How this works

Every model gets the same prompt for each list and answers through a fixed structured format: an ordered list of names. Open lists ask for a top 10; closed lists ask the model to order a fixed set of options. Each model answers each starting list 3 separate times and each visitor request 2 times, at its normal randomness setting. The first answer also gives a one-line reason for each pick. Software buying questions are each asked 4 ways (a list request, a buyer asking for advice, a search query, and whether it's worth paying for), and the brand pages pool every wording so no one phrasing decides a product's standing. Reasoning is turned off for models that allow it and set to its minimum for the rest.

Models write the same thing differently ("New York City", "New York, NY"). A separate model call groups spelling variants of each list into one name, and that grouping is cached and reviewable.

Each model's runs fold into one ballot by Borda count: a pick in position i of an n-item list earns n − i + 1 points. The ballots then fold the same way into the consensus, so every model has exactly one equal vote.

Agreement is rank-biased overlap (p = 0.9): the share of picks two lists have in common, checked at every depth and weighted toward the top. Comparing a model with itself across runs gives the noise floor; comparing different models shows how much of the gap is taste.

Whenever the published starting lists change (a new panel run, a reworded question, or a seat moving to a new model), the site saves a dated snapshot. Brand pages compare each product's share of answers with the previous snapshot, but only on seats that kept the same model, so a model swap never shows up as movement. A change counts as real only when it is larger than two standard errors of the difference; smaller ones are shown in gray as sampling noise.

Anyone signed in with a Google account can request a list from the home page. A moderation model first rewrites each request into a neutral ranking question and declines rankings of protected groups, private people, or harmful topics. Every accepted request is answered by the full panel, 2 times per model, and then published here for everyone. On the brands page you can also check a software product: the moderation model writes the buying question for its kind of product in all 4 wordings, never naming the product, or reuses a question already asked, and the product's page then shows whether the panel names it. A new question uses one free ranking and is published like any other. Each account gets 1 free ranking a day. Signing in stores your Google account's name and email address so the allowance follows you; requests are published without them.

Each lab is represented by a current general-purpose model. For Anthropic, OpenAI, and Alibaba that is their lower-cost tier (Claude Haiku, GPT Luna, and Qwen Flash), which keeps requested rankings free. Results reflect one prompt wording per list (software buying questions pool 4), and a lab's answers can shift when it updates a model.