Funnel analytics that explains the drop, not only the rate

A funnel is an ordered question: of the people or visits that reached one observable step, how many reached the next step inside a defined window?

The chart is easy. The hard work is defining steps that represent real value, preserving their meaning through redesigns and connecting a drop to evidence a team can act on.

MetricFold treats a funnel as a versioned operating contract. Every definition has a name, ordered steps, identity scope, time window, strict or loose order, filters, exclusions and owner. Reports show the sample and definition next to the result so a percentage cannot drift away from its meaning.

Start with one business question

“What is our conversion funnel?” is too broad. Choose a decision:

• Where do qualified visitors abandon signup?
• Which first product action predicts activation?
• Does reaching the pricing page lead to a verified payment?
• Which checkout step fails on mobile?
• Did the new onboarding reduce time to first value?
• Which source creates retained customers rather than cheap signups?

Then define the smallest sequence that answers it. A self-serve SaaS activation funnel might be:

1. signup_started
2. account_created
3. workspace_created
4. first_value_completed
5. subscription_paid

The first step may be automatic form-start evidence. Account creation and payment must be trusted server events. Workspace and first value should be domain events emitted after durable success, not clicks on buttons that may fail.

Use stable steps

URL steps are convenient for marketing paths. Event steps are better for product state. Use a URL when reaching the page is the outcome, such as a campaign landing page. Use an event when success depends on processing, such as a report being generated or an integration completing its first sync.

Avoid visible labels, CSS selectors and component names. They change during a redesign and split historical data. Semantic action attributes are appropriate for meaningful browser interactions, while trusted server events own money, entitlements and durable business transitions.

Every property should be allowlisted. Good examples are plan family, template type, integration family or product surface. Email, names, search text, free form errors and exact account IDs do not belong in a public analytics payload.

Define identity and time explicitly

A visit funnel answers whether steps happened during one browsing visit. A daily pseudonymous funnel can span visits on the same day. An authenticated product funnel may use a host-owned non-reversible subject reference across a longer window when the product's policy allows it.

Those are different measurements. MetricFold does not silently merge them. The report states its identity scope and window, and a privacy-minimized public visitor key is not stretched into a cross-month profile.

Time-to-convert is as useful as completion rate. Show the median and a bounded distribution between steps. A high final conversion with a three-day delay may signal a missing explanation, failed email or payment obstacle.

Move from drop-off to diagnosis

A step count alone says where measurement stopped. It does not prove why. MetricFold links the drop to adjacent evidence:

• dead or repeated clicks on the expected control;
• abandoned forms and validation-stage counts without field values;
• failed requests grouped by safe operation name and status class;
• client errors with scrubbed stable fingerprints;
• LCP, INP and CLS distributions for completers and non-completers;
• device, browser, country, source, campaign and page segments;
• paths taken immediately before exit;
• release and campaign annotations.

The system can then produce a bounded hypothesis: “Mobile Safari visits that reached checkout had a higher payment-step drop and a concentrated request failure after release 2026.08.1.” That is evidence for a reproduction step, not proof of causation.

Segment without manufacturing certainty

Inspect segments with enough observations and a reason to differ. Device class, browser family, new versus returning scope, source, campaign and plan are usually useful. Do not search hundreds of tiny cuts until one looks dramatic.

MetricFold places denominators beside rates, marks truncated or low-sample results and limits the number of AI-generated hypotheses. An apparent 100% improvement from two visits should not outrank a smaller change affecting hundreds.

Compare adjacent periods only after checking seasonality, campaign mix and release dates. An annotation is often more useful than a more complicated model.

Validate the implementation

Before trusting a funnel:

1. trigger every step once on desktop and mobile;
2. confirm failed actions do not emit success events;
3. repeat and reload to verify event IDs deduplicate retries;
4. test SPA back/forward navigation and SSR hydration;
5. complete a provider test payment, renewal and refund;
6. verify excluded bots and staging traffic never enter the denominator;
7. inspect the raw bounded event shape without exposing personal data;
8. compare strict-order and loose-order results and record which one the review uses.

The goal is not a beautiful tapering graphic. It is a definition the product, engineering and growth teams trust, plus enough linked evidence to turn the largest valuable drop into one measurable change.

Choose the funnel rule explicitly

Plausible's funnel analysis documentation shows page and custom-event goal steps, while PostHog documents ordered funnel analysis. The product team still needs to choose the identity, order and window that match its decision.

Design each step as an observable contract

Prefer completed semantics

Use account_created, project_first_run_succeeded and subscription_payment_succeeded rather than button clicks. A CTA click is useful as an attempt step, but the next durable outcome should come from the system that owns it. The product analytics guide explains how to separate client intent from server truth.

Keep properties bounded

Allowlist plan, route template, device class, experiment assignment and other dimensions required for a planned analysis. Reject names, email addresses, free-form messages and high-cardinality identifiers. Version the event when its meaning changes; do not reuse a familiar name for a new outcome.

Define eligibility and expiry

State who can enter the funnel and when the opportunity expires. A trial activation funnel may include accounts created in the reporting window and allow fourteen days to activate. The report must distinguish mature cohorts from recent entrants whose window is still open.

Read the funnel from denominator to intervention

Begin with instrumentation health: unknown events, excluded traffic, duplicates and trusted-outcome lag. Then identify the earliest step whose loss is both large and commercially important. Compare one or two pre-registered segments and inspect adjacent friction, performance and path evidence.

Quantify the opportunity

For each step, show entrants, completions, drop count, conditional conversion and end-to-end conversion. Estimate an opportunity range using affected actors and the downstream verified value rate; do not multiply every lost visitor by average customer value as if all would convert.

Choose one falsifiable change

Write the suspected mechanism and competing explanations. If request failures rise on mobile at the payment step, repair and test the operation before rewriting the pricing page. If no technical evidence exists, an explanation, field reduction or usability study may be the right experiment.

The friction analytics guide covers reproducible blockers and the customer journey guide covers valid detours that a single funnel cannot express.

A worked activation-funnel review

Assume 1,000 eligible trial accounts begin setup. Eight hundred connect a data source, 620 complete a successful first run, 410 return to repeat the core action, and 190 become verified paid accounts inside the mature window. The end-to-end paid conversion is 19%, but the first review should not simply choose the largest numerical loss.

The first step loses 200 accounts. Segment it by supported integration and request result. If one connector has a concentrated authentication failure, there is a technical reproduction. The first-run step loses 180. Compare time, validation class and device; a long successful operation may need progress and recovery rather than fewer fields. The repeat-use step loses 210, but that may reflect product cadence. Confirm whether repeat use is required before paid value and whether recent cohorts had enough time.

For every step, calculate conditional conversion from the previous step and end-to-end conversion from eligibility. Preserve raw actor counts. Compare with an equal mature cohort and annotate releases or campaigns. Do not add the step-level losses together as if they were independent opportunities.

Google's funnel exploration documentation is another example of ordered funnel analysis. MetricFold's responsibility is to keep the actor, order, identity, time window and trusted outcome visible alongside the graphic.

Frequently asked questions

Is funnel conversion visitors or visits?

It depends on the declared actor. A public marketing funnel may use bounded visits, while a product activation funnel may use an approved account key. The dashboard must show the denominator beside the rate.

How long should the conversion window be?

Use the shortest window that covers a normal decision cycle. Inspect the time-to-convert distribution, choose a documented boundary and treat still-open recent cohorts separately.

Should optional steps be included?

Only when the question requires them. Optional exploration belongs in path or journey analysis. Forcing it into a strict funnel can make healthy behavior look like failure.

Can the funnel prove a change caused conversion?

No. It describes observed progression. A controlled experiment with a valid assignment and guardrails is needed for causal attribution.

How should open cohorts be displayed?

Keep entrants whose conversion window has not expired out of a final failure denominator or label them “still eligible.” Otherwise a surge of recent signups can make activation appear to collapse before those accounts had a fair opportunity to complete.

Which comparison window is valid?

Use an equal cohort and observation window whose eligible actors had the same time to convert. Check campaigns, seasonality, releases, pricing and provider incidents before attributing movement to a product change. Preserve the former definition if the event plan changes; a new event version starts a new baseline.

What belongs in an AI funnel packet?

Include actor and identity scope, ordered steps, mature window, entrants, completions, step losses, samples, filters, exclusions, trusted-outcome status and truncation. Ask for competing explanations and a reversible next check. Do not let the model query unbounded raw journeys or rewrite the live funnel.