Methodology
Everything here is published on purpose. If you think a number is wrong, the prompts, the alias tables and the raw counts are all open. Check them, then open an issue.
What we run
Six prompt templates, applied identically to all 15 categories so scores are comparable. {cat} is the category noun, {leader} is the best-known incumbent in the category.
best {cat} for a solo foundercheapest {cat} that is actually goodalternatives to {leader}what should I use instead of {leader}? I want to stop payingbest AI-powered {cat} in 2026recommend a {cat}, I have a small budget and no team
Each prompt runs 3 times against each model (ChatGPT) to smooth out nondeterminism. Snapshots are taken monthly and never overwritten; the history is the point.
How mentions are counted
Answers are parsed against a hand-curated alias table per category: exact word-boundary matches only, no fuzzy matching. Product names that are also common English words (“Later”, “Linear”, “Wave”) are matched case-sensitively. If a product is missing an alias, its score is understated: that is a bug in the table, and a one-line pull request fixes it.
The visibility score
A mention is weighted by where it appears in the answer, then averaged over all runs:
| Position in answer | Weight |
|---|---|
| 1 | 1 |
| 2 | 0.85 |
| 3 | 0.7 |
| 4 | 0.6 |
| 5 | 0.5 |
| 6+ | 0.4 |
visibility = 100 × Σ weight(position) / total runs
100 means named first in every run. 0 means the model never says your name.
Honest caveats
- We query model APIs, not the consumer chat apps. Consumer apps add web search and personalization, so what your customer sees can differ from what we measure.
- Model answers are nondeterministic. Multiple runs narrow the error bars; they do not eliminate them. Treat small score gaps as ties.
- Category product lists are curated, not exhaustive. If a product belongs in a category and isn't tracked, open an issue.
- Vendors will try to optimize for this. Good: the prompt set is public, so gaming it is visible in the diff.