The second sale.

The challenge

For a subscription health business, a customer who leaves and comes back is worth close to a customer won for the first time, and costs a fraction as much to reach. Most marketing mix models never see them. Marketing mix modeling measures first orders, the plan optimizes against first orders, and returning subscribers sit in the revenue line with no channel attached to them. When acquisition slows, as it does in every maturing category, the measurement system, and the marketing performance analytics built on it, has nothing to say about the part of the business still growing. This marketing mix modeling case study shows how a direct-to-consumer subscription health brand measured its subscription marketing differently, using MMM to model reactivation as a KPI in its own right, and what the rebuilt models found.

Half the subscription growth sat outside the marketing mix model

The brand’s previous marketing mix model measured paid media against first orders. Returning subscribers were counted in revenue and reported weekly, but no channel carried a cost per reactivation and no coefficient estimated one. Over the same period, approved new customers fell 24% on a 22% reduction in media investment while reactivations rose 60%. The MMM could explain the first movement and not the second. Budget decisions were being taken on a marketing measurement view covering less than half the customers arriving each week.

Contribution to first orders by channel

Previous model, later half, new customers only

% of media spend% of first orders credited

Paid Generic Search

32%

41%

Paid Shopping

11%

17%

Paid Social

9%

12%

VOD

7%

9.5%

Television

19%

8.4%

YouTube

6%

7.1%

Audio

3%

5%

Four measurement gaps in the previous MMM specification

• Win-back had no model. Returning subscribers were counted but never modeled, so no channel carried a cost per reactivation. The plan optimized against acquisition while half the weekly volume came from elsewhere.
• CRM sat outside the measurement frame. Retention email was treated as an operational cost rather than a modeled driver, so no incrementality measurement existed for it and its budget was never argued on returns.
• Above-the-line spend was judged on in-month cost per acquisition. Television and video hold most of their effect for weeks after a flight ends, so the monthly view understated both and the plan cut them first.
• The efficiency spread across the mix was unknown. Blended figures hid a five-fold difference between the cheapest and most expensive parts of the budget, and the single largest line sat in the middle of it, neither defended nor challenged.

Each gap was visible in the previous output only as an unexplained residual. The rebuild addressed all four by modeling three outcomes instead of one, summarized in Table 1.

The solution

MASS Analytics rebuilt the measurement of its subscription marketing as a nested marketing mix modeling structure: three models rather than one. Brand Awareness was modeled first from above-the-line activity, then nested into two commercial models run in parallel, New Customers and Reactivations. Each carried its own channel coefficients, response curves and cost per customer, so a channel could be weak on first orders and still earn its budget on win-back. The build spans two and a half years of data, with a flat approval rate applied so cost per acquisition reflects patients actually treated.

Customer reactivation entered the nested model as a first-class KPI

The Reactivations model took the same channel set as the acquisition model and added the drivers that apply only to a returning customer: CRM send volume split by campaign type, discount depth against the category, average order value, and a trend term standing for the growing pool of lapsed subscribers. Every paid channel therefore produced two numbers, a cost per new customer and a cost per reactivation, and the two rarely ranked the same way. Brand Awareness entered both commercial models as a nested input, which is what nested modeling is for: crediting above-the-line activity upstream as well as downstream, on a log-linear transformation. Adstock and saturation were estimated separately in each model, because a channel’s memory differs depending on whether it is reaching someone for the first time or reminding someone who has already bought.

MMM data preparation made the second model possible

Weekly subscriber data covered two and a half years, split at source into first orders and returning orders so the two KPIs could be modeled against the same media and the same externals without double counting. CRM send volumes arrived by campaign type rather than as a single total, which is what allowed retention activity to be separated from acquisition messaging. Pricing carried both average order value and discount depth weekly, and category demand came from external trackers. The split at source is the part of MMM data preparation that cannot be fixed later: a second model is not possible if the outcome arrives as one number. MMM validation ran on both commercial models independently, since a shared input does not imply shared fit.

Results and Impact

The rebuilt models produced four findings that changed the following year’s plan. Reactivations grew 60% half on half while approved new customers fell 24%, and for the first time the MMM could say where that growth came from: media delivered 44% of it on an incremental basis.

Retention CRM went from noise to a measurable driver

In the earlier half, modeled CRM contribution to reactivations was roughly 300 customers, small enough to treat as noise. In the later half it passed 1,800, a sixfold increase and around 5% of all reactivated subscribers. Send volume roughly doubled over the same period, but contribution rose several times faster, which points to the campaign-type split rather than the send count. Because CRM had never been modeled, none of this was arguable internally. In context: the lapsed pool and category demand together contributed close to twice the net movement, competition and pricing took back most of that, and media and CRM delivered a little under half the growth between them.

$0.9M delivered 23% of all customers

Sorting every channel by blended cost per customer, across first orders and reactivations together, produced a better than fivefold spread. The cheapest tier, $0.9M or a tenth of the half-year budget, delivered 5,500 customers at $164 each. The most expensive tier, 26% of the budget, delivered 2,700 at $867. Generic paid search, the largest single line at 32% of spend, sat in the middle. Measured across a full carryover window rather than in-month, television and video each moved up a tier. The two-step marketing mix optimization that followed moved money inside the existing budget and protected the brand pool first, so conversion channels could not out-bid it on marginal ROI.