.
The Signatures, Side by Side
| Family | Signature | Highest-leverage asset | Primary risk |
|---|---|---|---|
| ChatGPT | Broad, incumbent-tilted, community-aware | Comparison hubs, community corroboration | Being outframed by better-known names |
| Claude | Narrow, canonical, stable, cautious | Deep official documentation | Never entering the consideration set |
| Gemini | Ecosystem-aligned, structured-data sensitive | Entity and schema completeness | Silent retrieval caps from inconsistency |
| Perplexity | Citation-dense, freshness-weighted | Current, data-rich, crawlable pages | Losing answers to fresher sources |
Behavioural signatures observed on a 50,000-prompt commercial benchmark panel. See the comparative research paper for full methodology.
Model families update frequently; absolute numbers age quickly. The signatures above have proven more durable than any point estimate, but per-model measurement on your own category panel remains essential.