Conversion funnel analysis: finding the step that actually broke

Conversion funnel analysis is how you go from "conversion is down" to "the mobile checkout step broke on the 9th". The headline rate tells you something is wrong. The funnel tells you where, and only once you know where can anyone fix it.

This post works through one example end to end: an online store spending $120K a month on ads, whose conversion rate dropped from 3.2% to 2.7%.

The headline problem

Here is the month as it first appears in the weekly review:

Ad spend held at $120K. Traffic held at 84,000 visits. Conversion fell from 3.2% to 2.7%. With an average order of $54, revenue fell from $145,152 to $122,472, a loss of $22,680.

Return on ad spend went from 1.21 to 1.02, and cost per purchase went from $44.64 to $52.91. Those are the numbers that make a marketing lead nervous, and the first instinct in the room is usually to blame the ads.

But traffic didn't move. The same $120K bought the same 84,000 visits. Whatever broke, it broke after people arrived.

This is the logic of a metric tree: revenue is traffic times conversion times AOV, two of the three held, so all of the drop is in conversion. That narrows things down. It doesn't tell you which part of conversion.

Step 1: rule out the measurement

Before diagnosing the funnel, check the ruler. A surprising share of conversion drops turn out to be tracking changes rather than anything customers did.

Three quick checks:

1. Compare purchases in the order system with the analytics tool. If your store's backend shows 2,268 orders and analytics shows 2,268, the drop is real. If the backend shows 2,688, your tracking broke.
2. Look for changes to tags or consent banners. A new cookie banner that defaults to "reject" can remove a chunk of tracked sessions or events overnight.
3. Check for duplicate or missing events. A purchase event that fired twice last month and once this month will look like a 50% drop that never happened.

In our example, the backend agrees with analytics: 2,268 orders this month, 2,688 last month. The drop is real.

Step 2: lay out every step of the funnel

Conversion rate is one number summarising four steps. Break it out:

The step rate is each step divided by the one before it, which is the number that matters. The overall rate (purchases divided by visits) blends all four together and hides where the change is.

Here, three steps are identical to the decimal. Checkout completion fell from 53.3% to 45.0%. All 420 lost purchases happened between "started checkout" and "purchased".

One table, and the search area has shrunk from the whole site to one page.

Step 3: check it isn't noise

Funnels wobble from month to month. Before raising an alarm, check that the change is bigger than normal variation.

A rough test that works for most funnel steps: the typical random wobble in a rate is about the square root of p × (1 − p) ÷ n, where p is the rate and n is the number of people at that step. Double it for a comfortable range.

For checkout completion last month: p is 0.533 and n is 5,040. The square root of (0.533 × 0.467 ÷ 5,040) is about 0.7 percentage points, so the normal range is roughly 53.3% plus or minus 1.4 points.

This month came in at 45.0%, more than eight points lower. That is far outside the normal range. Something changed.

Step 4: segment the broken step

Now split the broken step by the dimensions most likely to explain it. Device is usually first, then traffic source, then country or payment method.

Desktop didn't move at all. Mobile checkout completion fell from 54.0% to 40.1%, and the 420 lost purchases are all mobile.

That changes the conversation. This is no longer a marketing question or a pricing question. It is a mobile checkout question, and there's a short list of people who can answer it.

Step 5: line it up with what changed

With the step and segment identified, look at what changed in the same window: deploys, payment provider updates, new form fields, price display changes, a third-party script added to the checkout page.

In this example, the daily mobile checkout rate drops sharply on the 9th and stays down. The release log shows a checkout update that day, which moved the express payment buttons below the fold on smaller screens. People who used to tap one button now had to scroll and fill in a card form, and many didn't.

That is the root cause. The funnel analysis didn't find it on its own, but it reduced the search from "something on the site" to "mobile checkout, from the 9th", which takes an engineer about ten minutes to check.

Mistakes that send you the wrong way

Blaming traffic quality first. Sometimes conversion drops because the traffic changed: a new campaign brings in colder visitors, and every step converts worse. The tell is that early steps (product views, add to cart) fall too. In our example they held exactly, so traffic quality wasn't the cause. Always check the early steps before blaming the ads.

Reading only the overall rate. Overall conversion fell 0.5 points. That sounds small. The broken step fell more than eight points. Overall rates make big local problems look minor.

Mixing up step rates and overall rates. "Add to cart is 12%" means very different things depending on whether it is measured against visits (10,080 ÷ 84,000 is 12%) or product views (20%). Label every rate with its denominator.

Fixing the funnel average. If you had redesigned the whole checkout because "conversion is down", you might have spent a month on desktop changes that fixed nothing.

Writing it up for the review

The finding fits in three lines, and it should, because the people in the weekly review need the conclusion more than the method:

1. What happened. Conversion fell from 3.2% to 2.7%, costing $22,680 of revenue on unchanged traffic and spend.
2. Where. Entirely in mobile checkout completion, which fell from 54.0% to 40.1% starting on the 9th. Desktop and every earlier step held.
3. Why, and what next. The checkout release on the 9th moved express payment below the fold on small screens. A fix ships Thursday. We will confirm recovery in next week's numbers.

Put the funnel table and the device split underneath, for anyone who wants to check the working. Most people won't, and that is fine. The fact that the working is visible is what makes the three lines believable.

Building it so next month is faster

The work above takes a few hours the first time, mostly assembling the funnel table and the device split. The second time it should take minutes, if you keep the structure.

In a spreadsheet, keep one table per month with the step counts by device, and compute step rates with formulas rather than typing them. Add the noise range as a column. When a step falls outside its range, conditional formatting turns it red.

In FlowProofs, the same funnel sits on a canvas: each step a shape reading from your analytics or order export, step rates as formulas, and a filter that splits the whole funnel by device at once. When conversion moves, the why panel shows which step accounts for the change. It is built for presenting this to the team. It won't replace your analytics tool for the exploration itself. The data analysts page shows the setup.

Wherever you build it, the order of operations is what matters: confirm the drop is real, lay out the steps, check the change is bigger than noise, segment the step that broke, then look at what changed. If your funnel dropped this month, start by checking whether early steps moved. That one comparison tells you whether to call the marketing team or the checkout team. For the same method applied to revenue as a whole, see how to explain why revenue changed.