Measuring content to registration lag so editors stop judging pieces too early
Content to registration lag is the number of days between a prospect's first editorial page view and their registration. Measure it as a distribution across a cohort, using a survival estimate so people who have not registered yet are handled correctly, then report the median and the ninetieth percentile.
Seven days after a piece goes live, somebody pulls the traffic report. Two thousand sessions, four registrations. The format gets quietly dropped from the commissioning plan and the writer is put back on show previews. Nobody measured the content to registration lag first, so nobody knows whether four was the final number or the first four of forty.
For a show that happens once a year and opens registration five months out, the gap between reading something and filling in a form is measured in weeks. A day seven review is a measurement taken before most of the thing being measured has happened.
What a day seven review is actually measuring
Two things, and neither of them is editorial quality.
It measures whether the piece was delivered. Did it get indexed, did the newsletter actually send, did the social post go out on the day the calendar said. Those are real questions with same-week answers, and a day seven check is the right place for them.
It also measures the small subset of readers who were already most of the way to registering when they arrived. Somebody who has been to the show four times, reads the agenda announcement, and registers that afternoon was going to register anyway. Your day seven number is dominated by them, which is why the pieces that score well at day seven tend to be the ones aimed at people who need the least persuading.
Everything else the piece might do is still in flight.
Building the lag distribution from two columns
You need two dates per person: the first editorial page view attributable to them, and the registration timestamp. If your registration file and your content analytics both carry an email or a stable identifier, that join is an afternoon's work.
Take one edition. Pull everyone who viewed at least one editorial page in the six months before the show and who also appears in the registration file. Suppose that cohort is 4,300 viewers, of whom 1,120 registered. For each of the 1,120, subtract the first view date from the registration date. Sort the results.
A distribution from a real event brand tends to look lopsided. In the shape I would expect, 98 of the 1,120 registered within seven days, which is 8.8 per cent. By thirty days you are at 402, or 35.9 per cent. The median sits at 41 days and the ninetieth percentile at 118.
Those four numbers change how the day seven report reads. Eight point eight per cent means the day seven count captures roughly one registration in eleven. Multiply four by 1,120 divided by 98, which is 11.4, and the piece that produced four registrations in its first week is on course for about 46.
I would not use that multiplier as a forecast. Different pieces have different audiences and the multiplier is an average across all of them. Use it once, in the meeting where somebody proposes killing a format, to establish the order of magnitude of what the day seven number is missing.
Why is the naive lag average always too short?
Because of the people who have not registered yet.
Two groups sit outside the simple calculation. There are readers still inside their decision window, who might register next month. There are readers who left the observable population, by unsubscribing, bouncing, or changing employer. Dropping both from the denominator inflates your registration rate. Counting both as failures deflates it. Neither is correct, and the size of the error grows with how recently the piece was published, which means your newest content always looks worst.
The fix is older than any of this. Kaplan and Meier published the product-limit estimator in the Journal of the American Statistical Association in 1958, for exactly the situation where some observations are incomplete at the moment you stop watching. Their term is censoring, and a reader who has not registered yet is a censored observation.
The arithmetic is small enough to do by hand, and doing it once makes the behaviour obvious. Take ten viewers of one piece, followed for sixty days. Registrations happen on days 12, 19, 33 and 47. Two people leave the list, at day 25 and day 40. Four are still unregistered when you stop counting at day 60.
At day 12 there are ten people at risk and one event, so survival is 1 minus 1 over 10, which is 0.900. At day 19 there are nine at risk, so survival becomes 0.900 times 8 over 9, which is 0.800. Somebody leaves at day 25, dropping the at-risk count to seven without counting as a registration. At day 33, survival is 0.800 times 6 over 7, which is 0.686. Another departure at day 40 takes the at-risk count to five. At day 47, survival is 0.686 times 4 over 5, which is 0.549.
Estimated probability of registering by day 60 is 1 minus 0.549, or 45.1 per cent. The naive count says four out of ten, 40 per cent. The gap is five points on ten people, and it widens as the censored share rises.
Report the median from the survival curve, and quote the share of the cohort still censored at the cut-off next to it. A median of 41 days with 12 per cent censored is a number. A median of 41 days with 60 per cent censored is a guess with a decimal point.
One practical warning about the median. If fewer than half your cohort ever registers, the survival curve never crosses 0.5 and there is no median to report. On the 4,300 viewer cohort above, 1,120 registered, which is 26 per cent, so the median lag exists only among people who registered at all. Say which population you are describing every time, because the median lag of registrants and the median lag of viewers are different quantities and the second one is usually undefined.
What a ninety day lookback window hides
If you plan to read this out of a web analytics tool, check the window before you trust the shape.
Google Analytics 4 sets the lookback window for acquisition key events at thirty days by default, with seven as the alternative, and for all other key events at ninety days by default, with thirty and sixty as the other choices (Google Analytics Help, 2025). Ninety is the ceiling.
Set that against a ninetieth percentile of 118 days. Everything in the last tenth of your distribution is invisible inside the tool, and it is invisible in a specific direction: the slowest readers are the ones furthest from the show, which is usually the audience an editorial programme exists to reach in the first place. The fast registrations show up, get credited to the last email, and the long tail is attributed to direct traffic.
The workaround is to do the join yourself against the registration file, where the only limit is how long you keep the view log. That is also where counting content assisted registrations has to happen, and the two calculations share a table, so build the join once.
How long should an editor actually wait?
Three review points, each answering a different question, and the dates come from your own distribution.
Review at day seven for delivery. Indexed, sent, published where it was supposed to be, no broken canonical. Kill nothing at this review.
Review at the median for reach. If your median lag is 41 days, a check somewhere around day 45 tells you whether the piece found an audience at all. Sessions, scroll depth, newsletter clicks. A piece with 60 sessions at day 45 has a distribution problem and no amount of waiting fixes it.
Review at the ninetieth percentile for registration. On a 118 day tail, that is a check at four months, which feels absurd until you notice it still lands before the next edition's planning cycle. This is the review that carries a commissioning decision.
The uncomfortable consequence is that a piece published eight weeks before show open can never be judged on registration for that edition, because the edition happens before the window closes. Commission those pieces on a different basis and say so at the point of commissioning, so nobody grades them against a rule they cannot pass. That interacts with the balance of evergreen and timely commissions, because the timely half of the calendar lives permanently inside its own lag window.
Where this stops
A lag distribution describes the people who read something and later registered. It contains no claim that the reading caused the registering, and the shape of the curve cannot supply one. A show announcement that lands 41 days before a registration deadline will produce a 41 day median whether it persuaded anybody or not.
If the question in the room is whether the content programme is worth its budget, the lag distribution is the wrong instrument and withholding content from a randomised group is the right one. The lag work still comes first, because it tells you how long the holdout has to run before you read the result.
There is also a matching problem underneath all of this. The join between a content view and a registration depends on identifying the same person in two systems, and everybody who reads while logged out and registers with a different address falls out of the cohort silently. That loss is not random. It skews towards first-time attendees, which is the group whose lag you would most like to know.
Pull one edition's registration file and one edition's editorial view log, join them on email, and sort the day differences. You do not need the survival estimate to start. The median and the count above ninety days, on a single afternoon's join, are enough to tell whoever owns the editorial operation whether the day seven review has been deciding things it cannot see.
Questions people ask about content to registration lag
- How long after publication should an editorial piece be reviewed?
- Long enough that most of the registrations it will ever be associated with have happened. Measure your own lag distribution first. If the median is around six weeks and the ninetieth percentile is around four months, a review at day seven can only see delivery problems, and the review that carries a commissioning decision belongs somewhere past the ninetieth percentile.
- Why does a simple average lag understate the real gap?
- Because everyone who has not registered yet is either dropped from the calculation or counted as a failure, and both choices bias the answer downwards. People still being observed may register next month. A survival estimate keeps them in the denominator until the moment they leave it, which is the correction Kaplan and Meier published in 1958.
- Does a ninety day analytics lookback window cover the whole lag?
- Often it does not. Google Analytics 4 offers thirty, sixty or ninety days for key events other than acquisition, and ninety is the maximum. Any registration whose first editorial view happened more than ninety days earlier is attributed to something else inside that tool, so a long tail has to be measured against the registration file directly.
Related reading
- Counting content assisted registrations without pretending the content caused them
- Running a holdout test for editorial content when attribution will not settle it
- Choosing an evergreen and timely content mix for a seasonal event brand