Experiment specification

Experiment spec: delivery date and fee on the basket page

Hypothesis

Showing the delivery date and fee on the basket page, rather than at the final checkout step, raises the share of basket sessions that end in an order. We expect fewer last-step abandons because the fee is no longer a surprise. The counter-risk is more exits at the basket, so the primary metric must count both.

Randomisation

• Unit: logged-in user ID, not session or device. 31% of buyers use both app and web, and buyers average 3.8 sessions a month. Session-level assignment would show one person both experiences and make observations non-independent.
• Split: 50/50 by salted hash of user ID, the same arm on app and web. Users are assigned on first basket view.
• Independence: use a different salt from the Payments squad's "Pay later" test, so every arm is about 50% Pay-later (see Calendar).
• Checks: run an A/A check in the week before launch. Run a sample-ratio-mismatch (SRM) test throughout, and pause if p < 0.001.

Metrics

Primary: order conversion = basket-page sessions that end in an order ÷ basket-page sessions, with standard errors clustered by user.

Do not use orders ÷ checkout sessions. If the fee scares people off at the basket, they never become checkout sessions, the denominator shrinks, and the treatment looks better than it is.

Secondary: - Basket→checkout-start rate (expected to fall, and we need to see by how much). - Checkout-start→order rate (the squad's mechanism). - Revenue per basket session. - Orders per user over the window. - Platform split (exploratory only).

Guardrails: - Average order value (AOV, baseline $64): fail if the lower CI bound is below −2% (−$1.28). - Revenue per basket session: fail if significantly negative. - Fee-related support contacts per 1,000 orders. - Basket-page error rate and latency. - 14-day repeat rate among exposed users. The 30-day repeat rate (38%) can't be read in a 3-week test.

Baseline from the traffic file

• Weekly sessions are about 260k and daily volumes are stable, so I use p = 6.35%.
• Platform split: app 57–58% of sessions at about 7.1% conversion; web 42% at about 5.3%.
• Weekends run at about 65% of weekday volume, so the test must cover whole weeks.

Sample size

• MDE: +3% relative, which is +0.19pp (6.35% → 6.54%). That is roughly 490 orders a week, or about $31k a week at $64 AOV, which is worth shipping for.
• Settings: two-sided α = 0.05, power 80%, so (1.96 + 0.84)² = 7.84.
• Per arm, ignoring clustering: n = 2 × 7.84 × p(1−p) / δ²
• p(1−p) = 0.0635 × 0.9365 = 0.0595
• δ = 0.00191, so δ² = 3.63×10⁻⁶
• n = 0.932 / 3.63×10⁻⁶ ≈ 257k sessions
• Clustering: users contribute about 2.7 sessions in 3 weeks. I assume an intra-user correlation of 0.2, giving a design effect of 1 + 1.7 × 0.2 ≈ 1.3. That puts the requirement at 257k × 1.3 ≈ 334k sessions per arm.
• Duration: each arm gets about 130k sessions a week (260k ÷ 2), so 334k ÷ 130k = 2.6 weeks. Run 3 full weeks (21 days), which gives about 390k per arm.
• For comparison: a 2% MDE needs 578k × 1.3 = 751k per arm, or 5.8 weeks, which the calendar can't fit. A 5% MDE needs 92k × 1.3 = 120k per arm, under 1 week.
• Subgroups: web alone (42% of traffic) can only detect about 4.6% relative (3% × √(1/0.42)). Platform cuts are exploratory.

Data issues to resolve before launch

1. Buyer count doesn't match the traffic file. The data team's 412k buyers in 30 days is inconsistent with about 70k orders a month in the file (16.4k a week × 30/7). Their 3.8 sessions per buyer implies about 1.57M sessions, against about 1.11M in the file. I sized from the traffic file. The data team needs to reconcile the two, since the user-level assumptions (the design effect) depend on it.
2. Check the definition of `checkout_sessions`. I assumed it equals basket-page sessions. If it only counts sessions that reached checkout, the true basket-level baseline is lower and the test is longer. For example, at 4% conversion, n scales by (0.96/0.04) ÷ (0.9365/0.0635) ≈ 1.6×, which is about 4.5 weeks.
3. Re-run the sums on the real basket-session baseline before launch, and fix the duration then (3 or 4 weeks).

Calendar

• Black Friday: promotions start 20 Nov and code freeze starts 16 Nov. Promo traffic isn't representative, and anything not shipped by 15 Nov waits until after the freeze.
• Waiting for Pay later to end (9 Nov) doesn't work. That leaves 10–15 Nov, which is 6 days and not enough.
• Pay later (until 9 Nov) sits on the checkout page and affects the same outcome, so it can't be ignored. I'll run the test concurrently with an independent split. Each arm then gets about the same Pay-later mix, so the contrast stays unbiased. I'll pre-register a treatment × Pay-later interaction check, and report the effect by Pay-later arm. The Payments squad must freeze their allocation for the duration.
• Search ranking (until 30 Oct) only touches search results pages. It's balanced by randomisation and needs no action, but I'll note it.
• Schedule:
• Build, QA and A/A: 5–11 Oct.
• Run 12 Oct – 1 Nov (3 weeks).
• Readout 2–4 Nov, decision before 9 Nov.
• Ship ramp before the 16 Nov freeze.
• Buffer: a 4-week run (to 8 Nov) still ends before Pay later does.

Decision rule

This is fixed in advance, with one analysis at the end of the run and no peeking.

• Ship if the primary lift is positive with the 95% CI excluding 0, and no guardrail fails, and revenue per basket session isn't significantly negative.
• Do not ship if the CI includes 0. We'd conclude "no detectable effect at +3%". Any wish to ship for fee-transparency reasons is a separate product decision, not an experiment win.
• Kill if the primary is significantly negative, or if AOV or revenue per basket session breaches its guardrail. Basket exits are the main thing we expect to see here.
• Extend: only if an outage or traffic shortfall leaves fewer than 334k sessions per arm, up to 28 days. Never extend because the result is "almost significant".
• Interaction: if the effect differs materially by Pay-later arm, ship only to the configuration that will be live after 9 Nov, and confirm with the Payments squad which one that is.

Experiment spec: delivery date and fee on the basket page

Owner: Checkout PM · Status: Draft. Blocked on one data question (see 0).

0. Data question to resolve before launch

The two sources disagree badly:

• CSV: about 16,525 orders a week, or about 70,800 per 30 days.
• Data team: 412,000 people ordered at least once in 30 days, so at least 412,000 orders.
• Sessions: 412,000 × 3.8 = about 1.57M checkout sessions a month. The CSV shows about 1.12M.

The CSV may cover one region or a subset, or the definitions may differ. I've sized the test on the CSV because it is the conservative figure. If real traffic is about 6× higher, the test still runs the full 4 weeks to cover the weekly cycle. It would simply detect a smaller effect.

1. Hypothesis

Showing the delivery date and fee on the basket page removes the "surprise fee" at the final step. That reduces last-step abandonment and increases orders per shopper.

The counter-risk is that some shoppers leave at the basket instead. The test measures the net effect of both.

2. What's randomised and how

• Unit: account ID, 50/50. It is not the session. Buyers average 3.8 checkout sessions a month, and 31% use both app and web. Session-level randomisation would show the same person both versions and contaminate the result.
• Assignment point: first basket-page view in the test window, with the assignment held for the whole test.
• Logged-out web baskets: use the device cookie, then map to the account at login. Accounts that end up in both arms are excluded, and the rate is reported.
• Platforms: both, with app and web as pre-registered segments.
• Run first: a 2-day A/A check for sample-ratio mismatch.

Why the metric can't be "orders ÷ checkout sessions": the treatment changes who starts checkout. People who leave at the basket drop out of the denominator, which would inflate the treatment's rate. The denominator must be fixed at the point of assignment.

3. Metrics

Primary: orders per assigned account during the test window.

Secondary (diagnostic): - Basket → checkout start rate. Expected to fall; this is the risk the squad named. - Final-step abandonment rate. Expected to fall. - Basket → order conversion, split by app and web. - Share of baskets topped up to reach a free-delivery threshold, if one exists.

Guardrails (any breach blocks shipping): - Average order value (baseline $64): no drop of more than 2%. - Delivery-fee revenue per order: no significant drop. Must be explained if it moves. - Basket-page load time (p75): no increase of more than 100 ms. Basket-page error rate: no increase. - Customer contacts about fees or delivery date: no significant increase. - Sample-ratio mismatch: p > 0.001.

Note: the 30-day repeat rate (38%) can't be read inside the window. It becomes a follow-up read 30 days after the test ends.

4. Sample size and duration

Baseline from 8 weeks of CSV data (10 Aug – 4 Oct):

• Weekly volume is very stable: about 260,424 sessions and 16,525 orders a week.
• Weekdays run about 40% above weekends, so the test must run whole weeks.

Minimum detectable effect: a 3% relative lift, from 6.345% to 6.535% (+0.190 pp). At current volume that is about 496 orders a week, or about $31.7k a week at $64 average order value.

Calculation (α = 0.05 two-sided, 80% power):

• n per arm = (1.96 + 0.84)² × [p₁(1−p₁) + p₂(1−p₂)] / d²
• = 7.84 × (0.05942 + 0.06108) / 0.0019035²
• = 0.9447 / 3.623×10⁻⁶
• ≈ 260,750 sessions per arm, or 521,500 in total

Adjusting for repeat users: randomising by account means sessions from the same user are correlated. I've applied a design effect of 1.5 because the true clustering factor is unknown:

• 521,500 × 1.5 = 782,000 sessions
• ÷ 260,424 a week = 3.0 weeks

Planned duration: 4 full weeks. The extra week covers uncertainty in the clustering factor. It also leaves enough power for a directional read on app and web separately (each about 0.5–0.6 of total traffic).

The analysis will use account-level orders with variance computed directly from the data, so the 1.5 design effect is not baked into the result.

5. When it can run

Waiting until Pay later ends (10 Nov) leaves only 6 days before the freeze, which is too short. So we run alongside it.

Dates: - Build and A/A check: 5–11 Oct. - Test: Mon 12 Oct – Sun 8 Nov. That is exactly 4 weeks. It ends the day before Pay later finishes, so Pay later's 50/50 split is constant for our whole test. - Analysis and ship decision: 9–13 Nov, before the freeze.

6. Decision rule (pre-registered)

1. Ship if the primary lift is positive and significant (p < 0.05), no guardrail is breached, and sample-ratio mismatch passes. 2. Don't ship if: - the primary is significantly negative, or - any guardrail is breached.

Either way, investigate whether the basket-stage drop outweighed the last-step gain. 3. Flat result (confidence interval within ±3%, no guardrail breach): ship only if final-step abandonment fell significantly, on the grounds of transparency. Otherwise keep control. 4. Pay later interaction: if the effect differs significantly between Pay later arms, report the effect within each arm. The ship decision then follows Pay later's own launch decision. 5. No early stopping on the primary metric. A guardrail breach above a pre-set threshold may stop the test early.

Waymark checkout experiment spec

Hypothesis

Showing the delivery date and fee on the basket page will increase the share of eligible shoppers who place an order by setting expectations earlier and reducing last-step abandonment. It may also cause some shoppers to leave sooner; the primary metric must capture both effects.

Design and randomisation

Randomise eligible shoppers 50:50 when they first view a basket with delivery information available:

• Control: Current experience; date and fee appear at the final checkout step.
• Treatment: Date and fee appear on the basket page and remain visible through checkout.

Assign by stable customer ID across app and web, so shoppers who use both platforms see the same variant. Use a persistent anonymous ID for guests and merge it with the account on login. Stratify assignment by platform at first exposure.

Metrics

• Primary: Order conversion among eligible basket visitors: shoppers placing an order divided by shoppers randomised. This captures both earlier exits and improved final-step completion.
• Secondary: Checkout-start rate; completion among checkout starters; step-by-step abandonment; time to order; AOV; results by app/web.
• Guardrails: Basket-to-order conversion must not fall by more than 0.3 percentage points; monitor cancellations, fee-related support contacts, payment failures and site/app errors.

Sample size and duration

The traffic file covers 56 days. Summing its daily rows gives approximately 2.09m checkout sessions and 132k orders:

• Baseline checkout-session conversion: 132k ÷ 2.09m ≈ 6.3%
• Average traffic: 2.09m ÷ 56 ≈ 37.3k sessions/day, or 18.65k/arm/day at 50:50.

Power for a 5% relative lift (6.3% to 6.615%), with 80% power and two-sided 5% significance:

\[ n_{\text{per arm}}\approx \frac{2(1.96+0.84)^2(0.063)(0.937)}{(0.00315)^2} \approx 93{,}000 \]

Because shoppers can generate repeated sessions, conservatively inflate by the supplied 3.8 checkout sessions per buyer per month (worst-case design effect, assuming fully correlated sessions):

\[ 93{,}000 \times 3.8 \approx 353{,}000 \text{ sessions/arm} \]

\[ 353{,}000 \div 18{,}650 \approx 18.9 \text{ days} \]

Plan for at least 19 days, including full weeks and weekends, then allow time for order outcomes to mature. This is a provisional calculation: the file reports checkout sessions, not eligible basket visitors, and therefore cannot establish the primary metric’s baseline or power. Before launch, confirm basket-visitor volume and conversion, and validate the repeat-session assumption against deduplicated users. The supplied 412k monthly buyers and session-frequency figures should also be reconciled with observed traffic.

Calendar

Do not run concurrently with Payments’ checkout experiment, which ends 9 November. Starting 10 November would require about 19 days, crossing both the 16 November code freeze and 20 November Black Friday promotions. Defer until those changes and promotions are over and the freeze is lifted; the calendar does not specify those end dates. Confirm a stable window of at least three weeks before scheduling.

Decision rule

Ship only if the primary conversion lift is statistically significant at two-sided p < .05, the 95% confidence interval excludes zero, and the basket-to-order guardrail is not breached. Otherwise, do not ship; investigate platform and funnel-step results before proposing a follow-up. Don’t stop early for a positive result; stop only for a pre-agreed severe guardrail issue or data-quality failure.