When platform attribution, incrementality tests and marketing mix modeling point in different directions, the wrong response is to crown one universal winner. These methods often answer different questions. The practical response is to name the decision first, then design the video variants, identifiers, evidence and action threshold around it.
On October 20, 2026, IAB released its Cross-Channel Measurement Decision Guide for public comment through November 4. The draft organizes attribution, incrementality and MMM around decisions and explicitly addresses conflicting results. It is a consultation draft, not a final industry standard. It shows how one industry body is trying to make measurement more decision-led; it does not prove the market has already converged on one method.
For international brands working with Chinese supply chains and distributed marketing teams, the production implication is immediate: a video asset should not merely be ready to run. It should be a stable, comparable unit that can carry evidence back into the next production decision.
Why the three methods can disagree
Attribution assigns credit under a reporting rule
Attribution is useful for daily operations. It can compare click-through rates, conversion paths, audiences and placements, then help teams stop obviously weak executions quickly.
But attribution depends on windows, matching, cross-device visibility and the rules used to assign credit. It tells a team which touchpoint received credit inside that system. It does not, by itself, establish that the conversion would not have happened without the ad. Platform ROAS can support pacing and optimization; it should not automatically be presented as proof of creative incrementality.
Incrementality estimates what changed against a counterfactual
Incrementality experiments compare an exposed or treated unit with a control or counterfactual. They are better suited to questions such as: did this media-and-creative package create additional sales, leads or search behavior?
Experiments have boundaries too. User-level studies require stable assignment and sufficient data. Geo experiments can work across publishers, use aggregated data and help calibrate MMM, but they often bring more noise, longer durations and real opportunity cost. Google's publisher-agnostic Meridian GeoX documentation supports holdback, go-dark and heavy-up designs; its FAQs also stress power analysis, one primary KPI per run and sufficient geographic units.
MMM informs portfolio allocation
MMM examines the relationship between channels, spend, time and business outcomes. It is better suited to quarterly or annual allocation, channel mix and longer-term effects. IAB's Modernizing MMM guidance treats experiments as causal validators and attribution as a map of the funnel—not as interchangeable substitutes.
An MMM result may support increasing short-form video investment overall. It will rarely have enough creative granularity to decide whether the product, a person or a problem statement belongs in the first 1.5 seconds. Asking a portfolio model to choose a hook is like asking an annual financial statement to choose an edit point.
The expensive mistake happens before the cameras roll
Many teams postpone measurement design until after the campaign. Production changes the talent, hook, location, caption, offer and duration at the same time, then media asks which element worked. Even a large performance difference cannot identify the cause.
Asset identity is another hidden failure. The same master is renamed across languages, aspect ratios, markets and platforms. Production, media and analytics use different codes. The organization has plenty of data, but no reliable path back to the original script, shot or usage right.
This is where a production partner supporting a global launch content system should enter measurement design. The role is not to replace the statistician. It is to make comparability part of the production architecture.
The ONCE five-layer framework
We use a simple sequence: Decision–Variable–Unit–Evidence–Action.
1. Decision: write the resource choice in one sentence
Do not begin with “measure video performance.” Begin with a decision that can change work or spend:
- If product-first openings keep winning under matched delivery, allocate more of the next shoot to demonstrations.
- If a geo incrementality test misses the pre-agreed threshold, pause the media scale-up rather than changing only the attribution window.
- If MMM and experiments support a channel but individual assets fatigue quickly, increase the supply of variants instead of cutting the channel.
Without a decision sentence, measurement becomes retrospective explanation.
2. Variable: change one primary production variable per batch
Controllable variables include the opening hook, time to product, proof mechanism, demonstration order, type of speaker, call to action and duration. Each test batch should have one primary variable while the other material conditions remain as stable as practical.
Commercial production will never behave like a laboratory. The useful discipline is to separate the variable changed on purpose from execution conditions that had to change. That is more valuable than delivering ten visibly different videos whose differences cannot be interpreted.
3. Unit: give every asset a stable identity
Each delivered version needs a persistent internal asset ID linked to its master, language, market, aspect ratio, duration, hook, test variable, release date and usage rights. Platforms may rename the file or ad, but the internal ID should survive.
The point is not tidy naming. It is the ability to trace a result back to a production choice: which version, what changed, where it ran, how much exposure it received and which rights applied.
4. Evidence: choose the method for the question
A practical evidence ladder looks like this:
- To choose between two specific executions, run a controlled creative A/B test under matched conditions.
- To determine whether a media-and-creative package produced net-new outcomes, use a lift or geo experiment.
- To allocate across channels and time, use MMM and calibrate it with suitable experimental evidence.
- When the methods conflict, first check the question, time window, KPI, unit of analysis and population before commissioning another test.
Google's Meridian calibration documentation is explicit that experiments and MMM can estimate different quantities. Recency, duration, geography and population coverage all matter. A disagreement is first a diagnostic clue, not an instruction to choose sides.
5. Action: define the threshold before results arrive
Agree in advance which result will trigger which action. The threshold does not need to masquerade as a universal industry benchmark; it should reflect the brand's economics, risk tolerance and available sample.
At minimum, the action field should state whether the team will continue, scale, commission another test, reshoot, hold or stop—and name an owner and review date.
Three decisions require three production designs
Choosing an execution requires comparable versions
If the question is product-first versus person-first, the audience, offer, duration, aspect ratio and core claim should remain as comparable as possible. The shot list should identify fixed shots and variable shots. The edit template should concentrate the difference in an interpretable place.
Testing incrementality requires a renewable experiment kit
If the question is whether the advertising created a net-new outcome, production, media and analytics need to align before launch. The experiment may last longer than one asset remains fresh. A single hero film is therefore fragile. Build replacement hooks, ratios and language versions around the same claim so creative fatigue does not turn a causal test into a judgment on one aging edit.
Allocating a portfolio requires an asset ledger
MMM does not need every edit point as a model feature, but it does need stable data on channel, market, time and spend. Production should classify assets by proposition, content type, market, language and active period so a model result can inform next quarter's content supply—not just a media spreadsheet.
Put measurement fields into the production scope
A production brief or request for proposal can require five operational deliverables:
- A one-page test manifest with the decision, primary variable, constants, KPI and evidence method.
- A version matrix with asset ID, master, market, language, ratio, duration and rights.
- A media map connecting internal asset IDs to platform ad IDs.
- A freeze rule defining which assets cannot change once a formal experiment begins.
- A result-feedback session that writes wins, losses and uncertainty back into scripts, shots and the asset library.
These fields cannot guarantee a valid experiment. They do reduce the common failure in which the team has performance data but cannot use it to improve the next shoot.
When not to commission a large experiment
If a brand has low traffic, sparse conversions or too few geographic units, a costly geo experiment may be underpowered. A better sequence is to fix event definitions, asset identity and version control; use smaller creative tests to find obvious differences; then decide whether a causal study deserves the budget.
When attribution, incrementality and MMM disagree, do not immediately blame a platform or model. Return to the five layers. Are they informing the same decision, observing the same variable, identifying the same unit, covering comparable people and time, and leading to the same action?
The next upgrade in video production is not simply to hand more shots to an algorithm. It is to make every shoot leave behind evidence that improves the next decision. Which method should you trust? The one designed to answer the decision you actually need to make.
Sources and boundary
This article draws on IAB's October 20, 2026 public-comment announcement, the draft Cross-Channel Measurement Guide, IAB's Modernizing MMM guidance, and Google's documentation for Meridian GeoX, GeoX FAQs and ROI priors and calibration. The IAB cross-channel guide was still a consultation draft at publication and may change. Decision–Variable–Unit–Evidence–Action is an ONCE production decision framework, not an official IAB or Google model.
