Data & Decisions
Choose a Funnel Conversion Window That Matches the User Journey
Compare elapsed conversion windows on the same synthetic event log. See how order, repeat entries, object identity and observation time change the count, then use a completed decision record to choose a provisional window without inventing product improvement.
Choose a funnel conversion window from the task users are trying to finish, then test how the conclusion changes under a plausible alternative. Define the starting event, required order, counting unit and elapsed-time boundary before comparing rates. A longer window can recognize a legitimate return visit, but a higher resulting percentage does not prove that the product improved.
This LandOffer.ai guide compares 24-hour and 72-hour windows on one original synthetic event log, then completes a window-selection decision record. The offline calculation was actually run; the people, workflow evidence and product decisions are fictional. No analytics account, real user data or deployed instrumentation was tested. The authored calculation is not a reproduction of Amplitude's engine.

State what completion means before choosing a duration
A funnel answers a specific sequence question. For example: of users who begin preparing a room booking, how many review its details and confirm it within the allowed time? Counting everyone who fired any of those events during a calendar week answers a different question. Those users may not have completed the sequence in order, or even started it within the cohort period.
Amplitude's Funnel Analysis FAQ distinguishes funnels from event segmentation: funnel entrants perform the first event in the selected date range, then complete the required later events in order within the window. It also distinguishes unique-user counts from event totals and notes that a day window is elapsed time rather than a strict calendar date. These documented distinctions inform the checks below; our custom counting rule remains explicit and separate.
Write the business outcome in plain language. A confirmation click may be an attempt, while a server-confirmed booking may be the outcome you actually need. Decide which event is authoritative and what its timestamp represents. Check the outcome record rather than treating a displayed button as a completed booking.
Then specify the unit. A unique person, account, session and booking are different denominators. A person making two bookings might count once in a user funnel and twice in a booking-level measure. Neither is automatically wrong; they answer different questions.
Separate entry dates, elapsed windows and data availability
The entry date range selects who can enter this cohort; it does not require every later step to fall inside that same date range. The conversion window sets how long an entrant has to complete the task. The data cutoff tells you which events were available for analysis. Keeping all three in the report, with timezone, lets another reader distinguish a different cohort from a different opportunity to finish.
An entry at 23:30 followed by completion at 01:30 the next day takes two hours. Ending its opportunity at midnight would impose a thirty-minute window on that entrant while allowing almost a full day for an early-morning entrant. If you intend elapsed hours, calculate elapsed hours.
Amplitude's interpretation guidance defines the window from entry to completion and describes incomplete windows at the recent end of a conversion-over-time chart. For a real analysis, confirm both the chart configuration and whether entrants have had the entire opportunity to convert. A reporting tool's available data can still lag the events that occurred.
Our example selects entries on November 2, 2026, from 00:00 inclusive to November 3 at 00:00 exclusive, all in UTC. The saved synthetic log covers events through November 6 at 00:00. Every entrant therefore has at least 72.5 hours of observation, enough for both candidate windows. These are invented scenario dates, not a claim that future production observations already exist. The latest entry is November 2 at 23:30, giving 72.5 hours to the cutoff, which still covers both windows.
Worked asset: the same events under two windows
The fictional Orchard room-booking app records three events: A: preparation started, B: details reviewed, and C: booking confirmed. Each event carries a stable synthetic user ID and a booking ID. Every user in the fixture has one booking; all steps must retain that booking ID. Other unrelated events would be permitted between the required steps.
For this exercise, count each user once from their first qualifying A in the entry period. Ignore later A events as restarts. Require strictly increasing A, B and C timestamps, and require B and C to be no later than the chosen window after that first A. Equality at the final boundary counts here by author-defined policy. Confirm your vendor's actual boundary and repeat-entry behavior separately before applying this rule elsewhere.
| User | First A on November 2, UTC | B after first A | C after first A | 24-hour / 72-hour completion |
|---|---|---|---|---|
| Aster | 00:30 | 1 hour | 2 hours | Yes / Yes |
| Birch | 23:30 | 0.5 hours | 2 hours | Yes / Yes; crosses midnight |
| Clover | 09:00 | 20 hours | 26 hours | No / Yes |
| Dahlia | 12:00 | 48 hours | 70 hours | No / Yes |
| Elm | 15:00 | 1 hour | 73 hours | No / No; outside both windows |
| Fern | 16:00 | 2 hours | 1 hour | No / No; C precedes B |
| Hazel | 17:00 | 24 hours | 25 hours | No / Yes; second A at +1 hour ignored |
| Iris | 18:00 | 23 hours | 24 hours | Yes / Yes; inclusive boundary |
| Juniper | 19:00 | 70 hours | 72 hours | No / Yes; inclusive boundary |
| Linden | 20:00 | 1 hour | No C recorded | No / No |
The actual Python calculation checks all 30 authored event rows, including Hazel's extra A. It produces these counts without changing the entry cohort or adding users:
| Cumulative step reached within the rule | 24 hours | 72 hours |
|---|---|---|
| First A, unique users | 10 | 10 |
| Ordered A then B | 8 | 10 |
| Ordered A then B then C | 3 | 7 |
| Complete / entered | 30% | 70% |
The difference is 40 percentage points, entirely caused by the allowed window on the same synthetic data. It is neither a measured improvement in product behavior nor a change produced by a release; the original events are identical in both calculations. At 24 hours, three of the eight users reaching B finish C within the entry-based window: 37.5%. At 72 hours, seven of ten do: 70%. Label those step-conditional denominators separately from overall conversion.
The full CSV and per-user decisions are retained with the local article evidence. The script checks counts, ordering, equal-boundary cases, duplicate starts and mature observation time. It does not validate real identity resolution, timestamps, missing events, vendor computation or booking success.
Read the exclusions before diagnosing drop-off
Fern has all three event names, but no ordered A–B–C sequence. The script cannot treat the early C as a later confirmation. That could represent an instrumentation problem, a different permitted path or an invalid business assumption in real data. Inspect the underlying sequence before calling it user abandonment.
Hazel demonstrates why repeated entries need an explicit policy. Restarting the clock at the later A would place C at exactly 24 hours after that restart, changing this custom 24-hour result. Our first-entry policy deliberately does not do that. Amplitude's conversion-computation documentation describes its own earliest/most-complete logic and how holding properties constant can change the unit. Do not assume an author-written first-entry script and a vendor's unique-user chart must agree.
Elm confirms after 73 hours, so neither proposed window counts the completion. “Not converted within 72 hours” is accurate; “never completed” is not. Linden has no observed C, but that does not by itself prove a decision to abandon the booking. A missing event and an unfinished task need different investigations.
Avoid adding a session constraint casually. A legitimate review-and-return task may span sessions or devices. Conversely, joining steps only by user ID can combine different bookings. Verify the business object and identity rules, not just whether the total looks plausible.

Worked asset: choose a primary window from task evidence
Analyst Sabine prepares a selection record for product owner Oren. The stipulated workflow allows staff to prepare a booking, check room details with a colleague and return later to confirm. The fictional team's stipulated process notes describe occasional next-day review. They justify considering a longer window, but do not establish a population distribution of completion times or prove that 72 hours is optimal. The ten-row teaching fixture is not representative evidence for selecting a universal duration.
| Selection question | Filled record | Decision or remaining requirement |
|---|---|---|
| What is the outcome? | Authoritative booking-confirmed event for the same booking as preparation and review | Engineer Darien must verify the production event contract before any real report is approved |
| Why consider more than 24 hours? | The documented fictional task can wait for a colleague's next-day review | Use 72 hours as a provisional task-completion measure; do not describe all extra time as friction-free |
| What happens to the shorter window? | Three of ten synthetic users complete within 24 hours | Retain 24 hours as a speed diagnostic alongside the primary definition, not as a conflicting replacement headline |
| Which cohort is comparable? | Same eligible entry rules, UTC date boundary, user/booking identity and complete observation horizon | Compare only equally mature cohorts; label recent cohorts provisional |
| What would change the definition? | A changed review policy, evidence of multiple distinct tasks, or a different unit such as bookings | Sabine versions the metric before rerunning history; Oren approves the business interpretation; until then, revised definitions remain proposals |
| What may be concluded now? | Synthetic arithmetic demonstrates sensitivity; real event and task evidence are unverified | No real conversion baseline, product-success claim or release decision is approved |
Oren's completed fictional decision reads: “Adopt a provisional 72-hour definition for this review-and-return task, retain the 24-hour diagnostic, and hold production reporting until Darien reconciles the event contract and Sabine verifies a fully observed cohort. The teaching result of 70% is not our product conversion rate.”
The team also keeps a slow-completion question open. A wider window may fit the task while hiding an unnecessary delay between preparation and review. Review time-to-complete alongside conversion, with clear population limits. A duration summary calculated only among converters excludes people without an observed completion; it cannot describe how long everyone takes.
No choice here establishes causation. If a later release reports a higher rate, first check whether the window, entry population, eligibility, instrumentation or data cutoff changed. A definition change can move the number while user behavior stays exactly the same.
Preserve the definition wherever the number appears
Save a compact metric contract with event names, object identity, counting unit, order mode, repeat-entry policy, entry dates, timezone, elapsed window, boundary rule and observation cutoff. Include the script or chart configuration version. Document ingestion delay and timestamp checks when you have evidence for them; otherwise mark those checks pending.
Use sensitivity analysis to explain the decision, not to search for the most flattering setting. If the headline becomes 72-hour completion, label prior 24-hour results as a different metric and retain their history. Do not splice the two definitions into a seamless trend.
Finish with the action the evidence supports: verify the contract, wait for a mature cohort or investigate a specific delayed step. Keep later return behavior in a separate retention question. For more practical work records, continue with LandOffer.ai's workplace guides; keep this funnel's definition attached whenever you reuse its rate.
Sources and Further Reading
- Amplitude: Funnel Analysis FAQ — unique users versus totals, elapsed rather than calendar-day windows, and sequence/date-range distinctions.
- Amplitude: Interpret your funnel analysis — entry-to-completion duration, conversion denominator and incomplete observation windows.
- Amplitude: How Amplitude computes conversions through funnels — documented unique-user selection and property-holding behavior. All three current primary bodies checked October 27, 2026; exact publication dates are not displayed. No authenticated chart test was performed.


