Estimate Conversion Rate and the Cohort Problem

Why this matters

A shop that closes half the work it quotes can read a conversion rate of 15 percent every month for a year and never find out why. Nothing is broken in the selling. The figure is built so that it cannot show the truth, and the owner who acts on it retrains a good estimator, drops prices on work that was closing fine, or fires a salesperson. This card owns the mechanism for the group, because several other rates in the quote-to-cash chain are built the same way.

What the figure is actually made of

The usual construction: of the estimates created inside the window that got past draft, the share currently sitting approved.

Both halves are keyed to the date the estimate was raised. Neither is keyed to the date the customer decided. That single design choice is the whole problem, because raising an estimate and deciding on it are separated by days or weeks, and the window has a hard right edge.

An estimate raised three days before the window closes is counted in full as an opportunity. It has had three days to be approved. If your customers typically take three weeks, that estimate is in the denominator and cannot reach the numerator, and nothing about it is unusual.

The denominator gets time the numerator does not

Strip it to the mechanism with a hard decision cycle, which exaggerates the effect but isolates it cleanly. Say every estimate that will ever approve approves exactly N days after it is sent, and half of all sent estimates eventually approve.

Read a window of W days on the day it closes. An estimate sent on day d matures on day d plus N, so only estimates sent on or before day W minus N have had their chance. That is a matured share of (W minus N) over W. Everything sent after that is in the denominator with the outcome still ahead of it.

So the reported rate is the true rate multiplied by the matured share, and the understatement depends only on the cycle as a fraction of the window, not on either number alone. A 7-day cycle read on 30 days and a 21-day cycle read on 90 days land in exactly the same place, because 7 over 30 and 21 over 90 are the same fraction.

Two windows, two cycles

Same shop, true rate 50 percent of sent estimates, 60 estimates sent per month at 2 a day.

Work the worst cell all the way. On a 30-day window with a 21-day cycle, 18 of the 60 estimates sent that month have matured (9 days at 2 a day). Half of 18 is 9 approvals. The denominator is all 60, because all 60 got past draft. The rate reads 9 over 60, or 15.0 percent, against a true 50 percent.

The shop reading a rolling 30-day window sees a stable 15 percent every single month. Stability is what makes it convincing. There is no jump to investigate, no bad month to explain, just a flat number saying the shop loses roughly six of every seven quotes it writes, and it is wrong by a factor of more than three.

Notice also that no window removes the problem. Even 90 days against a 7-day cycle still reads 46.1 percent against 50. Lengthening the window shrinks the error and cannot delete it, because there is always an unmatured slice at the right edge.

Relax the hard cycle and the numbers soften. Real decisions spread across a range, so some estimates raised late in the window do approve inside it. The direction never changes: every estimate raised in the window sits in the denominator, and only the ones that have had time can be in the numerator.

Age the cohort, not the calendar

The fix is not a longer window. It is to stop reading a window that is still open.

Measure the estimates raised in a window that closed at least one full decision cycle ago, and state the allowance you used. With a 21-day cycle, on 1 May you read the cohort of estimates raised in March. That window closed on 31 March, so the youngest estimate in it has had 31 days, comfortably past 21. The March cohort reports 50 percent, because every estimate in it has had its chance.

Write the figure with its cohort attached, every time: "March cohort, read 1 May, 21-day cycle allowance, 60 sent, 30 approved, 50.0 percent." A conversion rate with no cohort label is not a rate, it is a snapshot of an unfinished race.

Two constraints that come with it:

• You are always reading old news. With a 21-day cycle you find out about March in May. That is the honest cost of a rate that means something, and it is the reason this figure is a quarterly review number rather than a weekly one. If you need a weekly signal on selling, count estimates sent and approvals received as raw counts, which have no maturity problem.
• The cohort needs enough records. Below about 30 sent estimates, roll windows together until you clear it, and say how many you rolled. A rate read off a dozen estimates moves several points when one customer changes their mind.

Finding your own cycle length

The cycle is not a guess and it is not the industry's number, it is in your records. Pull the elapsed days from estimate raised to job created on the estimates that converted, and read the distribution.

The allowance you need is the tail, not the middle. A sibling card teaches reading that distribution by its median and its shape, which is the right read for setting follow-up cadence. For a cohort allowance you want the point past which conversions have essentially stopped, so take roughly the 90th percentile of the conversion lag and round up to a clean boundary. A shop whose 90th percentile lands somewhere between 60 and 120 days does not have a 21-day cycle, and it needs a quarterly cohort read two quarters back.

If your two figures disagree, use the longer one and say so. Under-allowing reintroduces exactly the truncation you are trying to remove.

The estimate nobody ever marked

The quieter distortion sits in the same denominator. An estimate the customer silently declined, and nobody ever marked lost, stays in the population forever. It is past draft, it is not approved, and it will sit in every cohort read that includes its creation month for the life of the shop.

On the construction above it is not arithmetically wrong, since a dead estimate genuinely was not won. What it destroys is your ability to tell still deciding from already dead, and that is the input the cohort allowance depends on. A pile of unmarked estimates makes the lag distribution look longer than it is, which makes your allowance too long, which pushes the whole review further into the past for no benefit.

It also quietly poisons the pipeline figure, where an unmarked estimate keeps counting at full value indefinitely.

The practical rule: an estimate that has passed your own cycle allowance with no decision gets one contact that asks the closing question plainly, and then it is marked won or lost with a reason. Not "pending". Pending is not an outcome, and a status that never resolves is a status that carries no information.

What an aged rate is allowed to decide

Once the figure is honest it still only answers a narrow question, and it is worth naming the boundary before somebody stretches it.

It can decide whether a job type is worth quoting at all. Run the aged cohort per job type and a type sitting far below the others is either priced wrong for its market or being quoted to the wrong customers, and either answer is actionable. It can also confirm whether a price change worked, because a change made in one month shows up in that month's cohort and not in the ones before it, which is a clean before-and-after the blended figure cannot give you.

It cannot grade an estimator. Two estimators handling different job types carry different rates by construction, and the shop's mix decides who looks good. If you want to compare people, compare them inside one job type, on one aged cohort, with both counts printed, and accept that most small shops never have enough estimates in a single type in a single quarter to make that comparison honestly.

It also cannot stand in for booked value. A rate is a count share of counts. Thirty small approvals and three large ones can produce the same rate and very different quarters, which is why the rate belongs next to the value that actually got booked rather than on its own.

Two ways to compute it, erring in opposite directions

Most shops use one of two constructions, and they fail in mirror-image ways. Run both on the same aged March cohort of 60 sent estimates: 30 approved, 22 marked lost, 8 never resolved.

• Approved over all sent. 30 over 60, or 50.0 percent. The 8 unresolved sit in the denominator as losses. This construction is the one truncation attacks, and it errs low whenever the cohort is young or estimates go unmarked.
• Approved over decided only. 30 over 52, or 57.7 percent. The 8 unresolved are dropped. This construction is immune to truncation, and it errs high by exactly as much as your marking discipline is bad, because every unrecorded loss leaves the denominator while every win is always recorded.

The two figures differ only because of the unresolved estimates, so the gap between them, 7.7 points here, tells you how much of the cohort has not finished deciding. A shop with clean marking and a properly aged cohort sees the two figures converge, which is the only reliable signal that the number is readable at all. A shop whose two figures sit 20 points apart does not have a selling problem to diagnose yet, it has a records problem, and no amount of sales training moves either figure until that is fixed.

Pick one construction, name it beside the number, and never let a review compare a rate computed one way against a rate computed the other.

References

• See related: Days From Estimate to Job and What a Long Tail Means - where your cycle length comes from and how to read the tail
• See related: Average Estimate Value Counts Drafts and Conversion Does Not - the figure most often printed beside this one, over a different population
• See related: Pipeline Value Is Not a Forecast Until You Weight It - stage weights derived from an aged cohort rather than from the open list
• See related: The Close Rate Improved Because They Stopped Quoting - what happens when this denominator moves for reasons that have nothing to do with selling
• See related: Lead Conversion Rate Excludes the Leads You Never Touched - the same rate one step earlier in the funnel, with a different exclusion