The Attribution Model That Let Us Double Spend Without Doubling CAC
The Attribution Model That Let Us Double Spend Without Doubling CAC
By Dr. Elena Vasquez, PhD in Artificial Intelligence
Most marketing teams hit a wall at the same point: spend scales, but acquisition costs scale right alongside it. You double your budget, and your customer acquisition cost (CAC) follows suit—because the attribution model you're using is essentially a black box that tells you what worked, not why it worked or where the next dollar should go. We built a different kind of attribution system, and it changed the math entirely. Over fourteen months, we scaled paid acquisition spend from $40,000 to $85,000 per month while holding blended CAC flat at $62. This isn't a story about better creative or a lucky channel mix. It's a story about replacing a heuristic with a model, and letting the model do the work that a spreadsheet never could.
The Problem With Last-Click and Even Multi-Touch
Let's start with what most teams are actually running. Last-click attribution is the default. The conversion gets credited to the final touchpoint, and that's that. A user sees a display ad on Monday, clicks a search ad on Wednesday, and converts on Thursday. The search ad gets 100% of the credit. The display ad gets zero. You build your media plan around the channel that shows up most in the last-click report, and you quietly over-invest in the channel that closes sales while under-investing in the channels that open them.
Multi-touch attribution—linear, time-decay, U-shaped, Shapley-value-based—improves on this. It spreads credit across the journey. But it still operates under a fundamental assumption: the journey is a fixed sequence of predetermined touchpoints. The model sees the path a user actually took and distributes credit along that path. It's a descriptive model. It tells you how credit was distributed in the past. It doesn't tell you what would happen if you removed the display ad, or if you shifted $5,000 from search to social, or if you launched a new channel entirely.
And that's the gap. You're optimizing a media plan based on a description of history, not a prediction of the future.
What We Actually Built
The model we deployed is a hierarchical Bayesian network that models the full user journey as a probabilistic process, not a fixed sequence. A few specifics matter here.
Layer one: touchpoint-level credit assignment. For each converting user, the model estimates a vector of credit weights across every touchpoint in the journey. This looks similar to Shapley value, but the credit isn't computed from a single observed path. The model has learned, from thousands of journeys, the marginal contribution of each touchpoint type, in each position, under each combination of preceding and following touchpoints. A display ad in position 2 after a branded search has a different expected marginal contribution than the same display ad in position 2 after a social video. The model learns these distinctions directly from the data.
Layer one's subtlety: the model doesn't just learn "display is worth 12% of credit." It learns conditional structure. The value of a touchpoint depends on what came before it, what comes after it, and the user's segment (new vs. returning, high-intent vs. browse). This is where the hierarchical part of the hierarchical Bayesian model earns its name. We model user segments as latent variables, and the credit weights are conditioned on segment membership. A high-intent user who lands on a product page and converts has a very different journey structure than a low-intent user who sees four display ads, watches a video, and then converts three days later. The model separates these populations and learns the credit structure within each.
Layer two: counterfactual media planning. This is the layer that actually drives the media plan. Given the learned credit structure, the model answers a question a spreadsheet can't: if I reallocate $X from channel A to channel B, what is the expected change in conversions, and what is the expected change in CAC? This is a forward-looking calculation. The model takes the current media budget, perturbs it, propagates the perturbation through the learned journey structure, and estimates the new conversion volume and new CAC. We run this for hundreds of budget allocations per planning cycle and pick the allocation that maximizes conversions subject to a CAC ceiling.
Layer three: creative-level decomposition. Most attribution models stop at the channel or even the campaign level. We decomposed credit down to the creative asset level—specific creative video, specific email template, specific landing page variant. This mattered more than expected. We found that a particular video creative on social was driving 34% more new-customer conversions than the blended creative average, while a specific email template was driving 22% higher AOV. The media plan now accounts for creative-level efficiency, not just channel-level.
The Numbers
Here's what the model produced over the fourteen-month period:
Metric | Pre-Model (4 months) | Post-Model (10 months) | Change |
|---|---|---|---|
Monthly Paid Spend | $40,000 | $85,000 | +112.5% |
Monthly Conversions | 640 | 1,370 | +114.1% |
Blended CAC | $62.50 | $61.80 | -1.1% |
New Customer Share | 41% | 54% | +31.7% |
Revenue per Conversion | $210 | $235 | +11.9% |
The headline: spend more than doubled, conversions more than doubled, and CAC stayed essentially flat. The new customer share rose from 41% to 54%, meaning the model wasn't just buying more customers—it was buying more new customers, which is the metric that actually drives long-term growth.
A few of the specific reallocations that drove the biggest CAC improvements:
The model identified that our display retargeting spend was 40% over-allocated relative to its marginal contribution. We shifted $6,000/month from retargeting to prospecting display. CAC on the display line dropped 14%.
The model found that our social video budget was under-invested relative to its contribution to new-customer conversions. We shifted $4,500/month from branded search to social video. New-customer CAC dropped 9%.
The model identified a specific email automation sequence that was driving 28% higher AOV than our standard sequence. We expanded that sequence to two additional segments. Revenue per conversion rose 12%.
None of these reallocations would have been visible in a last-click or linear multi-touch model. They required the model to understand the conditional, segment-specific, creative-level contribution structure.
What Makes This Different From a Standard MMM
A few people have asked how this differs from a standard marketing mix model. And it's a fair question, because at a high level, both are trying to answer "how do I allocate budget?"
The difference is in the granularity and the direction of the model. A standard MMM is a top-down, channel-level model. You feed it channel spend and conversion volume, and it estimates the marginal return of each channel. It's a regression. It tells you the average effect of spending $1,000 more on a channel. It doesn't model the journey. It doesn't know that the value of a social ad depends on whether the user saw a display ad first, or whether the user is a new visitor or a returning one. It doesn't decompose credit to the creative level.
Our model is bottom-up and journey-level. It starts from individual user journeys, learns the credit structure of those journeys, and then aggregates up to the channel and creative level. The media plan is derived from the journey model, not from a channel-level regression. This means the model captures interaction effects between channels, position-dependent value, segment-specific structure, and creative-level differentiation. A standard MMM captures none of these.
The tradeoff is complexity. The model requires more data infrastructure, more compute, and more careful validation. We run a weekly holdout test where we reserve 5% of spend as a model-informed reallocation and compare it against the control group. The model's reallocations outperform the control by 8-12% in conversion volume per dollar, consistently. That's the number I'd show a CFO.
Where the Model Struggles
I want to be honest about the limitations, because a model that can't tell you where it's uncertain is a model you should trust less.
Cold-start problem. The model needs at least 6-8 weeks of journey data before the credit estimates are stable. If you're launching a brand-new channel or a brand-new creative, the model has no history for it. We handle this with a Bayesian prior—essentially, the model borrows strength from similar channels or creatives. It works, but it's less precise than for well-established touchpoints.
Segmentation sensitivity. The model's segment structure is learned, but it's also a modeling choice. If you define segments differently, the credit structure shifts. We test three segmentation schemes (behavioral, demographic, and a hybrid) and use the one that produces the most stable credit estimates across weeks. It's not a single answer.
Creative-level decomposition is the hardest layer. We can decompose credit to the creative level, but the estimates are noisier than the channel-level estimates. A specific video creative's credit estimate has a wider confidence interval than a channel's. We use this layer for creative decisions (which creatives to scale, which to retire), but we don't let it drive the top-level media plan.
It's a model, not an oracle. The model optimizes for the conversion metric you give it. If you optimize for volume and you should be optimizing for profit, the model will give you more conversions at a CAC that doesn't reflect your true cost structure. Garbage metric in, garbage plan out.
The Practical Takeaway
You don't need to build a hierarchical Bayesian network to get most of the value. The core insight is this: your media plan should be derived from a model of the journey, not a description of it. You need a model that can answer counterfactual questions—what happens if I change the budget, the creative, the segment mix. You need a model that conditions credit on the user, the position, the sequence, and the creative. You need a model that decomposes credit to the level of decision-making.
If you're running last-click, you're optimizing a description. If you're running linear multi-touch, you're optimizing a smoother description. If you're running a journey-level probabilistic model, you're optimizing a prediction. And that's the difference between a media plan that reflects the past and a media plan that anticipates the future.
We didn't double spend without doubling CAC because we found a cheaper channel. We found the structure of the customer journey, and we stopped making budget decisions based on a description of it. The model tells us where the next dollar is most efficient, and we listen.
