Most marketing teams now own more than one measurement method. In BCG's 2025 survey, 46% of surveyed marketers reported using marketing mix modeling for planning, incrementality tests for causal insight and multi-touch attribution for daily optimization, all at once. That is a survey self-report from mid-2025, not a market census, but it points at a practical problem: having three methods is common, and having a plan that connects them is rare. This guide is that plan. It walks a CMO through five steps: an inventory of the decisions marketing actually makes, a rule for assigning each decision to a method, a sequence of experiments that keeps attribution and MMM honest, a data readiness check, and an ownership and review cadence that marketing and finance agree on before any number is produced.
GPI's position is that measurement debates usually start in the wrong place. Arguing whether attribution or MMM is more accurate ignores that accuracy only matters relative to a decision someone has to make and how much is riding on it. Start by writing down what you will reallocate, how often and with how much money at stake, then pick the method that can answer at that speed with that confidence. Attribution can rank creatives; it cannot prove a channel caused sales. An experiment can prove causation for one channel and period; it cannot plan a year. MMM can plan, but it needs experiments to stay honest. The same test applies to any partner's claims: which decision was the number built for, what method produced it, and what can it not establish.
Why a Marketing Measurement Strategy Starts With Decisions, Not Methods
Reporting tells you what happened; measurement tells you what marketing caused
A dashboard that shows spend beside conversions is reporting. Measurement, as LatentView frames it, is the work of connecting campaigns, channels and spend to business outcomes such as revenue, pipeline and acquisition, and answering whether marketing caused those results rather than merely coincided with them. That distinction is the whole reason a strategy is needed. Anyone can read a report; a causal claim has to come from a method built to produce it.
A useful first exercise: for each dashboard your team relies on today, write one sentence stating the causal claim it can support and one stating the claim it cannot. Anything that only supports the second sentence is reporting, and it stays out of the decision inventory in Step 1 until a method is attached to it.
The three-method stack is common, but the connective tissue is not
The BCG survey cited above found that nearly half of respondents run MMM, incrementality testing and MTA together, each for a different job. The same report notes that more than 70% of respondents use AI somewhere in their measurement toolkit, yet nearly one in three still name evaluating effectiveness across channels as their biggest challenge. Both figures come from one survey sample, so treat them as directional. Read together, they suggest that tooling has moved faster than the operating model around it. Teams have the methods; what they lack is the agreement on which method answers which question, in what order, and who acts on the answer.
What a roadmap adds that a method comparison cannot
If you still need the conceptual comparison of attribution and marketing mix modeling, GPI's decision guide on that topic covers definitions, strengths and limitations, and this article will not repeat it. The roadmap here assumes you know roughly what each method does and focuses on connecting them. It proceeds in five steps: a decision inventory, a method assignment with a reconciliation rule, an experiment sequence, a data readiness diagnostic, and an ownership and review cadence. A short section on durability follows, then guidance on how GPI applies the same standard when reading an agency's measurement claims.
Step 1: Build the Decision Inventory Before Choosing a Method
Open a blank table and list every recurring decision that moves marketing money or creative before you discuss a single method. The inventory is the foundation for everything that follows, and it is the artifact finance will recognize as theirs as much as marketing's.
List the recurring decisions marketing actually makes
Most teams make a small number of decisions repeatedly: the annual budget split across channels, the quarterly reallocation between channels or regions, weekly bid and creative rotation inside a channel, and the occasional go or no-go on a new channel or partner. Each of these is a row. List choices rather than metrics. A row should be something a named person can act on by a date, with a business outcome attached. LatentView's definition is useful as a gate here: if the row cannot be tied to revenue, pipeline or customer acquisition, it is probably a reporting habit rather than a decision.
Record owner, cadence, stakes and the evidence standard each decision needs
For each row, capture who owns the call, how often it is made, how much annual spend it moves, how easily it can be reversed, and what standard of evidence the organization will accept before acting. The last two fields carry the most weight and deserve the most argument. A weekly creative rotation that moves a few thousand dollars and can be undone next week can run on directional platform data. An annual commitment of a large share of budget to a channel that cannot be unwound for months needs causal evidence. Stakes and reversibility, not team preference or tool availability, set the standard.
Supermetrics illustrates why the evidence-standard column exists: a platform report may show a campaign driving 1,000 sales while a large share of those sales would have happened without the campaign. The report is accurate as a count of sales touched by ads; it is silent on how many the ads caused. A row that needs a causal answer cannot accept that report as evidence, however precise it looks.
Separate strategic, tactical and validation decisions
Sort rows into three groups. Strategic decisions allocate money across channels or years. Tactical decisions allocate money or attention within a channel that has already been shown to work. Validation decisions ask whether a channel adds anything at all. The grouping matters because the tactical group can live with correlation, while the strategic and validation groups need something closer to causation. Step 2 will assign methods along exactly this line.
A worked inventory row
The template below shows the fields and one filled row. The row is illustrative, constructed for this article, and does not describe any measured campaign.
| Decision | Owner | Cadence | Annual spend at stake | Reversibility | Evidence standard required | Business outcome named |
|---|---|---|---|---|---|---|
| Reallocate up to 15% of paid social budget between two platforms | Performance lead | Quarterly | Illustrative: the 15% tranche of the paid social line | High: reversible next quarter | Geo or audience holdout on both platforms within the last six months | Incremental first orders |
| Rotate ad creative within an active campaign | Channel manager | Weekly | Low: creative production and a week of delivery | Very high | Platform-reported directional performance | Cost per platform-reported conversion, as a proxy only |
| Commit to a new channel for the coming year | CMO with finance partner | Annual | High: a new budget line | Low: contracts and hiring involved | Randomized lift test with independent readout before commitment | Incremental revenue against the counterfactual |
The evidence-standard column is anchored on the platform-report illustration from Supermetrics and the outcome column on the LatentView definition of measurement as a link to business outcomes.
To check completeness, reconcile the inventory to the budget: every planned dollar should trace to at least one row. Money that traces to no row is being spent without a decision anyone owns. A practical way to build the first version is a 90-minute session with the media lead, the analytics lead and the finance partner, filling the table for the next four quarters and then sorting by spend at stake and reversibility. The top of that sorted list is where Steps 2 and 3 begin.
Step 2: Assign Each Decision to Attribution, Experiments, or MMM
Assign a method to each inventory row by reading its cadence and evidence-standard columns, not by asking which method your team prefers. The question buyers often ask, which of MTA, MMM or incrementality is best, has no answer at the level of the company; it only has an answer at the level of a row.
Match cadence and stakes to the method that can answer at that speed
Daily and weekly optimization within a channel already shown to work maps to attribution or platform data, because those are the only sources that update at that speed. Channel-level go or no-go and questions about scaling spend map to incrementality experiments, because they need a causal answer. Annual and quarterly allocation across channels maps to MMM, because it is the only method that sees the whole mix at once. This mirrors the division of labor BCG observed among the 46% of surveyed marketers running all three: MMM for planning, experiments for causal insight, MTA for daily optimization.
Assignment is per decision, never per team or per channel. The same paid social channel can legitimately appear in three rows with three methods: weekly creative ranking on platform data, a quarterly holdout for its budget tranche, and a coefficient inside the annual MMM.
| Decision type from inventory | Typical cadence | Assigned method | Why this method fits | What it cannot tell you | Fallback if data readiness is missing |
|---|---|---|---|---|---|
| Creative and bid rotation within a validated channel | Daily to weekly | Attribution or platform reporting | Updates fast enough to act on | Whether the channel causes sales | Platform reporting with a wider tolerance for error |
| Channel go or no-go, budget scale-up | Quarterly or on trigger | Incrementality experiment | Isolates causal effect for that channel and period | Cross-channel trade-offs or a full-year plan | Delay the decision or run a platform lift study with independent readout |
| Cross-channel allocation | Quarterly and annual | MMM calibrated to experiments | Sees the whole mix and external drivers | Which creative to run tomorrow | Experiment-informed allocation on the largest lines only |
The method roles in this table follow the split reported by BCG; the causal claim for experiments rests on Haus.
Where attribution is still the right tool
Attribution and platform reporting remain the right input for ranking within a channel: which creative, which audience, which bid. They are fast and granular. Their limit is that they cannot separate sales the ads caused from sales that would have happened anyway, which is why they never serve as the evidence standard for a scale decision.
Where only an experiment will do
When the question is whether spend adds sales, only a design with a control group answers it. Haus describes incrementality tests as isolating marketing's true impact through randomized controlled trials. Any row whose evidence standard reads causal gets an experiment.
Where MMM earns its place
MMM earns the cross-channel rows because nothing else models the mix together. It is also the method most dependent on the others: without experiments to constrain it, its channel estimates are only as good as the historical variation in the data.
Reconciling conflicting answers on the same decision
Agree the reconciliation rule before any results exist. When methods disagree on a row, the experiment result is the reference, MMM is recalibrated toward it, and attribution is used only for ranking within the channel. Return to the Supermetrics illustration: a platform reports 1,000 sales, and if a holdout suggests around 70% would have occurred regardless, the weekly row keeps using platform data to rank creatives while the quarterly budget row uses the roughly 300 incremental sales as its input. Add two columns to the inventory, assigned method and reference when methods disagree, and have finance sign them.
Step 3: Sequence Experiments to Calibrate Attribution and MMM
Run experiments before you trust either of the other two methods, and run them in an order set by the inventory. Randomized designs are the only part of the stack that produce a causal anchor on their own; Haus's description of incrementality tests as randomized controlled trials that isolate true impact is the reason they sit at the front of the roadmap rather than the end.
Start with the largest reversible line item
The first test should target the inventory row with the largest spend at stake and the highest reversibility. Size matters because the value of a test is the size of the reallocation it can change. Reversibility matters because a result is only useful if someone can act on it this quarter. This also answers the cost question: measurement spend is justified by the decision it informs, so budget each test against the spend-at-stake column of the row it serves, and skip tests whose row could not move enough money to pay for them.
Design for the decision, not for statistical elegance
The sequencing process, as an ordered routine:
- Pick the inventory row with the largest spend and highest reversibility that has no causal evidence within the last six to twelve months.
- Choose the design your data readiness allows: a geo holdout where markets can be split, an audience holdout where they cannot, or a platform lift study where neither is available, with independent readout wherever possible. Step 4 covers the readiness dependency.
- Pre-register the decision on one page: hypothesis, design, the minimum effect the team cares about, the action that fires at each outcome, and the date MMM and attribution will be recalibrated.
- Run the test for the pre-agreed window and read it out independently of the platform being tested.
- Feed the measured incremental effect into MMM as a constraint on that channel's coefficient and rescale attribution's share for that channel to the incremental total.
- Record the result against the inventory row and schedule the next row.
The pre-registration in step 3 is the easiest item to skip and the most expensive omission. Without a pre-committed decision, a test result rarely changes a budget.
Use each result to recalibrate the other two methods
Calibration is the mechanism that turns three methods into one description of reality. An experiment yields an incremental effect for one channel and one period. MMM takes that effect as a constraint, so the model's coefficient for that channel is held near the measured level instead of floating on historical correlation alone. Attribution's share for that channel is then scaled so its total matches the experiment's incremental total, and its within-channel ranking is kept for creative and bid decisions.
Using the Supermetrics illustration as the worked example, with its numbers understood as the source's teaching example rather than a measured result: the platform reports 1,000 sales, and the holdout suggests roughly 300 were incremental. The MMM coefficient for that channel is constrained toward the 300 level, the attribution total for the channel is rescaled to 300 while its creative ranking is preserved, and for that quarter all three methods agree on what the channel added.
Build a rolling 12-month test calendar
Maintain a calendar with four quarters ahead. Each quarter carries one or two live tests, each tied to an inventory row and to the decision date it must land before. Add a retest cadence for channels whose effect may drift because of creative fatigue, auction changes or seasonality; a result from a year ago is context, not a current input. The calendar should sit next to the review dates from Step 5 so that readouts arrive before the meeting that acts on them.
Common sequencing mistakes
- Testing small channels first because they are easy to isolate, which produces clean answers to questions that move little money.
- Running tests with no pre-committed decision, so results are debated rather than executed.
- Treating one lift study as permanent truth and never retesting.
- Letting the platform under test run the holdout and read the result without an independent check.
- Skipping calibration, so the experiment result sits in a deck while MMM and attribution continue to disagree.
BCG's survey suggests the combination of methods is already widely used; the sequencing and calibration discipline is what makes the combination worth having.
Step 4: Data Readiness for a Hybrid Measurement Stack
Score your data readiness for each method before assigning any inventory row to it. A method whose inputs are missing cannot answer anything, so readiness acts as a gate: rows are assigned to a method only when its inputs exist, and an interim method covers the row until the gap closes. This section is a conceptual framework derived from what each method needs to run, not a benchmark study, and specific tooling choices are out of scope.
What attribution needs
Attribution needs consistent conversion definitions across platforms and the site, and identity resolution to the extent it is legally and technically available in your markets. The common gap is definitional: the same word, conversion, meaning different events in different systems, which makes any cross-platform ranking unreliable.
What experiments need
Experiments need the ability to split exposure by geography or audience and to read outcomes from a source that is independent of the platform being tested. Haus's framing of incrementality tests as randomized controlled trials implies both conditions: a control group and an unbiased readout. The common gap is readout independence, where the only outcome data available comes from the platform whose effect is in question.
What MMM needs
MMM needs roughly two to three years of weekly spend and outcome history by channel, plus the external drivers that move demand: pricing, distribution, seasonality, competitor activity. The common gap is history that is too short, too aggregated or broken by a platform migration.
Diagnosing your readiness level
| Method | Minimum data inputs | Readiness signal (green) | Common gap (amber or red) | Owner to close gap | Interim method while gap is open |
|---|---|---|---|---|---|
| Attribution | Consistent conversion definitions; identity resolution where available | One conversion taxonomy used by every platform and the site | Conflicting definitions; heavy modeled conversions | Analytics lead with channel managers | Platform-native reporting with explicit tolerance for error |
| Experiments | Exposure split by geo or audience; independent outcome readout | Ability to hold out a region or audience and read sales from first-party or warehouse data | No geo control; readout only from the platform | Measurement lead with data engineering | Platform lift study with the result treated as directional |
| MMM | Two to three years of weekly spend and outcomes by channel; external drivers | Complete weekly history, channel-level, with price and seasonality variables | Short or aggregated history; missing external drivers | Planning analyst with finance | Experiment-informed allocation on the largest lines only |
The boundary between reporting and measurement follows the LatentView definition; the experimental requirements follow Haus.
In data terms, warehouse tables that record spend and conversions are reporting. Measurement needs a design, the holdout, or a model, the MMM, built on top of those tables. Clean tables are necessary but not sufficient.
Interim answers while gaps are closed
Score each method red, amber or green, then record the owner and target quarter for every cell that is not green directly in the inventory. As a conceptual example, not a client case: a brand with clean weekly spend history but no way to split markets would score green for MMM and red for experiments. Its roadmap would start with an MMM refresh and an audience-holdout pilot on the one platform that supports independent readout, rather than the geo test it would prefer, and the geo capability becomes a dated task for data engineering. The inventory rows that need geo evidence carry an interim method and a note that their evidence standard is temporarily unmet.
Step 5: Ownership and Review Cadence Across Marketing and Finance
Assign one named owner to each method and one named owner to each inventory decision, and publish both on a single page co-signed by the finance partner. This is where measurement programs tend to break down, more often than in the statistics.
Who owns each method and each decision
A workable split: the performance lead owns attribution inputs and the conversion taxonomy; the analytics or measurement lead owns experiment design, readout and calibration; a planning analyst who sits close to finance owns the MMM and its scenarios. Decision ownership is separate. The owner of the quarterly reallocation row is accountable for making that call on the review date, using whatever the assigned method produced, and for recording the decision. Method owners supply evidence; decision owners act. If an agency will own part of this stack, see how GPI documents a paid media partner's evidence in a profile such as AB Marketing Group, and hold the agency to the same owner-and-evidence standard you apply internally.
The three review rhythms: weekly, quarterly, annual
| Review rhythm | Cadence | Inputs reviewed | Decisions in scope | Owner | Finance role |
|---|---|---|---|---|---|
| Optimization review | Weekly | Attribution and platform reporting for validated channels | Creative, bid and audience rotation | Performance lead | None beyond visibility |
| Reallocation review | Quarterly | Latest experiment readouts; MMM refresh calibrated to them; signal-change check | Budget shifts between channels; go or no-go rows due this quarter | Decision owners with measurement lead | Co-signs the numbers used and confirms they meet the agreed evidence standard |
| Planning review | Annual | MMM scenarios; the year's test learnings; readiness diagnostic | Next year's budget split and new-channel commitments | CMO with planning analyst | Co-owns the plan and the evidence standard for the following year |
The review rhythms align to the cadence column from Step 1 and to the method roles observed in the BCG survey.
Getting finance to sign the evidence standard, not the number
Finance rarely distrusts marketing arithmetic; it distrusts numbers whose provenance it did not agree to. The fix is to have finance sign the evidence standard and the reconciliation rule from Step 2 before any result exists, then present experiment-calibrated figures with their stated limitations rather than platform totals. A CFO who agreed in January that a geo holdout is the standard for the paid social row has little reason to dispute the holdout's answer in April.
What the review meeting actually decides
Each review decides the rows due on that date and nothing else. The weekly meeting rotates creative. The quarterly meeting moves money between channels using the calibrated inputs and confirms the next quarter's tests. The annual meeting sets the split and commits to new lines. Method choice is not on any agenda; it was settled in Step 2. BCG's finding that more than 70% of surveyed marketers use AI in measurement while nearly one in three still struggle with cross-channel effectiveness is a useful caution: in that sample, adding tools did not resolve the question. Ownership and cadence are the layer that was missing, and no tool supplies them.
Warning signs the cadence is failing
- The quarterly review re-litigates which method to trust.
- The MMM owner and the experiment owner present competing numbers for the same row.
- Tests finish with no decision attached and no reallocation recorded.
- Finance maintains a parallel model and quietly plans from it.
- Review dates slip until readouts arrive after the decisions they were meant to inform.
Making the Roadmap Durable as Signals Change
Why the decision inventory outlives any tracking method
The decisions a company makes about its budget change slowly; the signals available to inform them change often. That asymmetry is why the inventory, not the method, anchors the roadmap. When a platform, browser or regulator changes what user-level data is available, the rows in the inventory stay the same. What changes is the readiness column and, for some rows, the assigned method.
Weight the stack toward aggregate and experimental evidence
User-level tracking changes hit attribution inputs hardest. Experiments depend on splitting exposure by geography or audience and comparing groups, and MMM depends on aggregate weekly history; neither requires following an individual across sites. Haus's description of incrementality tests as randomized controlled trials explains the resilience: the design isolates impact through the control group, so a loss of individual identifiers degrades the daily row without touching the quarterly one. As a conceptual illustration, a weekly optimization row that leaned on user-level attribution might move to platform-reported directional data, while the quarterly reallocation row keeps its geo-holdout standard unchanged.
Re-run the readiness diagnostic when a platform or policy changes
Add a standing signal-change item to the quarterly review. Whenever a platform materially changes its measurement or targeting rules, re-score the readiness table from Step 4 and revisit assignments for affected rows. Treat it as maintenance rather than a rebuild. Durability here is a property of the design; this article makes no forecast about specific policy timelines.
How GPI Reads a Measurement Roadmap When Evaluating Partners
The roadmap reduces to five artifacts a CMO can ask for from an internal team or an agency: the decision inventory, the assignment matrix with its reconciliation rule, the rolling test calendar, the readiness diagnostic, and the review RACI. Any partner can say it runs MMM, experiments and attribution; the BCG survey suggests that combination is already common among respondents. The differentiator is whether a partner can show how the three connect to your decisions.
GPI applies the same standard when documenting agencies. A partner's measurement claims are read against the evidence, methodology and stated limitations behind them, not against reported ROAS alone, because a reported total is reporting rather than measurement until a method ties it to an outcome. A practical evaluation step: ask any prospective agency to walk through your inventory and mark which rows it can support with experiments and which only with reporting, then record the answer. For an example of how GPI records a marketplace-focused partner's documented evidence, review the AMZ-Marketing profile.
Frequently Asked Questions
How many experiments should be live at once when we are starting the roadmap?
One, on the largest reversible row, until the team has completed a full cycle from pre-registration through calibration. Step 3's calendar allows one or two tests per quarter once the routine is established. This is a framework recommendation based on operational load, not a benchmark from measured programs.
What do we do when an MMM refresh contradicts a recent geo test?
Apply the reconciliation rule agreed in Step 2: the randomized result is the reference for that channel and period, and the MMM coefficient is constrained toward it. Haus's framing of these tests as randomized controlled trials is the basis for that ordering. If the gap is large, check whether the test period was unusual before treating the result as the new prior.
Can we run this roadmap with an agency owning attribution and an in-house team owning MMM?
Yes, provided the RACI from Step 5 names one owner per method, one owner per decision, and a single reconciliation rule both parties have signed. The risk is two owners presenting competing numbers, which is a listed failure signal. Decision ownership should stay in-house even when method ownership is shared.
How long before finance should expect calibrated numbers rather than platform totals?
After the first experiment on a major row has been read out and fed into MMM and attribution, which the sequence in Step 3 places within the first one or two quarters if readiness permits. Until then, present platform totals labelled as directional with the interim-method note from Step 4. The article offers this as a planning frame, not a measured timeline.
Do small-budget channels ever justify an incrementality test?
Only when the row's spend at stake and reversibility would move enough money to pay for the test, or when the channel is a candidate for scale and the go or no-go row needs causal evidence. Testing small channels first for convenience is a sequencing mistake listed in Step 3.
How often should the decision inventory itself be revised?
Review it at every annual planning meeting and touch it at the quarterly review whenever a signal-change or a new channel adds or removes a row. The fields from Step 1 rarely change; the readiness and assigned-method columns change more often, which is the intended design.
