Decide whether to move $20,000 a month on a last-click report, recompute social's cost per order under a stated assumption, and name the test that settles it.
critiquehardExercise · 3 min · 32 of 60
On this page
An agency wants to move $20,000 a month on the strength of a last-click report, and the report contains its own counter-evidence. Attribution and its limits explains why the channels nearest the order always look cheapest.
A consumer brand spends $125,000 a month: $60,000 on paid social, $40,000 on paid search ($25,000 of it on brand keywords) and $25,000 on display retargeting. The last-click report for last month's 2,400 orders:
Channel
Share
Orders
Spend
Cost per order
Paid search
48%
1,152
$40,000
$35
Retargeting
22%
528
$25,000
$47
Paid social
18%
432
$60,000
$139
Organic and direct
12%
288
—
—
The agency proposes moving $20,000 from social to retargeting and brand search on those costs per order. Two more facts: 70% of retargeting conversions follow a site visit that came from social within 24 hours, and brand-search volume rose 40% in the quarter social spend doubled.
No. Retargeting fires only after a site visit, and brand search only after someone already knows the brand, so last-click credits both by construction: it rewards the step nearest the order, not the cause. eBay's experiment found that brand-keyword clicks mostly arrive anyway through the unpaid link. And both extra facts point at social: 70% of retargeting's buyers came from social within the previous 24 hours, and brand searches rose 40% when social spend doubled.
Recompute under one stated assumption. Search orders split in proportion to spend, so brand search has 1,152 × 25/40 = 720 orders. If half of retargeting's orders (264) and half of brand search's (360) are downstream of social, social has 432 + 264 + 360 = 1,056 orders, and its cost per order is $60,000 ÷ 1,056 ≈ $57, not $139. Retargeting's rises to $25,000 ÷ 264 ≈ $95. The half is an assumption, and that is the point: this report cannot say which figure is right.
To the agency: the budget holds for four weeks while a test decides. Run a four-week geo holdout on social in 20% of regions, with brand search also paused in half of those, and count orders in our own data, not the platforms'. Keep social's budget if the test puts its incremental cost per order under $47: retargeting's last-click $47 almost certainly flatters it, so social beats it on any reading. Above $47, test retargeting the same way next, and move the $20,000 only where an incremental cost per order comes in below social's.
Keep the last-click report for what it does well: comparing creatives within each channel and pacing daily spend.
"Attribution is always wrong; build a mix model." A mix model needs two or more years of weekly data and cannot resolve a four-week decision; it also ignores the two facts in the situation that already point at social.
"Reject it and move $20,000 into social instead, since its real cost is $57." The $57 rests on the "half" assumption, so this swaps one unmeasured belief for another. The recomputation shows the report is unreliable; only the test says which way to move money.
"Accept it, but cap retargeting's frequency." It treats the costs per order as measured and tunes the channel the report flatters, while cutting the one feeding it.
