recurring.fyi

how to read our numbers

Methodology

recurring.fyi is not a peer-reviewed survey. It is a publicly-contributed dataset of weekly meeting hours, collected from verified employees who voluntarily submit. That phrasing matters — both halves of it.

This page tells you exactly how the data is collected, the biases we know are baked in, the structural choices we made to limit them, and how to interpret the published numbers fairly. If you're a journalist, analyst, or curious skeptic, this is the page to read first.

How we collect

  1. A user submits their work email. We HMAC-hash it and never store the raw value.
  2. A 6-digit OTP is sent to that email. Verification proves the submitter controls the address.
  3. The email's domain determines the company. We auto-titlecase a display name; manual review queue cleans up edge cases.
  4. The verified submitter enters two integers for the current ISO week: meeting count and total hours.
  5. Re-submission overwrites; up to 3 edits/week then locked. Schema CHECK constraints cap at 100 meetings and 80 hours.
  6. A company is hidden from public pages until at least 5 of its employees have verified (k=5 anonymity gate).

Known biases

Selection bias (self-selection)

People who feel strongly about their meeting load — usually negatively — are more likely to submit than people who feel neutral or positive. Our sample is therefore not a random draw from the population of any company's employees. Expected effect: our per-company averages probably over-state the true average for that company.

Response bias (exaggeration)

"That felt like a three-hour meeting" → reported as 3 hours when the calendar invite was 90 minutes. Even honest submitters round to gut-feel rather than calendar arithmetic. Schema caps stop the most blatant inflation (no one has 100 meetings totalling 80 hours), but soft-pressure on the upper range remains.

Engagement bias

Filling out a form takes ~60 seconds. People drowning in meetings often don't have those 60 seconds. The most overloaded employees may be invisible to our dataset, pulling the average down — opposite direction to selection bias. The two effects partially cancel, though we don't claim to know by how much.

Sampling bias

Companies on the leaderboard are those with 5+ employees who already know about the site. That's not random. Companies whose first 5 employees come from "people who heard about recurring.fyi on social media" skew toward tech, media, and orgs with reflective cultures. Mining and manufacturing are statistically under-represented.

Structural mitigations

  • k=5 anonymity gate. Below 5 verified employees, no company appears publicly. This forces multiple data points per company before any claim is made, dampening the effect of a single highly-opinionated voice.
  • Schema CHECK constraints. Hard upper bounds (100 meetings/week, 80 hours/week) make the dataset survivable against troll submissions or pranks. These bounds are at the schema level — application validation alone wouldn't be enough.
  • 3-edit-per-week lock. A user can only edit their weekly submission 3 times before that week's row locks. Stops gaming via repeated up-revisions.
  • Research baselines visible inline. Every page that shows a crowdsourced number also shows external research figures from independent sources (HBR, Atlassian, Microsoft, Otter.ai, Doodle) — see /sources. If our crowdsourced figure diverges wildly from all independent estimates, that divergence is the bias signal — and we make it visible by design.
  • Open methodology + open challenge. Every figure on this site is one DM away from being scrutinized. If you find a methodological problem we haven't addressed, tell us; we'll publish corrections under docs/ in the repo.

How to read our numbers

Treat the per-company averages as "what employees who care about their time say" — not as a peer-reviewed estimate of the company's true meeting burden. Specifically:

  • A high number is a signal that at least some employees experience the company that way — not a population mean.
  • A low number indicates absence of complaint, but could also indicate absence of participation. Always check the n.
  • Direct between-company comparisons are only valid when both companies have similar sample sizes and similar opt-in mechanics.
  • Trend over time within one company is more meaningful than a snapshot, because selection bias is roughly constant across weeks.

What journalists should know

Quote our numbers with their n, link this page in your caveat, and you're playing it straight. Don't strip the methodology from the number. We can also export aggregated (non-identifying) data on request to hello@recurring.fyi — normally same-day turnaround.

Methodology version

v0.1 — 2026-05-23

Future versions will add statistical bias correction (e.g. external-baseline-anchored adjustments) once sample sizes are large enough to make them stable. Until then, we'd rather publish the raw skewed-but-honest data than a "corrected" number whose error bars are bigger than the correction.