Funnel diagnostics: find the stage, not the symptom.

A funnel is a series of gates, so the numbers only make sense in order.

Every rate in a funnel is computed on a population that the stages above it selected. That single fact is why diagnosis has a direction, and why reading out of order reliably costs a cycle.

Take the most common example. You buy cheaper traffic, the registration rate holds up, and the show-up rate collapses. Read bottom-up and the reminder sequence is the obvious culprit — you rewrite it, add a send, and nothing improves, because the sequence never changed. The room changed. The show-up rate was reporting a composition problem in the language of a content problem, and only the number above it could say so.

So the procedure is: read the stages in order, stop at the first one that is out of line with your own previous run, and do the work there. Everything below the first bad number is a symptom until proven otherwise. Once that stage is fixed, read the whole thing again from the top — fixing an upstream stage changes the audience every stage below it receives, so the numbers you took before are no longer the numbers you have.

Out of line with what, exactly

With your own last run of the same funnel. Not with a benchmark. Published conversion figures average across industries, price points, audience temperatures and traffic sources that have nothing to do with yours, and using one as a target produces confident action in an arbitrary direction. If you have no previous run, this pass is not a diagnosis — it is the baseline, and its only job is to be recorded accurately enough to compare against next time.

Seven stages, seven numbers, and what each one cannot tell you.

One metric per stage is a deliberate constraint. Two metrics per stage produces an argument; one produces a decision. The blind spot column matters as much as the metric — most bad diagnoses are a number asked a question it cannot answer.

Two of these deserve a note. Traffic quality is read per tagged source and never pooled, because pooling is exactly the operation that hides it — one weak source drags the overall rate down and the page takes the blame. And refund rate is a lagging check that cannot diagnose the run it belongs to, but it is the only number here that can invalidate a result that otherwise looked like a win.

If your funnel has more stages than these — an application step, a booked call, a multi-session sequence — the method is unchanged. Give each stage one number, define it precisely, and keep the order. Customer journey mapping is the exercise for working out what your stages actually are when the shape is not a standard one.

Check the instruments before you trust the dials.

A surprising share of funnel problems are measurement problems wearing a performance costume: a reminder that never sent, a tag stripped by a redirect, an attendance definition that changed between runs. Establish the readings are real, then read them.

One change per cycle, and it has to be big enough to see.

The rule is not about discipline for its own sake. It is the only way a result tells you anything you can reuse.

Change the headline, the price and the reminder cadence in the same week and the number moves. You have learned that something worked, which is not knowledge — you cannot keep it, cannot apply it to the next funnel, and cannot tell whether one of the three changes was actively hurting and being carried by the other two. The cycle cost you a run and returned a rumour.

The second half of the rule is under-appreciated: the change has to be large enough to produce a difference you could detect at your volume. Swapping a word in a subject line is a real change and an undetectable one for most people reading this. Rewrite the headline as a different promise, cut three form fields, move the offer ten minutes earlier — changes with enough amplitude that if they work, you will be able to tell.

Two exceptions, both legitimate

1. 01Broken, not underperforming. A dead checkout link, a reminder that silently failed, a page that does not render on a phone. Fix every broken thing at once and immediately. You are repairing, not testing, and there is no experiment to protect.
2. 02Replacing a stage rather than tuning it. A new registration page rather than a new headline; a new offer rather than a new price. The whole stage is the one change — just do not also change the stage next to it, or you are back to the rumour.

Write down what you are changing before the run, not after it. Written afterwards, the note reliably records the change you remember rather than the change you made, and it is the after-the-fact note that quietly ruins a change log.

Most of the time, the honest answer is “not enough data”.

This is the part conversion advice skips, because it is the part that tells you to do nothing. At the volumes most funnels actually run at, the majority of stage-to-stage differences between two runs are indistinguishable from noise.

The arithmetic is worth doing once, in your head, so it stops being abstract. It is illustrative — substitute your own numbers. If forty people reach a stage, one person is two and a half points of that stage’s rate. Eight conversions is twenty percent, ten is twenty-five, and the difference between them is two people who could have been having a better Tuesday. Nothing about that comparison licenses a conclusion, and yet “we lifted conversion 25%” is exactly how it gets written down.

The same arithmetic run the other way is the useful part: to detect a small relative improvement you need the number of people reaching that stage to be large enough that one or two of them cannot move the rate meaningfully. Which is why the earliest stages are the only ones most funnels can honestly test early — they have the largest denominators, and every stage downstream has a smaller one than the stage above it.

What to do when the numbers are too thin

• →Pool runs. Three runs of the same funnel, read together, beat one run read alone — as long as nothing structural changed between them. This is the single most effective move available at low volume.
• →Work the largest denominator. Test the stages with the most people in them. It is also usually where the biggest absolute losses are, so the priorities agree.
• →Switch to qualitative evidence. Five replies from people who left, or one recorded session watched end to end, will tell you more than a rate computed on eleven people. Small samples are bad at rates and perfectly good at reasons.
• →Look for broken, not suboptimal. At low volume you cannot detect a ten percent improvement, but you can absolutely detect a checkout that fails on mobile. Spend the cycle hunting failures rather than tuning.
• →Write down “inconclusive”. It is a real finding and it protects the next cycle from building on a coin flip. A change log full of honest inconclusives is worth more than one full of confident stories.

Symptom, stage, likely cause, what to change.

For the post-mortem, once the readings are taken. Find the row that matches the earliest stage that looked wrong — the causes below are the common ones, not the only ones, and they are starting points for a look rather than verdicts.

Notice how many rows point somewhere other than where the money visibly was not. The checkout is the most rebuilt stage in conversion work and one of the least often guilty, because it is the last place you watched a sale fail to happen. Read the stages in order and the checkout gets rebuilt when it deserves it.

Sometimes the funnel is fine and the answer is no.

A funnel can only fail to be the reason somebody bought. It cannot create demand for something people do not want, and a great deal of conversion work is spent refusing to believe that.

The distinction is clean enough to act on. The funnel is at fault when a stage loses people who were on their way to wanting the thing. The offer is at fault when people arrive, understand it, and decline. Four signs point at the offer, and they are worth checking before a third cycle of page rewrites:

1. 01High stay rate, low click rate — repeatedly. They listened to the whole thing and did not move. Once is a bad transition; across several runs and several audiences it is the offer, the price, or who it is for.
2. 02You sell well and refund heavily. The funnel is doing its job too well. Something in the session implies a product the product is not, and the gap is being paid for at the end instead of the beginning.
3. 03The same objection every time. Read the replies. If the recurring sentence is about the thing itself — the price, the scope, the timing, the format — no page change addresses it. If the recurring sentence is about the page, keep working the funnel.
4. 04Upstream wins that never reach revenue. More of the right traffic converting better should raise revenue per registrant. When several genuine upstream improvements leave that number flat, the constraint is downstream of everything a funnel can fix.

When it is the offer, the levers are different: the price and what sits behind it, the scope of what is included, the delivery format, the risk the buyer is being asked to carry, and who the thing is for. Pricing digital products and refund policies and guarantees are the two most under-used, because both feel like concessions and both are usually conversion decisions.

One line per run, or cycle four repeats cycle two.

Everything above compounds only if the runs are comparable, and runs are only comparable if somebody wrote down what was different. This is the least glamorous habit in conversion work and the one that separates four cycles of learning from four cycles of activity.

The log is small on purpose. Per run: the date, the seven readings with their counts as well as their rates, the one thing that changed, anything that changed without being chosen — a new traffic source, a holiday, a price test somebody else was running — and the conclusion, including “inconclusive” when that is the truth.

Record what did not change, too. A stable number across five runs is a finding: it tells you that stage is not your constraint and buys you permission to stop looking at it. Half the value of a change log is the stages it lets you stop worrying about.

Two entries in and the log starts answering questions you did not write it for: which change actually held, whether last quarter’s improvement survived, and whether the thing you are about to try was already tried in March. Without it, funnel work has no memory, and a funnel with no memory gets rebuilt in a circle.