Funnel analytics, done properly

A funnel is a fixed sequence of steps with the drop-off measured between them, and its entire value is in the drop-off. The absolute numbers tell you almost nothing. The step where half your people leave tells you where to spend next week.

Funnels are also the most commonly misbuilt report in analytics, usually in one of four specific ways, all of which produce a number that looks authoritative and means nothing. This page is about building one that does not.

Designing steps that mean something

Three rules and the rest follows.

The steps must be in a genuine order. A funnel is a sequence, not a list of things you care about. If somebody can reasonably do step three before step two, the percentages between them are meaningless arithmetic. This sounds obvious and is the single most common error, usually introduced when somebody adds a step to an existing funnel because it seemed related.

Keep it to four or five steps. Every step multiplies the attrition, so a ten-step funnel ends at a number so small that normal variation swamps the signal. If the journey genuinely has ten stages, build two funnels.

Anchor the first step to something real. Starting a funnel at all traffic to the site means your conversion rate is mostly a measure of how much unqualified traffic you have, which moves for reasons that have nothing to do with the funnel. Start it at the first action that indicates intent, such as reaching the pricing page or viewing a product.

The time window, which is where most funnels lie

A funnel needs a period long enough for people to complete it, and this is where honest-looking funnels most often mislead.

Run a seven-day funnel on a product with a thirty-day evaluation cycle and you will see a catastrophic drop-off at the final step that is not happening. The people are still deciding. They convert in week three and land outside your window, so they are counted as lost.

The symptom is a funnel whose last step improves every time you widen the window. If that happens, the window was too short, not the page.

The opposite error is a window so wide that unrelated behaviour gets stitched together, and a visitor who came back six weeks later for a different reason is counted as having progressed. Match the window to the real decision cycle, write down what you chose and why, and stop moving it to make a chart look better.

The four things that break a funnel

Counting the same person twice. A thank-you page that people refresh, or an event that fires on load on a page reachable by the back button, inflates every step downstream. Convert yourself and then refresh. It takes two minutes and catches it.

Consent loss, which compounds. This is the one people miss. If your analytics needs a banner, every step of the funnel is drawn from the subset who accepted, and the visitors missing from the top are missing from the bottom too. The *rate* may be roughly salvageable; the volumes are not, and any comparison with a period before the banner design changed is worthless.

Bots in the first step. Automated traffic enters the top of the funnel and never reaches the second step, so your first drop-off is inflated by whatever share of arrivals are crawlers. On a site where that is a third of traffic, your headline conversion rate is understated by roughly a third and it moves whenever a new scraper appears.

Reading a funnel without segmenting it. An aggregate funnel averages two populations that behave completely differently: people arriving from a branded search who already intended to buy, and people arriving from a listicle who did not. Segment by source at minimum, and the useful finding is usually in one segment rather than the average.

The funnel tells you where, not why

This is the limit of the whole technique and it is worth being blunt about, because it is where most funnel projects stall.

A twenty per cent drop between checkout started and payment entered is a fact. Whether it is a shipping cost revealed too late, a form field that rejects valid postcodes, a payment method people expected and did not find, or a button that does not look like a button, is a question a funnel cannot answer and never will.

Answering it needs behaviour tooling: a scroll map showing nobody reaches the reassurance copy, a click map showing people repeatedly clicking a heading they think is a link, a recording of somebody failing to complete a field. Teams that own both a funnel and a heatmap on the same step close this loop in an afternoon. Teams that own only the funnel produce a quarterly report identifying the same drop-off every time.

The practical consequence for buying: check whether your funnel tool can also show you the page, or whether answering the obvious follow-up question means a second subscription and, usually, a consent banner.

Who has funnels, and what they cost

Funnels are almost always a paid-tier feature, and the pricing varies more than the capability.

Matomo includes them at every level, including the free self-hosted build, which makes it the cheapest real funnel on this list if you can operate a PHP and MySQL application. Pirsch puts them on its $12 Plus tier, which is the cheapest hosted funnel here. Plausible and Absolutely Analytics both put them on a Business tier at £19 and $19 respectively. Fathom does not offer funnels at any price, which is worth knowing before you buy it for that purpose.

Every product analytics platform has them and goes further, adding open path exploration, which shows the routes people actually took rather than the one you predicted. That distinction matters: a funnel tests a hypothesis, a path exploration generates one. If you keep discovering your funnel was the wrong shape, you want the second kind of report.

The rest of the field

Common questions

Related guides