CAC Apportionment Model
DTC and Amazon CAC with TV and Meta spend apportioned by incrementality tests
Rolled up vs fair view
Total spend unchangedChannel CAC on DTC
Agreed view (test-calibrated)| Channel | Spend | To DTC | To Amazon | DTC customers | DTC CAC |
|---|---|---|---|---|---|
| Meta | $180,000 | $162,000 | $18,000 | 1,900 | $85.26 |
| Paid search | $60,000 | $60,000 | $0 | 700 | $85.71 |
| TV (linear and CTV) | $150,000 | $75,000 | $75,000 | 1,500 | $50.00 |
| Podcast | $40,000 | $40,000 | $0 | 400 | $100.00 |
| Amazon sponsored ads | $45,000 | $0 | $45,000 | Amazon | $46.00 |
| Organic and direct | $0 | 700 | Baseline |
Each source weighted, then scaled to match incrementality tests. The one view everyone signs off on.
How This Model Is Built
Spend from every ad platform and TV, orders from Shopify and Amazon, subscriptions from Recharge, and discount codes and survey answers land in one warehouse. Each attribution source is kept side by side. The agreed view and the TV to Amazon split are rules everyone signs off on, re-set after each incrementality test.
Frequently asked questions
How do you apportion TV spend to Amazon?
Run an incrementality test, such as a geo holdout, and measure how much Amazon new-to-brand sales rise where TV runs. Move that share of TV spend into Amazon CAC, and keep the rest on DTC. Re-test on a set schedule so the split stays informed by data.
Why does Amazon CAC look so cheap?
If you divide only sponsored ads spend by Amazon new customers, every customer that TV, podcast, or social sent to Amazon looks free. Amazon CAC goes up and DTC CAC goes down once that halo is credited to Amazon.
What is marginal CAC?
Marginal CAC is what the next dollar buys: extra spend divided by the extra customers it brings. It rises as a channel saturates, so it can be far above the average CAC for the same channel.
Which attribution model should we use?
Keep several side by side: platform reported, last click, discount codes and survey, and a test-calibrated view. Agree on one as the official view, and use the others as filters to see how much the answer depends on the model.