Marketing mix modeling
Long-term contribution, carryover, saturation and marginal return across the whole portfolio. Bayesian MMM with saturation curves, adstock, and competitive and macro controls. Answers: where should the next dollar go?
Measurement & Incrementality
Measurement is a decision system, not a report
Know which media created demand — and which simply harvested demand you already had.
No single number explains a B2B journey
A B2B purchase begins in memory, months before anyone is in market. It moves through search, social, publishers, peers and now AI answers. Six to ten people touch it. It ends in a form, a call, or a procurement process that no pixel ever sees.
Any method claiming to explain all of that in one number is overreaching. Last-click flatters search. Attribution models invent precision they cannot support. Marketing mix models are powerful at portfolio level and blunt at campaign level. Each is useful for a different question, and dangerous when treated as universal truth.
So we don’t pick one. We assign different methods to different questions, and use experiments to check what the models claim.
Four methods
Each answers a question the others can’t
Attribution
Shapley attribution and exposure-to-outcome matching, explaining observable journeys and interaction effects at campaign level. Answers: how do channels assist each other, and where does the journey leak?
Incrementality testing
Randomized holdouts, platform conversion lift, matched-market geo tests and synthetic controls. Answers the only question that really matters: did this media cause an outcome that would not have happened anyway?
Brand & demand studies
Brand lift and demand indicators tracked between commercial readouts. Answers: is mental availability actually moving, and is the message understood?
AI visibility
Presence, citations and recommendation stance in AI answers, fed into the same system rather than isolated in a separate scorecard.
How we measure it →Four connected layers
These methods aren’t parallel reports. They operate on different clocks and support different decisions, feeding one another as evidence accumulates.
Operating
Daily & weekly
Spend, delivery, site, conversion and audience-quality reporting.
Pacing, bids, frequency, exclusions, creative rotation, journey friction.
Diagnostic
Weekly & monthly
Shapley attribution, exposure-to-outcome matching, brand lift, cohort progression.
Channel interaction, sequencing, assists, segment performance.
Causal
Test dependent
Randomized holdouts, conversion lift, matched-market geo tests, synthetic controls.
Incremental customers, revenue and ROAS — and validation of what the models claimed.
Portfolio
Quarterly & annual
Bayesian MMM, saturation curves, adstock, competitive and macro controls, scenario simulation.
Brand and performance ranges, channel allocation, next-dollar planning.
Experiments calibrate the models. Modeled credit is checked against a randomized holdout or a geo test rather than accepted on model fit alone.
How we work
Six principles
-
Business outcomes first
Optimize toward incremental customers, revenue and predicted economic value — not impressions, clicks or form fills in isolation.
-
Method follows question
MMM for allocation, attribution for diagnostics, experiments for causality, brand studies for memory. Never one model asked to do all four jobs.
-
Experiments calibrate models
Modeled credit is validated against randomized holdouts, conversion lift or matched-market tests, rather than trusted because the model fits.
-
Decision before dashboard
Every analysis names the decision it supports, who owns it, the threshold, and the action. Reporting that doesn’t change a decision doesn’t get built.
-
Uncertainty is visible
Readouts carry confidence intervals, assumptions and limitations, and state plainly whether evidence is directional or decision-grade.
-
You own the truth layer
Data definitions, code, features, models and validation history stay transparent and portable. If you change agency, the measurement system goes with you.
The uncomfortable part
Most reported ROAS is not incremental
The gap between what media is credited with and what it actually caused is the single largest source of wasted B2B budget we see. It is also the easiest thing for an agency to avoid testing, because the honest answer is often smaller than the reported one.
- Branded search is the usual culprit. It converts beautifully and frequently causes very little. A geo holdout settles it in weeks.
- Retargeting flatters itself. It reaches people already heading toward you. Incrementality testing separates influence from observation.
- Upper-funnel gets under-credited for the opposite reason, which is why brand effects belong in the same system rather than a separate deck.
- We publish the result either way. A test that shows our own recommendation didn’t work is worth more than one designed not to.
Common questions
What is incrementality testing?
Incrementality testing determines whether media caused an outcome that would not have happened otherwise. It works by withholding media from a comparable group — a randomized holdout, a matched set of geographic markets, or a synthetic control — and comparing results. Unlike attribution, which allocates credit among touchpoints that were observed, incrementality measures causation.
What is marketing mix modeling, and when should B2B use it?
Marketing mix modeling (MMM) is a statistical method that estimates how much each channel contributed to an outcome over time, accounting for carryover, saturation, seasonality and external factors. It is well suited to B2B because it does not rely on user-level tracking and can absorb long sales cycles, but it needs sufficient history and spend variation, and it answers portfolio questions rather than campaign ones.
How is MMM different from attribution?
They answer different questions. MMM works top-down across the whole portfolio over months or years, and is best for allocation. Attribution works bottom-up on observed journeys, and is best for diagnosing sequence, assists and friction. MMM cannot tell you which creative to rotate; attribution cannot tell you whether your total brand investment is too low. Used together, and calibrated by experiments, they cover each other’s blind spots.
What is a matched-market geo test?
A matched-market test pairs geographic markets with similar historical performance, then changes media in one set and holds the other as a control. Because assignment is deliberate rather than observed, the difference in outcomes is a causal estimate. It is one of the few clean ways to measure channels where user-level tracking is weak, which in B2B is most of them.
Can you measure media that drives phone calls or offline sales?
Yes, though not with pixels alone. Calls, pipeline and closed revenue enter the system through first-party data and CRM integration, and the causal layer uses market-level tests precisely because they do not depend on tracking an individual from impression to signature.
Do we keep the measurement system if we leave?
Yes. Data definitions, code, features, model documentation and validation history stay transparent and portable. A measurement system that only works while one agency holds the keys is a commercial lock-in, not a decision system.
Find out what your media actually caused
We’ll walk through your current measurement setup, show you where credit is likely being misassigned, and outline the first test worth running.
Book a 30-minute consultation30 minutes · Virtual · No obligation