Spillmark
Demo · made-up numbers
OverviewAlerts4Approvals
EExample Coffee Co.

Brands you can access

Example Coffee Co.
Data current · oldest 9h ago

When each source last synced

  • Amazon Seller Central3h ago

    Amazon fees are estimates until orders settle (about 14 days).

  • Amazon Ads5h ago

    Amazon Ads attributes purchases up to 7 days after a click, so the last week can still rise.

  • Amazon Attribution9h ago

    Amazon Attribution keeps adding purchases to each day for 14 days.

  • Meta Ads2h ago

    Meta revises the last few days of results as delayed conversions arrive.

  • Shopify45m ago
Manage connections →
DemoGet started

Tests

What should the next holdout test be, what will it cost, and can its results be trusted?

Loading tests…
OverviewAttributionNeeds Amazon's written approval to combine Amazon Ads and Meta data (Amazon Ads policy 4.9).TestsProductsCohortsInventoryAlerts4ApprovalsReportsConnectionsSettings

Next holdout test

Planned on Amazon orders through Oct 24: about 31 a day at $46.80 each.

Move the dates

Meta → Amazon pays for itself at a 7.5% lift in Amazon orders ($38.00 a day of spend, first orders only). Holding out 40% of orders for 8 weeks can detect 7.2% or more, enough to tell whether Meta pays back.

Dates
Nov 2 – Dec 27
Ads off in
27 states, 40% of orders
Sales missed
39 orders, $1,832likely 16–62 orders
Net cost
−$170$681 contribution missed − $851 Meta spend saved

Missed orders use this brand's own estimate of what Meta → Amazon adds, with its range; no other brand's data is used. A negative net cost means pausing saves more than it loses.

Calendar and stock check

  • During the test: Holdout test exp-mock, plus 14 days washout · Nov 2 – Dec 13Two holdouts at once pause ads in overlapping states and blur each other's results. Start after this one and its washout end.
  • During the test: Cold Brew Mocha 2 lb likely out of stock at FBA until a new shipment lands · Nov 7 – Dec 9Days without stock give ads nothing to sell in either group, which weakens the test and can hide the effect. Restock first or start after the shipment lands.
  • During the test: Black Friday week through Cyber Monday · Nov 19 – Nov 30Order volumes, ad prices and shopper behavior all shift, so the result wouldn't describe a normal month. Start after it, or end before it.
  • During the test: December holiday shopping · Dec 1 – Dec 24Order volumes, ad prices and shopper behavior all shift, so the result wouldn't describe a normal month. Start after it, or end before it.
  • In the baseline: Holdout test exp-mock, plus 14 days washout · Jul 20 – Aug 30Ads were off in some states then, which distorts the baseline. Start later, or leave those dates out of the baseline.
  • In the baseline: Prime Big Deal Days (usually early to mid October; check Amazon's announced dates) · Oct 1 – Oct 20

Earliest start that avoids the serious ones: Dec 28.

Every option · holdout size against test length
Holdout test options
HoldoutWeeksDetectsOrders missedContribution missedMeta spend savedNet cost
30%410.9%15$255$319−$63.71
30%68.9%22$383$479−$95.57
30%8
States in each group
Ads off (27 states): exclude these from every Meta ad set for the test
AK AR DC DE HI IA ID KS KY LA ME MO MS MT ND NE NH NM NV OK RI SC SD UT VT WV WY
Ads on (24)
AL AZ CA CO CT FL GA IL IN MA MD MI MN NC NJ NY OH OR PA TN TX VA WA WI

Test results

Each result is computed from the locked design and added to the ledger. Earlier results are never edited.

  • Jul 20 – Aug 16 · 30% held out
    Lift in Amazon orders
    7.1%95% range 1.2% to 13.0%
    Orders Meta adds a day
    1.10.2–2.0
    Profit a day after Meta spend
    Needs every product's cost
    Design hash
    9f2c4e1a7b03d5c8…

Prediction track record

Each prediction is locked, with a hash, before its result is known, then scored against what happened.

Track record: no predictions scored yet; the first is due Nov 1.

  • Raise the Meta→Shopify sample ad set (broad) from $16 to $24 a day for four weeks.Scored Nov 1locked Oct 1
    • Shopify first orders per day: predicted 3.1 (80% range 2.4–3.8)
    • Sample → full bag at 30 days (%): predicted 22 (80% range 16–28)

Verifiable ledger

Every test design, result and prediction is chained by hash, so changing or deleting any earlier record would show.

Verified

4 entries, latest Oct 1. Each record still matches the hash recorded for it.

Chain head (SHA-256)
c1d7a3f09e2b48d6a5f3e1c7b9d2a4f6e8c0b3d5a7f9e1c2b4d6a8f0e3c5b7d9
Independent timestamp
Entries 1–4 timestamped Oct 1RFC 3161, by freetsa.org
The download is published with the next measurement.Has every design, result and prediction with the chain and timestamps, plus the steps to check them without Spillmark. Share it with investors, acquirers or agencies.
7.7%
29
$511
$638
−$127
40%410.2%20$341$426−$84.95
40%68.3%29$511$638−$127
40%recommended87.2%39$681$851−$170
50%410.0%24$426$532−$106
50%68.2%37$639$798−$159
50%87.1%49$852$1,064−$212

“Detects” is the smallest lift the test finds 80% of the time at 95% confidence, checked against placebo tests on this brand's past orders. Greyed out: can't see the 7.5% target.