Marginal CAC, attribution and incrementality

Theo

I could put up to $4,000 a week into paid social. In August I raised the budget week by week to see what happens.

Alex Tutor

Good experiment. Every channel has a channel capacity: the cheapest users come first, so each extra dollar buys fewer. Look past the average and ask what the last $1,000 bought.

Alex Tutor

Dataset · 4 rows

Paid-social budget ramp, August 2026

One week at each spend level; installs as reported by attribution.

Download CSV

Columns
week: A week of August 2026
weekly_spend_usd: Paid-social spend that week, USD
paid_social_installs: Installs attributed to paid social that week

Data

week	weekly_spend_usd	paid_social_installs
2026-08-w1	1000.00	1149
2026-08-w2	2000.00	1815
2026-08-w3	3000.00	2291
2026-08-w4	4000.00	2683

You

Average CAC goes from 87 cents at $1,000 to $1.49 at $4,000. That doesn't look too bad.

Alex Tutor

The average hides the edge. The step from $3,000 to $4,000 added only 392 installs. That's the marginal CAC: extra spend divided by extra installs.

Alex Tutor

Question 1

What was the marginal CAC of the last step of the ramp, from $3,000 to $4,000 a week? Round to cents.

Your answer

Units: USD, e.g. 1.25

Alex Tutor

Now compare cost with value. Paid-social users make $1.46 of gross profit in their first 12 months, less than Halves's average, because fewer of them subscribe.

Theo

So $1.46 is my ceiling per install?

Alex Tutor

Not quite. Each paid user also sets off a chain of invites: 0.31 + 0.31² + … = 0.31 ÷ 0.69, about 0.45 invited users over all generations. And don't assume what those are worth: the M7 cohorts split by channel show invite users made $1.52 of gross profit in 12 months, a little more than paid-social users.

Alex Tutor

\text{break-even CAC} = \text{LTV}_{12}^{\text{paid}} + \frac{k}{1 - k} \times \text{LTV}_{12}^{\text{invited}}

Invitees are valued at the measured 12-month LTV of invite users (\$1.52 against \$1.46 for paid social), not assumed equal to paid users.

You

And stopping at 12 months keeps it cautious. If invited users had turned out worth less, we'd plug in the smaller number.

Alex Tutor

Question 2

Paid-social users' 12-month LTV is $1.46, invite users' is $1.52, and k is 0.31. What is the break-even CAC per paid install, counting the invited users it brings over all generations? Round to cents.

Your answer

Units: USD, e.g. 1.25

Alex Tutor

Against $2.14: the second $1,000 cost $1.50 per install, well inside the line. The third cost $2.10, only just inside; the fourth $2.55, above it.

Theo

Store search ads cost more per install than paid social did at $1,000. Should I drop them?

Alex Tutor

Their users subscribe more, so their value is higher too. A channel fits when its marginal cost sits under its own users' value, not under the cheapest channel's cost.

You

So each channel gets its own ceiling and its own curve.

Alex Tutor

Question 3

Judged on the ramp's attributed installs and the $2.14 break-even, which weekly paid-social budget does the data support?

Choose one answer$1,000: anything above it raises the average CAC$2,000: the second $1,000 costs $1.50 per install, well under the line$3,000: the third $1,000 costs $2.10 per install, just under the line; the fourth costs $2.55$4,000: even the full budget has an average CAC under $2.14

Maya

How do we even know those installs came from the ads?

Alex Tutor

On phones, attribution now comes from the app stores' privacy-preserving frameworks: counts per campaign, aggregated, a day or more late, with no user-level trail.

You

So we can't join an install to that person's later payments, or build an LTV curve per ad.

Alex Tutor

Not per user. And attribution only says an ad came before an install, not that the install needed the ad.

Alex Tutor

Question 4

Which statements about attribution through the app stores' privacy-preserving frameworks are true? Select all that apply.

Choose all that applyResults arrive aggregated per campaign and with a delayEach attributed install can be joined to that user's later subscription paymentsAn attributed install is not proof that the ad caused itA geo holdout can check how many attributed installs the ads really addedIf attribution counts are right, a holdout test is unnecessary

Alex Tutor

To learn what ads add, turn them off somewhere. In September paid social kept running in four regions and paused in four similar ones. Before that it ran everywhere.

Alex Tutor

Dataset · 8 rows

Geo holdout test, paid social, September 2026

Eight regions; four weeks before the test and four test weeks.

Download CSV

Columns
region: Region
group: test: ads kept on; holdout: ads paused during the test
installs_pre_4w: All installs in the 4 weeks before, ads on everywhere
installs_test_4w: All installs in the 4 test weeks
paid_social_spend_test_usd: Paid-social spend in the test weeks, USD
attributed_paid_installs_test: Installs the store frameworks attributed to paid social in the test weeks

Data

region	group	installs_pre_4w	installs_test_4w	paid_social_spend_test_usd	attributed_paid_installs_test
R1	test	1480	1536	622.00	657
R2	holdout	1390	1020	0.00	0
R3	test	1210	1247	508.00	537
R4	holdout	1260	920	0.00	0
R5	test	980	1010	412.00	435

Show 3 more rows

R6	holdout	1050	762	0.00	0
R7	test	1330	1379	559.00	590
R8	holdout	1300	949	0.00	0

You

The holdout regions fell to 73% of their earlier installs. That's what the test regions would have done without ads.

Alex Tutor

Right. Expected installs without ads are the test regions' earlier installs times that ratio. What's left above it is the incrementality of the ads.

You

And the attributed count is what the frameworks gave the ads credit for.

Alex Tutor

Question 5

Use the geo test. What was paid social's incremental CAC: test-region spend divided by the installs above what the test regions would have had without ads (their pre-period installs × the holdout regions' test ÷ pre ratio)? Round to cents.

Your answer

Units: USD, e.g. 1.25

Maya

Attribution said 2,219 installs, the holdout says about 1,521. So does $3,000 a week still hold?

Alex Tutor

If only 69% of attributed installs are extra at higher budgets too, the second $1,000 costs about $2.19 per extra user, just above the line, and the third about $3.07, far above. The geo test ran at a lower spend, so that's a guess to test, not a fact.

Theo

So what do I ask for?

Alex Tutor

Question 6Short answer · AI-checked task

Write Theo's budget recommendation in 3–5 sentences: what weekly paid-social budget to run this autumn, why, and what to test next.

Write your answer and get a score with feedback from our AI reviewer.

Log in to get AI feedback

The rule: judge a channel at the margin, against its own users' value including invites, and measure what it adds with a holdout. Attribution tells you who was there; a holdout tells you who came because of it.

Alex Tutor

That’s the lesson. You answered every task — nicely done.

Next lesson: Capstone: the unexplained spike