how to read a funnel drop-off and find what caused it

The short version

• The biggest percentage fall is often not the biggest opportunity. Read the count sitting next to it before you pick a step to care about
• Intent climbs as people move down, so falls should shrink. A fall that gets bigger further down is the highest-signal thing in the whole report
• Half of all wrong drop-off numbers are mid-funnel entrants: people who came in at step three and were counted as if they had walked the whole path
• If one traffic source falls twice as hard as the rest, the page is probably fine and the promise upstream is wrong
• Under a few hundred arrivals a step, the honest answer is that you cannot tell yet, and watching the number daily makes it worse

To read a funnel drop-off you need three things the bar chart will not give you: the size of the fall next to the step above it, the intent of the people who arrived, and the number of visitors it was built from. A signup step losing 60% of the people who reach it is a normal Tuesday in one position and an emergency in another.

Every guide on this subject hands you the same list of reasons people leave. Form friction, a price that surprised them, a slow page, not enough trust. All real, and none of them tells you which one is yours.

So here is the order I work through instead. It is five moves, it works in any tool that can draw a funnel, and two of the moves end in "you cannot tell yet", which is a legitimate answer and one nobody selling an analytics platform will write down for you.

Find the fall, then read the count next to it

Finding the fall takes about four seconds, because it is the shortest bar on the page. Deciding whether it is the fall worth your week takes longer, and the reason is that a percentage and a headcount are two different questions.

Say 12,480 visitors reach /pricing and 30% of them go on to /signup. Say 90 people reach /checkout and 20% of them finish. The checkout step has the uglier percentage by a distance. The pricing step is where 8,700 people went missing.

Percentages tell you how well a step works. Counts tell you what fixing it is worth. Both are on the same screen, and it's remarkable how often people optimise the one with the dramatic number and the tiny population.

So shortlist on volume and diagnose on rate. Rank the steps by how many people are lost between one bar and the next, take the top two, then ask why each of them leaks. A 90% fall on a step that 40 people a month reach is a curiosity. A 25% fall on the step everybody passes through is a business.

How big does a funnel drop-off have to be before it means anything

There is no universal number, and any article that gives you one is selling a benchmark it collected from somebody else's funnel. A fall means something in two contexts only: next to the step above it, and next to how much intent the people arriving had.

Intent is the part people skip. Somebody on your pricing page wants something more specific than somebody who just landed. Somebody halfway through creating an account wants it more than that. The further down, the more each person has already invested, so the more surprising it is when they leave.

The last row is the one to internalise. People who have typed a card number are not casually browsing, so a fall there is not a preference, it is a fault.

When the falls get bigger further down

Because intent climbs, the falls should shrink as you go down. A funnel that leaks harder at the bottom than the top is telling you something specific: people were surprised after they had already decided.

That surprise is nearly always one of four things:

• A cost that appears later than the page which promised the price
• A required field nobody expected, usually a company name or a phone number
• A verification step that hands the person to their email client, which is a door out of your funnel
• An error that only fires for some people, one browser, one card type, one country

All four are cheap to check and all four are invisible in an aggregate number. Which is why the next three moves are about breaking the aggregate up.

Check who actually started at the top

Not everybody enters a funnel at step one. People land on /pricing from a search result, arrive mid-flow from a link a colleague sent them, or come back a week later on their phone and carry on. If those arrivals are counted alongside the people who walked the whole path, your step-one number is inflated and every percentage underneath it is wrong.

This is the most common reason a drop-off number is false, and I have never seen it covered in a guide to funnel analysis.

Two questions are hiding in one report, and they want opposite treatment:

• "How well does my funnel convert?" wants mid-funnel entrants out. You are measuring a path, so only people who took the path belong in it
• "How many people bought this month?" wants them in. You are counting outcomes, and a sale is a sale however the person arrived

Tiny Funnel excludes mid-funnel entrants by default and gives you a toggle to bring them back, because both views are legitimate. Whatever tool you use, find out which one it's doing before you trust a number, and split the first bar by entry page if you can. A first step that turns out to be 40% direct arrivals on /pricing is not a first step, it is two audiences sharing a bar.

One person can also arrive looking like two, which inflates the top of the funnel a different way. Somebody reads your launch email on a phone, then buys on a laptop that evening. Without something stitching those together you have two visitors, one of whom bounced. That is how one visitor's journey is kept in one piece, and it changes the denominator of everything.

Is it the page, or is it the traffic?

A single drop-off number is an average across everybody who arrived, and two groups behaving completely differently produce a perfectly ordinary looking bar. So narrow, one filter at a time, and watch whether the fall moves.

1. Referrer domain. Who sent them, and what did that page promise
2. UTM source. Which campaign, at which level of intent. Worth knowing that utm_source is a label you wrote yourself and the referrer is what the browser reported, so they disagree more often than people expect
3. Entry page. Where they came in, which is frequently not step one
4. Country, if currency, language or a payment method changes anything about the flow

If one source falls twice as hard as the rest, the step is probably fine. The expectation being set upstream is wrong. People arrived believing something the page then contradicted, and the fix is in the ad, the listing or the newsletter, not in the page you were about to redesign.

If every source falls about the same, it is the page. Less comfortable, more actionable.

There is a third answer worth knowing, and it catches people out. Sometimes nothing changed except the mix. If a source that converts at 2% doubles its share of your traffic in a month, your blended rate falls without a single step getting worse. The tell is that each individual source is flat while the total moves. When that happens, the thing to investigate is the campaign, not the funnel.

Check whether you caused it, then read one journey

A drop-off that appeared three weeks ago is a different problem from one that has always been there, and the fastest way to tell them apart is to look at what you changed. If your tool plots edits to your own watched pages on the same axis as conversion, that's a five-second check. A headline change on Tuesday and a fall starting Tuesday is not proof, but it is a very short list of suspects, and it beats scrolling back through deploys trying to remember what shipped.

Then read one person's actual path. Numbers are good at telling you where and genuinely bad at telling you what happened.

One journey proves nothing. Five of them usually make the answer obvious, because people fail in patterns: they hunt for a link that is not there, they open the same page twice, they reach a form and go back to the pricing page. That last one almost always means the price and the plan did not match up in their head.

Five is not an arbitrary number. Jakob Nielsen's case for testing with five users has held up since 2000 for the same reason it works here: the first few people surface most of the problems, and the tenth mostly repeats the third. When you turn click capture on, the trail has more in it, because you can see what somebody reached for as well as which pages they got to.

When the honest answer is "not enough visitors yet"

Small funnels produce numbers that move on their own. If 12 people reached your checkout and four finished, your conversion rate is 33%. If one more had finished it would be 42%. Nothing about your site would have changed.

That is the trap in checking a drop-off every morning. You're watching noise and reacting to it, and each reaction gets credited with whatever the number did next. Evan Miller's warning about stopping a test when the result looks good is about A/B tests, and the arithmetic is identical here: peeking repeatedly at a small sample and acting on the first interesting reading produced false positives up to 26.1% of the time against an assumed 5%.

Rough guidance, and it is rough on purpose. Under about 100 arrivals at a step, treat the rate as decoration and read the journeys instead. Between 100 and a few hundred, a fall has to be dramatic before it earns a week of your time. Past that, differences between traffic sources start to be readable.

Three things to do while you wait:

• Widen the date range rather than the funnel. Three months of a small funnel beats seven days of it
• Watch counts, not rates. "Nine people reached checkout, up from four" is a fact. "Conversion is up 12%" is the same fact wearing a costume
• Make sure you are counting everybody. A sample already thinned by a consent wall makes a small funnel smaller, which is one reason everyone gets counted here: cookie-less and first-party, so there is no accept-or-vanish gate in front of the measurement

Then go and fix the thing you already know is bad. Every funnel has one, and it does not need statistical permission.

Frequently asked questions

How big does a drop-off have to be before it is worth chasing? There is no threshold that travels between funnels. The useful test is whether the fall is bigger than the one above it, since intent should be rising as people move down. After that, rank by how many people are lost rather than by percentage.

What is a good funnel conversion rate? The only benchmark worth having is your own funnel last month. Published benchmarks average funnels with different steps, different traffic and different prices, so a number from one of them cannot tell you whether your /pricing step is healthy.

Should mid-funnel entrants be counted in a conversion rate? Not when you are measuring how the path performs, because they did not take it. Count them when you are counting outcomes, like how many people bought this month. Check which one your tool does by default, since it changes every percentage in the report.

How many visitors do I need before a funnel is worth reading? Roughly a hundred arrivals at the step you care about before the rate is worth an opinion, and a few hundred before you compare traffic sources to each other. Below that, read individual journeys, which stay useful at any volume.

What does it mean if the drop-off is worse at the bottom than at the top? That something surprised people after they had already decided to buy. Look for a cost that appears later than the page which promised it, an unexpected required field, a verification step that sends people to their email, or an error that only some browsers hit.

Reading a funnel drop-off, in five moves

Rank the steps by how many people you lose, not by percentage. Read each fall against the step above it and against the intent of the people arriving. Take the mid-funnel entrants out. Split by source until it stops being an average. Then read five journeys.

That is the whole method, and it is four or five moves rather than an afternoon of building a report first. It works in any tool.

It is faster in one where the funnel is already drawn when you log in, which is the entire reason this one exists. One script tag, then you click through your own flow to record the steps, and the page opens on where visitors came from, where they went and how many bought. Cookie-less, so nobody has been filtered out of the count you are reading.