Measurement & analytics.

Exposed audiences surveyed against matched, unexposed controls to quantify what the campaign moved: awareness, consideration, intent, association. Methodology is the whole game: matched cohorts, independent fielding, instruments locked before exposure, and sample sizes powered for the lift you actually care about.

Randomized experiments, run as geo holdouts, audience holdouts or placebo cells, answer the only question that settles budget debates: how many of these conversions would have happened anyway? Design and analysis are locked before launch, read against live spend, and the result is a number you can defend.

What it answers. Causation, not correlation. An incrementality test is the closest advertising gets to a clinical trial: one matched group sees the media, one does not, and the difference is the effect. It settles the arguments attribution cannot, including whether branded search and retargeting are harvesting conversions that were coming anyway.

What it needs. Enough conversion volume for statistical power, a market or audience structure that can be split cleanly, and patience: a well-built geo test typically reads over several weeks. Power calculations run before launch, so a test that cannot reach significance never starts. It does not need user-level identity, which is why the method survived the identity crackdown intact.

What the output looks like. A lift number with a confidence interval, pre-registered so the analysis cannot drift toward the hoped-for answer, plus the cost of the incremental outcome. Reads reconcile against the MMM so the two methods argue with each other in the open.

When to choose it. When one channel or tactic carries enough budget that being wrong about it matters. US marketers rate incrementality testing second on cross-channel measurement reliability, at 34.1% against MMM’s 46.9% (eMarketer/TransUnion, July 2025), and the two work best as a pair: the model allocates, the test validates.

Bayesian models read spend, exposure and outcomes at aggregate level, then estimate each channel’s causal contribution with carryover and saturation built in. No cookies, no IDs, no platform pixel. As click-based attribution thins out, this is the read everything else gets checked against.

What it answers. Marketing mix modeling, also called media mix modeling, is the cross-channel allocator: which channels drove the outcome, at what saturation, with what carryover into later weeks, and what happens to the total if budget moves between them. It reads television, audio and out-of-home on the same footing as clickable channels, which no click path can.

What it needs. Two or more years of weekly spend and outcome history where it exists, or a hierarchical model that borrows strength across markets where it does not. Externalities go in explicitly: seasonality, pricing, distribution, promotion calendars. What it does not need is any user-level identifier, which is why US marketers rate it their most reliable cross-channel read at 46.9% (eMarketer/TransUnion, July 2025).

What the output looks like. Channel-level contribution and response curves, a marginal-return ranking that says where the next dollar works hardest, and scenario planning against the curves. Holdout validation is part of the build, and the full argument for anchoring 2026 plans on it is in our MMM research.

When to choose it. Whenever the plan runs more than two or three channels. The model is the frame the other methods hang off: incrementality tests calibrate it, attribution feeds its within-channel detail.

Credit distributed across the real journey, from the display impression weeks ago to the social click and the branded search, on transparent Shapley-value math instead of a black-box score. Journeys resolve at the household, so the read doesn’t reset every time someone switches screens.

What it answers. The within-channel questions the aggregate models are too coarse for: which creative, which placement, which sequence, which frequency band. MTA is a diagnostic microscope. It stopped being credible as the top-line arbiter when identity fragmented, and last-click MTA now sits bottom of the marketer reliability ranking at 19.4% (eMarketer/TransUnion, July 2025).

What it needs. Dense digital touch data on identity that actually connects, which today means household-level resolution rather than a per-cookie fiction, and honest boundaries: walled gardens report themselves, so cross-platform journeys are stitched with assumptions that should be stated, not hidden.

What the output looks like. Fractional credit across the journey on stated math, refreshed continuously, feeding pacing and creative decisions inside the flight while the slower methods judge the flight as a whole.

When to choose it. For in-flight optimization within addressable channels, always with the caveat attached. We run it reconciled to the MMM rather than as a rival source of truth: when the two disagree, that disagreement is usually the finding.

Viewability says an ad could have been seen; attention measurement quantifies whether it was, using eye-tracking panels, biometric sampling and screen-class-weighted dwell models. It reprices exposure quality across formats and feeds planning the way a metric should: validated against lift studies rather than vendor assertion.