Methodology
How every number here is made.
A statistic you cannot check is an opinion with digits in it. This page is the standing definition of every platform-level figure we publish — what counts, what is excluded, and the volume below which we refuse to show a number at all. If a figure elsewhere disagrees with this page, this page wins and the figure is wrong.
The gate, in order
Every figure on this platform is computed from readings that passed all five of these. A session is dropped at the first one it fails, and the thresholds below are the same constants the code enforces — not a description of them.
Not a bot
Link unfurlers, mail scanners and crawlers, detected at ingestion and stored with a bot mark so the filtering itself can be audited.
Not the author
You checking your own link. Excluded from platform figures, kept in your own view count, so your numbers still add up.
At least 3 seconds
Opens too short to be a read. A glance is not attention.
At least one section observed
Sessions where nothing was actually on screen long enough to record.
Tracker schema 4 or later
Readings recorded before the current definition, which cannot be compared with today’s.
A session that passes all five is still capped at 30 minutes. It is not discarded — the time above the cap is discarded, because a tab left open overnight is not attention.
What a “reading” is
Every statistic on this platform is computed from qualified readings: sessions on a tracked CV that were not a bot, were not the CV’s own author checking their link, lasted at least 3 seconds, observed at least one section, and were recorded by tracker schema version 4 or later. Sessions longer than 30 minutes are truncated at that cap — a tab left open overnight is not attention.
Link unfurlers, mail scanners and crawlers are detected at ingestion and stored with a bot mark so the filtering itself can be audited; they are excluded from every figure. The author’s own opens are excluded from platform aggregates but kept in their personal view counts, so their own numbers still add up.
What a session has to be, in seconds
Sessions shorter than 3 seconds are not readings. Sessions longer than 30 minutes count, but only their first 30 minutes. The rule is drawn on a compressed scale so both ends fit — it makes no claim about how long readings typically are.
Benchmarks
A benchmark cell — for example “skills sections, in engineering CVs, read on mobile” — is published only when it holds at least 100 qualified readings from at least 25 distinct CVs owned by at least 20 different people, with no single CV contributing more than a tenth of the cell. The reading floor makes the median precise to about two seconds; the people floors are anonymity, not statistics, and are not configurable.
When your exact cell has not met its floor, the comparison steps down an explicit ladder — your role without the device split, then the whole platform — and every surface says which rung it is standing on. Below the last rung the product says “not enough readings yet” and shows nothing.
Aggregates are computed on a schedule from a recent sample, cached, and served from the cache; a request from your browser never touches raw reading rows. Every published figure carries its sample size and the date it was computed.
What a benchmark cell must hold before it is published
A cell publishes only when all three floors are met at once and no single CV supplies more than 10% of it. The bars show the floors, not how full any cell is — that is not known to this page.
Where a comparison stands
- your exact cellsection × role × device
- your rolewithout the device split
- the whole platformevery qualified reading
- “not enough readings yet”and no number at all
A comparison steps down one rung at a time until it finds a cell that has met its floors, and every surface says which rung it is standing on. It never averages two rungs, and it never guesses.
Outcome statistics
Comparisons involving application outcomes — response rates with and without a link opened — exist only above 300 labelled applications with at least 75 on each side of the split, and every published rate carries its sample size; a doubling of a one-in-ten response rate is detected with the conventional 80% power at about two hundred per side, which the figure below draws. Below that the aggregate document itself contains no rates, so there is no underpowered number to leak.
These are correlations inside a self-selected population. A CV that held attention and got interviews proves the two travel together, not which caused which — some of that is the CV, some is the candidate, and no honest analysis can split them. Every outcome statement we publish carries that sentence with it.
How many labelled applications it takes to see a real doubling
Power of a two-sided test at α = 0.05 to tell a 20% response rate from a 10% one, by applications per side. At the 75-per-side floor it is 40%; it reaches 80% at about 199 per side. The 300-application gate is where a rate may be published at all, always with its sample size beside it; the per-side floor stops an imbalanced split from hiding a side too small to mean anything. This is a calculation, not a measurement — you can recompute it.
The attention forecast
The forecast applies published benchmark medians to a draft CV’s structure — section kinds, lengths, order — and shows a range, never a point. It activates only past 10,000 qualified readings platform-wide, names the number of readings behind each prediction, and its calibration (predicted against observed dwell, on held-out readings) is published on this page when the feature is live. Below its gate the editor shows nothing at all.
What the forecast is allowed to show
Before 10,000 qualified readings platform-wide, the editor shows nothing. After it, a range with its sample size — never a single number. The shapes are the form of the output, not a prediction.
What is never used
No reader identities: nothing in any aggregate can say who read a CV, and nothing collected could reconstruct it. Do Not Track and Global Privacy Control are honoured before a session is measured at all. When a user deletes a CV or their account, its readings leave the corpus at the next scheduled aggregation run; figures already published as aggregates are not recomputed retroactively, which is the standard erasure boundary for anonymous statistics — see the data collection page for the deletion cadence.
Where a deleted CV’s readings go
- a sessionno identity is ever recorded; DNT and GPC are honoured first
- the corpusqualified readings only
- scheduled aggregationa recent sample, computed on a timer
- published figureswith sample size and date
Deleting a CV or an account removes its readings from the corpus at the next scheduled run, so the next aggregation is computed without them. Figures already published are not recomputed — the standard erasure boundary for anonymous statistics.