How to Segment a Funnel Before You Optimise Anything
How to segment a funnel: a blended rate averages audiences that behave differently, so split by traffic source first, then device, before you change a thing.
To segment a funnel, split the same steps by the groups of people who pass through them, starting with traffic source and then device, before you optimise anything. A single overall conversion rate is an average across audiences that behave completely differently, and it hides the two findings that matter most.
The blended number is the one everybody quotes because it is the easiest to produce. It is also the one least likely to tell you what to do. This post covers what a blended funnel conceals, which cuts to make first, and how to check that a split is large enough to mean something.
How to segment a funnel: what the average hides
Imagine a single conversion rate for a funnel. It looks like a fact about your product. In truth it is the combined result of many groups of people, each arriving for a different reason, in a different way, with a different level of readiness. Averaging them produces a number that describes none of them.
Two findings disappear inside the average, and they are the two that change decisions.
| What is hidden | What it tells you | What to do with it |
|---|---|---|
| A segment that converts far better than the rest | Who to go and get more of | Direct more effort towards people like them |
| A segment that converts terribly | What may be dragging your whole number down | Often the cheapest thing to fix, or to stop paying for |
Both findings are usually cheap to act on once they are visible. Neither can be seen while everyone is in one pile. That is why the guide says to segment before you optimise: otherwise you may spend a month improving a step for an audience that was never going to convert, or leave a strong audience undiscovered.
Make the first cut by traffic source
The first cut is usually by traffic source. Where people came from is often the biggest difference between them. Someone who arrived from a source that matches your offer is in a different position from someone who arrived by accident, and the funnel will show it if you let it.
This cut pairs naturally with the work of attribution. Attribution tells you which channel produced a customer. The funnel tells you what happens to the people that channel sends. A channel with poor conversion may be sending the wrong people, not be a bad channel, and you cannot distinguish those without both views. Spending decisions made on channel volume alone reliably fund whichever source produces the most people who were never going to buy.
So when you split by source, you are asking a pointed question of each one: of the people you send, how many get to the end? The answer often changes what you would have done.
Make the second cut by device
The second cut is usually by device. People on different devices meet different versions of the same steps. A step that works well for one group may be awkward for another, and a step that never fires properly on one kind of device will look, in the blended number, like a modest overall weakness instead of a total failure for that group.
It is a cut that serves two purposes at once. It shows you where the experience is worse for one group, and it flags where the measurement itself may be broken. Both are worth knowing, and a device split is a quick way to find out which you have. If the drop appears for one device only, check whether the tracking works there before concluding that the people on it are different.
Other cuts worth knowing about
Source and device are the usual first two. The guide also lists other differences that can matter, and any of them may become your third or fourth cut once the first two have taught you something.
- Customer size. Larger and smaller customers may behave very differently along the same path.
- New versus returning. A person seeing you for the first time is in a different position from one who has been before.
- Different traffic sources. Beyond a first split, you may want to look at individual sources separately.
You do not need to cut by everything. Choose the groups most likely to behave differently in a way you could act on. Each cut multiplies the number of groups you have to read, and a long list of cuts quickly becomes a source of confusion instead of insight.
Check that each segment has enough volume
A split is only useful if each group is large enough for the numbers to mean something. Every cut divides your volume, and a small group produces percentages that jump around from week to week without telling you anything.
The same point applies to funnels in general: more steps and more splits both reduce the size of each group. So ask, before you read a segment, whether there are enough people in it for the result to be worth trusting. If not, the honest conclusion is that you cannot yet read that cut. Combine groups, wait for more data, or make larger changes whose effect would be obvious.
| Situation | What it means | Sensible response |
|---|---|---|
| Segments large enough to compare | The differences may be real | Read them, and look for the best and worst |
| A segment too small to read | Its percentages are noise | Merge it with a similar group or wait |
| Every segment looks the same | The blended number may be an honest summary | Move on to the step itself |
Is it the step or the audience?
Segmenting also protects you from fixing the wrong thing. When you have found a step that loses many people, how to find where a funnel loses people covers how to rank it. Before you commit to a fix, split that step by your chosen cuts and look at who is leaving.
If the loss is spread evenly, the step itself is the likelier problem. If it is concentrated in one audience, the step may be fine and the trouble lies in who you are sending there. In that case the response is to change the targeting or stop paying for that traffic, not to redesign the step. That connects directly to the three causes of drop off. A loss that belongs to one source often points towards mismatch, which is fixed upstream, and how to fix funnel drop off explains why working on the step cannot repair it.
Decide your segments before you start
It is tempting to slice the data until something interesting appears. That is a good way to find patterns that are not there. Instead, decide in advance which segments you will look at first, and write them down with the funnel definition. Traffic source is one of them, and the guide's checklist asks you to know what your blended rate hides before you start quoting it.
Setting the groups in advance has another benefit. It lets you compare month to month on the same basis. If the cuts change every time, you will never know whether a difference came from the funnel or from the choice of cut.
- Choose the segments you will look at first, with traffic source among them.
- Check that each has enough volume to mean something.
- Note what the blended rate hides before you quote it.
- Compare the best and worst segments, and ask why they differ.
- Decide whether the finding points to the step, the audience or the measurement.
What to write down after each split
A split that is not recorded is soon forgotten. After each one, write down three things in plain sentences: which segment looked best, which looked worst, and what you think explains the gap. Keep the sentences short and dated.
The best segment tells you who to go and get more of, so name it. The worst segment tells you where the number is being dragged down, so name that too, and note whether it is cheaper to fix or to stop paying for. Then record the one decision the split changed. If a split changes no decision, it was interesting and not useful, and you can drop that cut from your routine without regret.
Where segmentation fits in the whole sequence
Segmenting comes after the funnel is defined and counted correctly, and before you change anything. If the steps are not sequential, or the counts are not cohort-based, then a segment-level view will only break the same errors into smaller pieces. The foundations are covered in how to define funnel steps.
It also belongs in your regular rhythm. Each time you review the funnel, look at the segments again. A segment that was healthy last month may have changed, and a fix that helped one audience may have done nothing for another.
Frequently asked questions
Which segment should I look at first?
Traffic source. It is usually the biggest difference between people and it pairs well with channel-level decisions. Device is usually the second.
Why not just use the overall conversion rate?
Because it averages groups that behave differently. It hides the segment worth doubling down on and the one dragging the number down.
What if my segments are too small?
Then you cannot yet read them. Combine similar groups, wait for more data, or make larger changes whose effect would be obvious.
Can a segment tell me the step is fine?
Yes. If the loss belongs to one audience and the others pass through the step normally, the trouble is likely who you send, not the step.
The full method, with the checklists, is in The Funnel Playbook, a free 30-page playbook.