A fictional retailer spends heavily across paid search, social and email. Which channels actually caused the growth?
12 min read
The Question
Of the £2m spent across paid search, paid social, display and email in 2025, how much revenue was actually caused — not just correlated?
The Hypothesis
Last-click attribution will materially overstate paid search and email, and understate upper-funnel channels. We expect ~30% of reported "paid" revenue to be incremental.
The Data
78 weeks of channel spend, impressions and revenue (2024–2025), joined to 14 geo-holdout experiments and a synthetic control group.
The Analysis
fit a bayesian media mix model with adstock and saturation per channel, validated against four geo-holdout experiments. compared marginal roas to the platform-reported roas to size the gap.
The Decision
Cut display by 40% and reallocate to paid social prospecting. Keep brand search flat. Pause two underperforming affiliates.
The Outcome
Over the following two quarters, modelled incremental revenue rose 11% on flat total spend.
Further reading
Most marketing teams know their dashboards lie a little. This investigation is about how much.
A fictional UK retailer — let's call them Northwind — spent just over £2.0m on paid media in 2025. Their dashboard said the campaigns drove £8.4m in revenue. Their finance team said total revenue grew £3.1m. Somewhere between those two numbers is the truth.
reported channel mix (last-click)
Spend and last-click revenue in £000s. The dashboard view.
the question
If we turned each channel off tomorrow, how much revenue would actually disappear? That number — the incremental contribution — is what the budget should be sized against. Last-click attribution answers a different question: which touchpoint happened to be closest to checkout.
the approach
Three methods, run independently, then reconciled:
Media mix model (MMM) — Bayesian regression on 78 weeks of spend, impressions and revenue, with adstock and saturation curves per channel. Built in PyMC, sense-checked against Google Meridian.
Geo holdouts — 14 paired-region experiments across paid social, display and affiliate. Two-week on/off cycles.
Synthetic control — for the brand-search ablation we couldn't run live, a synthetic control built from 9 comparable retailers.
the contradiction
Once the three views were on the same axes, the gap between reported and incremental was hard to unsee.
reported vs incremental revenue
Incremental = MMM posterior median, reconciled with geo-test lift. £000s.
Brand search was the biggest single overstatement: ~70% of the revenue it "drove" would have happened anyway. Email was a close second — most of the audience were people who'd already decided to buy. Display was almost entirely retargeting people who would have converted regardless.
The interesting result was paid social. Almost none of it was last-click — but the geo tests showed it was doing real work upstream.
the lift curve
Geo holdouts let us trace what happens when you actually stop spending.
geo holdout: paid social pause
Daily revenue per 1,000 households. Pause begins day 0.
Within a week, revenue per thousand households in the test cells had dropped ~22% relative to controls. The decay was steady, not immediate — which matches a brand-and-consideration channel, not a direct-response one.
The display test ran the opposite way:
geo holdout: display pause
Test cell revenue lift vs control, %. Flat = no causal effect.
Two weeks without display advertising. No measurable effect on revenue.
return on what's actually working
Once you index spend against incremental revenue, the league table inverts.
Email's iROAS looks great until you remember its absolute size is small. Brand search is the inverse: a huge line item with a tiny incremental return.
the weekly story
Stitching the model back together across the year shows where the noise was.
weekly incremental revenue by channel
Modelled incremental £000s per week, 2025.
what changed
The team didn't slash the budget. They moved roughly £400k away from display and brand search, into paid social prospecting and a small creative testing fund. Two underperforming affiliate partners were paused; one was kept on a fixed-fee deal instead of CPA.
A quarter later, total spend was flat. Modelled incremental revenue was up 11%. The single line that mattered most was an internal one: finance and marketing now agreed on which number to look at.
what to take from this
Dashboards measure what was seen. Decisions need what was caused.
Three methods that disagree are more useful than one that you trust.
Most reallocation wins come from finding the £180k line item that does nothing — not from picking the next big bet.