Benchmarks ask made-up questions. We didn't make ours up.
Verso sees the same prompt answered by two models, and it sees which answer the person kept. Not annotators. Not synthesised prompt sets. People doing their actual work, choosing.
What the dataset is
Paired comparisons
One prompt, two models, one revealed preference. The pairing removes prompt variance, which is why a modest number of prompts still resolves a difference between models.
Brand mentions in answers
Named brands and products as they appear in model responses, not in user prompts. This is the unit GEO platforms buy, and it is far denser than purchase intent.
Real query taxonomy
How people actually phrase a question — the formulations, the follow-ups, the distribution. Usable as a seed layer for anyone generating prompt sets today.
Longitudinal drift
A fixed prompt panel, replayed as model versions change. Stability is the point: it measures what moved, and when.
Coverage
Distinct prompts per cell. Cells with enough distinct prompts support rates; thinner cells support model-vs-model deltas only. We say which is which.
| Vertical | Language | Distinct prompts | Supports |
|---|---|---|---|
| General advice | EN | TODO | TODO |
| Code & engineering | EN | TODO | TODO |
| Product & brand choice | EN | TODO | TODO |
| Product & brand choice | FR | TODO | TODO |
| Travel | FR / ES / DE | TODO | TODO |
Sponsored panels
If your vertical is thin in the table above, we don't pretend otherwise. We recruit for it. Consenting users, your category, a few weeks — funded by the engagement. You get depth where you need it instead of breadth you can't use.
What we will not sell you
Raw conversations. Not as a policy position — as a design constraint. The panel is European, the data is personal, and a dataset that can't survive scrutiny isn't an asset. What leaves Verso is derived signal: preferences, mentions, distributions.