Method v2026-11-17.1
How ranking works
Every listing gets a SharkScore from 0 to 100, built only from public evidence. The method is the same for everyone and published here in full.
Money never changes the rank
- Fast-track pays for a faster review by a person. Same rules, same score, and a small “paid review” note on the listing.
- Sponsored slots are ads. They sit in their own labelled box, never carry a rank number, and never appear in our API or MCP results.
- The score has no input for payment — not as a factor, not as a tie-breaker.
The five parts
- Adoption30%
- Is it really used? Monthly web visits, GitHub stars, package and model downloads — whatever exists for the kind of listing.
- User reviews25%
- Verified reviews, averaged with a pull toward the category average so a handful of 5-star reviews can't beat hundreds of 4-star ones.
- Maintenance20%
- Is it alive? How recently it shipped something, and whether it answered our uptime checks over the last 30 days.
- Trust & transparency15%
- Public pricing, a free tier or trial, open code or docs, a maker who proved ownership, and an API or MCP server.
- Momentum10%
- Growing or fading? Growth in GitHub stars over the last 90 days compared with before.
The maths, in plain words
- Each listing is compared only with listings in its own category — a small MCP server is never compared with ChatGPT. Once a category has 8 or more listings with a signal, we use percentiles (“better than 82% of MCP servers”); before that, fixed scales.
- Reviews use a Bayesian average: a listing needs about 10 reviews before its own average outweighs the category’s.
- A part with no evidence doesn’t count as zero; its weight is shared out. But thin evidence pulls the score toward the middle (50), so a listing we know little about can’t jump to the top.
- A listing needs evidence in at least 2 parts covering 40% of the weight to get a rank number. Until then it shows “Not ranked yet”.
- Ties are broken by number of reviews, then by name. Never by payment.
- Scores are recomputed daily and the history is kept. Makers can’t remove past scores.
Worked example: Cursor
Score 82.8 with evidence covering 90% of the weight.
- Adoption
- 71 · 30%
- User reviews
- 90 · 25%
- Maintenance
- 95 · 20%
- Trust & transparency
- 100 · 15%
- Momentum
- no data · 10%
How we check reviews
Only signed-in people can review, once per product. Reviews are labelled Verified user, Incentivised or Vendor-invited. Reviews from a maker’s staff or partners are shown but never counted. Positive and negative reviews are moderated by the same rules, and the default order is most recent — never “best first”. When we remove fake reviews, scores are recalculated the same day. The full review policy is published before reviews open.
Where the evidence comes from
GitHub (stars, releases), npm and PyPI (downloads), Hugging Face (downloads), OpenRouter (model usage, CC BY 4.0), Tranco (web traffic rank), and our own uptime and MCP health checks. Each input shows its date on the listing. We don’t publish the details of our fraud checks, so they can’t be gamed.
Method changelog
- 2026-11-17.1 — first published method: five parts, category-relative, Bayesian reviews, coverage shrink.