How ConversionBench computes percentiles
Every benchmark on ConversionBench is one of three things: a verified reading from a primary source, a derived interpolation that says so on the page, or research-pending that 404s rather than fabricates a number. This page documents the rules. Who applies them is a separate page.
1. Source selection
Each metric anchors to a primary source. The shortlist: Baymard Institute, WordStream, Adobe Digital Insights, Mixpanel, Contentsquare, Littledata, HubSpot, and VWO.
Each source gets a credibility score from 1 to 10 based on three things: sample size (a 487-site Baymard panel scores higher than a 50-customer competitor blog), methodology transparency (the report shows the formula and excludes outliers), and recency (a 2024 reading scores higher than a 2019 one). Sources scoring 6 or below are not cited as primary.
2. Row state
Every benchmark row in our dataset is in exactly one of three states:
- Verified — a direct reading from a credibility-7+ primary source. Cited inline. Fully indexable. The 25/50/75 numbers come from the source verbatim or with disclosed adjustment.
- Derived— interpolated from the primary source plus adjacent verticals. Disclosed on the page with a “derivation notes” section. Used when a source publishes a category average but not a per-vertical breakdown we’re confident reproducing.
- Research pending — no credibility-7+ source yet. The page returns 404. We never publish a placeholder number to fill the matrix.
3. Percentile interpolation
The calculator uses piecewise linear interpolation between p25, p50, and p75 anchor points, with diminishing returns above the 75th. For lower-is-better metrics (cart abandonment, churn, bounce rate), the curve flips so that low rates rank highest.
Concretely: if your rate sits between p25 and p50, your percentile is linear on that segment. Between p50 and p75 it’s linear again. Above p75 the slope flattens — an industry top performer doesn’t get rewarded with unbounded ranking for being twice the median. The math lives in lib/calc/percentile.ts and is covered by an API endpoint any tool can hit.
4. Refresh cadence
Source data refreshes once per year, anchored to when primary publishers ship their annual reports (typically March-May for the prior calendar year). Every benchmark page shows both the source-data year (when the underlying study published) and a last verified stamp (when ConversionBench last re-checked the row against the source).
Currently anchored to: source data 2024 · last verified .
5. What we don’t do
- No fabricated numbers. The matrix has gaps. We show them as gaps.
- No competitor-blog citations. Other people’s benchmark roundups are not primary sources, even when they’re well-written.
- No paid placement. Sources are cited because their data is good. Nobody pays to be on this site, and we don’t take affiliate commissions on tools mentioned in benchmarks.
- No newsletter, no email capture, no popups, no “subscribe to get the methodology”. The methodology is this page.
Primary sources
Full source list with credibility scores at /source.
Spot a problem?
If a benchmark looks wrong, a source has been superseded, or a derivation should be reconsidered, email [email protected]. Methodology corrections are append-only — we log what changed and why.