Which AI Assistant Has the Broadest Local Shortlist?

First place hides the size of the field

Two assistants can agree on the leading provider while constructing completely different markets underneath it. One may repeat five names; another may surface dozens. Rank alone cannot show whether a buyer is seeing a stable shortlist or a broad discovery set.

Viclaro calls this recommendation diversity: how widely an assistant distributes named-provider appearances across a defined prompt corpus. The simplest observable is the number of distinct resolved firms. Better measures also account for response volume and how evenly recommendations are distributed.

Cosmetic surgery shows the widest split

In the reconciled NYC cosmetic-surgery corpus, Perplexity named 296 distinct resolved firms across 779 extracted citations. Claude named 177 across 572, Gemini 75 across 366, and GPT-4o 69 across 286.

Perplexity’s leading practice held 4.2% of its citation pool. Gemini’s leader held 27.9%, GPT-4o’s 18.5%, and Claude’s 16.4%. Perplexity therefore produced both the broadest observed extraction field and the least concentrated model-level leader in that category.

That does not prove Perplexity is universally more diverse. It describes one category, prompt set, and snapshot. It does show how much information disappears when four providers are reduced to an aggregate rank.

The pattern recurs in cosmetic dentistry

Perplexity named 194 cosmetic-dentistry firms, compared with 111 for Claude, 86 for Gemini, and 56 for GPT-4o. Its leader received 5.9% of Perplexity citations. The leaders on GPT-4o, Gemini, and Claude received 13.8%, 24.6%, and 16.8% of their respective pools.

In divorce, GPT-4o named the widest field at 81 firms, followed by Perplexity at 73, Claude at 71, and Gemini at 19. The breadth ordering is therefore category-dependent; it should not be turned into a permanent personality claim about any model.

Why distinct-firm counts are only a first pass

A model that produces more names per response will naturally create a larger firm pool. A model with fewer usable outputs may look concentrated because its denominator is thin. Prompt refusals, non-commercial advice, retrieval availability, and extraction quality also affect the observed field.

A defensible diversity comparison should report distinct firms, total recommendations, recommendations per response, leading-firm share, and an evenness measure such as normalized entropy. It should exclude or flag cells too small to support comparison.

For that reason, Atlas uses diversity as a diagnostic rather than a quality score. A broad shortlist may improve discovery, introduce noisy entities, or both. A compact shortlist may reflect strong consensus or an overly repetitive answer set.

What recommendation breadth means for a firm

A wide model-level field offers more possible entry points but less share per appearance. A narrow field makes inclusion more valuable and displacement harder. Firms should compare their position with the assistant’s shortlist shape before choosing a target.

Measure diversity beside the consensus gap. Breadth describes the size and distribution of the field; consensus describes whether the same firm travels across assistants. The NYC Index supplies the category context around both.