Before you start
- The explorations tutorial, completed. This tutorial assumes you can open Explore, add dimensions and metrics to the variables panel and build a basic funnel; the GA4 explorations tutorial covers that and is the prerequisite.
- The event schema. A list of the events the site sends, their parameters and where each fires, from the tracking plan or from the events and parameters guide. Funnel steps are conditions on events, and guessing at names wastes the hour.
- A defined conversion. The event that counts as the outcome —
purchase,generate_lead, a form submission — and the business's own count of it for the period, for reconciliation. - A period inside data retention, long enough to cover full weeks and the seasonality the business has, and short enough that the exploration is not heavily sampledWhen an analytics tool calculates a report from a subset of the data and scales the result up, rather than processing every event..
- Access to the property with rights to create explorations, and knowledge of who should see the result when it is done.
- A question. "Where does the checkout lose people, and does it differ on mobile?" is a question; "let's look at the funnel" is not.
Define the steps from the journey, not the events
Write the journey as a person experiences it before touching the tool: arrive, view a product, add to basket, start checkout, enter delivery details, enter payment, confirm. Each step should be a decision point the business cares about, not every event the site sends; a funnel with a dozen steps spreads the loss so thinly that nothing stands out, and one with three hides where inside the checkout the loss sits. Then map each step to the event that marks it, with the parameter that narrows it where needed — a
page_viewon the delivery page, anadd_shipping_infoevent, abegin_checkout. Write the mapping down as a table; it is the specification for the exploration and the record of what the funnel means.Check that every step's event exists and fires once
Before building anything, confirm each event in the mapping is actually arriving. The events report shows the names and counts for the period; DebugView shows a live session. Walk the journey on the site with debug mode on and watch each event fire once, at the right moment, with the parameters the step condition will use. The faults this finds are familiar: an event that fires on every page instead of one, a step that fires twice on reload, a redesigned template that stopped sending its event, a parameter that is empty on mobile. A funnel built on a broken event measures the breakage, and fixing it afterwards does not repair the history, so the check comes first. Reconcile the conversion event against the business's count for one closed day while you are there.
Build the funnel exploration
In Explore, start a new exploration and choose the funnel technique. In the tab settings, add the steps in order, each as a condition on an event name, narrowed by a parameter where the mapping says so. Google documents a limit on the number of steps a funnel can hold; a journey that needs more than the limit is two funnels. Name each step with the journey word, not the event name, so the chart reads as the journey. Set the date range to the period chosen earlier and check the sampling indicator; if the exploration is sampled heavily, shorten the period or accept the approximation and say so when sharing. Add the conversion step last, so that the final rate is the one that matters to the KPIA metric tied directly to an objective..
Decide between an open and a closed funnel
Google's funnel exploration can be open, where people may enter at any step, or closed, where only people who completed the first step are counted at later ones. The difference changes the reading. An open funnel can show more people at a later step than at an earlier one — visitors who landed directly on a product page from an ad never saw the step before it — and reading it as a sequence produces impossible rates. A closed funnel reads as a sequence but excludes the people who entered mid-way, which for a shop may be most of the traffic. Use the closed funnel for step-to-step rates and the open one to see how many people the first step actually captures. The same panel sets whether each step must directly follow the previous one or may be indirectly followed with other events in between; for checkouts, indirect is usually right, because people go back and forth.
Segment by device
Add device category as the breakdown dimension, and the funnel splits into rows for each device. Read the step rates, not the totals: the question is which step loses a larger share on mobile than on desktop, because that step has a mobile-specific problem — a form that is hard to complete on a small screen, a payment method missing on mobile, a layout that hides the continue button. A step whose rate is similar across devices has a problem that is not about the device. Note the differences per step; they go into the hypothesis later. Google also documents an elapsed-time option that shows how long each transition takes, and a transition that is much slower on mobile is a hint even when the rate is similar.
Segment by source
Change the breakdown to session source or the default channel group, or build segments for the main sources and apply them as comparisons. Traffic from different sources enters with different intent: paid search visitors who searched for the product convert through the early steps far better than social visitors who were interrupted by an ad, and comparing them on the final rate alone says nothing useful. Look for the source whose visitors drop at a step where others do not; that is often a mismatch between what the ad promised and what the page shows, or an audience the page is not built for. Keep the source breakdown separate from the device one at first, because a table with both dimensions splits the sample into cells too small to read.
Compare periods
A funnel for one period shows where the loss is; two periods show whether it is changing. Build the same exploration for the previous period of the same length, or use a date comparison where the technique allows it, and lay the step rates side by side. A step whose rate fell between periods points at a release, a template change or a tracking change on that step, and the change log should name it. A step whose rate is stable is a structural problem rather than a regression. Seasonality complicates this: compare like periods — the same weeks last year, or a period before and after a known change — and avoid comparing a sale week with an ordinary one. Record both periods' step rates in the mapping table.
Find the biggest drop and do the arithmetic
With the step rates in hand, resist the instinct to work on the step with the lowest rate. For each step, take the number of people entering it and the improvement in its rate that a fix could plausibly produce, and estimate the extra conversions at the downstream rates; the conversion rate calculator does the rate arithmetic. The step with the largest estimate is the biggest leak in the sense that matters. Late steps often win because everyone there has declared intent and the fix is on a page the business controls. Use the next-action panel on the losing step to see what people who left did instead — went back, searched, left the site — and the option on a drop-off to build a segment of the people who left, which is the audience for recordings and surveys. The checkout optimisation guide catalogues the usual causes when the leak is inside the checkout. If the losing step is limited by something behind the page, such as stock or the capacity to answer enquiries, the theory of constraints for marketing helps decide whether to fix the page or the capacity behind it.
Write the hypothesis
A funnel finding is a location, not a cause. Before changing anything, combine the drop with evidence about why: recordings of the segment that left, a poll on the losing step, a usability session walking through it. Then write the hypothesis in the form the hypothesis-writing tutorial sets out — the evidence, the problem, the proposed change, the expected effect on the step's rate and the metric that will judge it. The step rate is the primary metric for the test, the overall conversion rate (CVRShare of visits or clicks that convert.) and order value are the guardrails, and the funnel exploration, saved and named, is the baseline the test will be read against. Because a funnel stops at its last step, a change that lifts a step rate by letting through people who never pay only shows up later, which is why the comparison of cohort and funnel analysis pairs each funnel fix with a reading of the cohorts it produced.
Share the result
Explorations can be shared with other users of the property, but a shared exploration is not a finding. Write a one-page summary: the journey as defined, the step rates for the period and the comparison, the device and source differences that matter, the biggest-leak arithmetic, and the hypothesis with the test that follows. Include the sampling status and the date range, because anyone who rebuilds the funnel later with different settings will get different numbers and needs to know why. Save the exploration with the question in its name, and record the mapping table in the measurement documentation, so the next funnel analysis starts from a definition rather than a blank page. The funnel and checkout hub is where the finding leads next.
Reading drop-off by segment
The explorer below shows a six-step e-commerce FunnelModel of the customer journey stages. on demonstration data, with the step-to-step rate beside each bar and the biggest step loss highlighted. Switching between all traffic, mobile and desktop shows the pattern the device step of this tutorial is looking for: the overall funnel hides a leak that one segment carries almost entirely. Use it to practise the reading — rate per step, not count per step, and the leak as the step whose loss would be most worth recovering — before opening the real exploration.
Funnel drop-off explorer
A furniture retailer's checkout funnel
A furniture retailer defines a closed, indirectly followed funnel for four weeks: product view, add to basket, begin checkout, add shipping information, add payment information, purchase. The all-device funnel shows 182,000 product views, 21,800 basket adds, 11,600 checkout starts, 8,900 with shipping entered, 7,400 with payment entered and 6,950 purchases. Broken down by device, the shipping step converts 84% of checkout starts on desktop but 69% on mobile, while every other step is within a few points across devices. Elapsed time on that transition is nearly twice as long on mobile. Recordings of the mobile drop-off segment show the delivery-date picker failing to open on one widely used browser. The hypothesis names the picker; the test's primary metric is the shipping step rate on mobile, with purchase rate and average order value as guardrails. The retailer estimates that bringing mobile to the desktop rate on this step alone would add around 900 orders a month.
Funnel checklist
| Check | Done when |
|---|---|
| Steps defined from the journey, each mapped to one event and parameter | The mapping table exists and someone else can rebuild the funnel from it |
| Every step event verified to fire once, at the right moment, on every template and device | A DebugView walk and a one-day reconciliation against the business count are recorded |
| Funnel built with the conversion as the last step | The final rate matches the period's conversion rate within the sampling margin |
| Open and closed readings both taken | The closed funnel supplies the step rates; the open one shows mid-funnel entry |
| Device breakdown read per step | The step with the largest device gap is named |
| Source breakdown read per step | The source with an unusual drop is named, with a candidate cause |
| Period comparison done on like periods | Step rates for both periods are in the mapping table; regressions matched to the change log |
| Biggest-leak arithmetic completed | The step with the highest estimated extra conversions is chosen, with the working shown |
| Hypothesis written with a primary metric and guardrails | It follows the hypothesis template and names the test |
| Result shared with sampling status and date range | The exploration is saved and named; the summary is with the stakeholders |
Where the funnel shows more arrivals with no more completions, why traffic grows while sales stay flat helps decide whether the step, the traffic source or the counting is at fault. When the leak sits before checkout, at product view or basket, the guide to product and merchandising analysis in GA4 reads the same events by item, list and promotion.
When the whole rate falls on a single day rather than leaking steadily at one step, the diagnosis for a conversion rate that dropped suddenly uses the same breakdowns to separate tracking, traffic, bug, payment, price and bot causes.
Verify it worked
- Open the mapping table. Someone else should be able to rebuild the funnel from it and reach the same step counts.
- Check the event walk. Each step's event fires once, at the right moment, with the parameters the step condition uses.
- Compare the final step rate with the conversion rate for the period. They should agree, allowing for sampling.
- Read the closed funnel. No step should hold more people than the one before it; only the open funnel shows mid-way entry.
- Read the device breakdown per step. One step should stand out, as shipping did in the furniture example, where mobile fell well behind desktop.
If it goes wrong
- A later step shows more people than an earlier one. The funnel is open, so people entered mid-way. Switch to the closed funnel for step-to-step rates.
- A step count looks wrong, or the final rate misses the business's count. An event probably fires twice, everywhere or not at all. Repeat the event walk and fix the event first.
- The sampling indicator shows heavy sampling. Shorten the period, or keep the approximation and say so when you share the result.
- Device and source cells are too small to read. You split by both at once, so return to one breakdown at a time.
- No step stands out by device. The problem may not be about the device, or the funnel has too many steps. Try the source breakdown and check each step is a real decision point.
Next steps
- Read the checkout optimisation guide when the biggest leak sits inside checkout, to turn the segment findings into candidate fixes.
- Write the finding up with the hypothesis tutorial, so the leak becomes a testable statement with a primary metric.
- Return to the events and parameters guide if an event failed the check, to correct names and parameters before rebuilding.
Common mistakes
- Building the funnel from the events list. Steps are decision points in a journey; the event is how each is observed.
- Trusting an unverified event. A step that fires twice or not at all defines the "finding".
- Reading an open funnel as a sequence. Later steps bigger than earlier ones are entry, not magic.
- Working on the lowest rate. The leak is the step where a fix recovers the most conversions, which is usually later.
- Two breakdowns at once. Device by source by step produces cells too small to read.
- Comparing unlike periods. A sale week against a normal one measures the sale.
- Treating the drop as the cause. The funnel says where; recordings, polls and usability sessions say why.
- Not saving the definition. The next analysis rebuilds a different funnel and the numbers "change".
