An MMM needs a trustworthy business outcome, a usable record of the activity competing to explain it, and context for material changes that were not caused by current media. The exact source list follows the business and decision. It is not a request to connect every available system.
Start with the decision, then the outcome
Before listing sources, write the allocation decision down. Is the team deciding whether to protect upper-funnel video, move the next dollar from paid social to Search, or set a more realistic target for a demand-capture channel? The decision determines the outcome and level of detail the model needs.
For many DTC businesses, the primary outcome is daily orders, sales, or revenue. Finance or the commerce system should own that commercial record. Platform-reported conversions, GA4, and an attribution tool can add useful perspective, but they should not quietly replace the outcome the business is trying to explain.
The core model inputs
1. The business outcome
- Orders, sales, revenue, or leads by day. Choose the outcome that matches the decision, then keep its definition consistent across the history.
- Major sales sources. Document where the business outcome is recorded and how the total reconciles to the model input.
- New-versus-returning treatment. Include it when the distinction is material to the question, and document what the business means by each label.
Do not begin with a platform conversion total simply because it is easier to export. An MMM can only provide a useful portfolio view if its outcome represents the business result leaders will act on.
2. Media investment and delivery
- Daily spend by channel or tactic. Useful lines may be Meta and Google, but can also be awareness versus conversion, YouTube versus Search, or other independently variable tactics within a platform.
- A stable channel taxonomy. Preserve definitions and note when campaigns, objectives, or naming practices changed. A channel that changes meaning halfway through the history is not one uninterrupted input.
- Available delivery context. Impressions, reach, frequency, and campaign structure can help investigate a change or decide whether to aggregate a line. They are not a reason to make every campaign its own model variable.
- Meaningful non-clickable media. TV, radio, podcasts, video, CTV, and similar activity still need a dated activity or spend record when they are part of the allocation question.
The model needs variation it can learn from. A channel that ran for a week, spent at almost the same level every day, or always moved in lockstep with another channel may not be scoreable on its own. That is a data limitation to document, not a reason to create a precise-looking ROI.
Context records to consider
These sources belong when they could plausibly move demand or the measured outcome independently of the media currently under review. Keep the records dated, specific, and close to the event. A retrospective explanation added after the result is less useful than an operating record maintained as the business changes.
3. Your promotion and merchandising calendar
- Promotion dates, offer details, and the intended merchandising effect.
- Major product, price, or availability changes when they could affect sales or conversion.
- Campaign and offer changes that materially alter what customers saw or could buy.
A sales increase during a promotion may reflect the offer and related demand, not a newly launched ad. If the promotion record is missing, the model may be asked to assign that movement to whatever media changed nearby.
4. Baseline demand and market conditions
- Seasonality and external events. Include the events or demand periods that are material to the category and the decision period.
- Brand, competitor-brand, or category search interest. These can be relevant demand proxies when they reflect a real condition the team needs to distinguish from paid activity.
- Weather. Use it only when the category is meaningfully weather-sensitive.
- Competitor promotions or near-term market activity. Include credible records when the team has them; do not invent a competitor calendar from hindsight.
These inputs help the model consider alternative explanations. They do not eliminate uncertainty or prove why an individual sale occurred.
Useful supporting evidence, not substitute outcomes
5. GA4 and first-party event data
GA4 can provide continuous site history, traffic, sessions, and behavior that help diagnose a result. First-party event data can make repeated visits and on-site behavior more inspectable where it is appropriate and privacy-compliant. GA4 exported to a client-owned BigQuery environment is a strong setup, but manual exports or other usable records may still support a review. Neither source captures every exposure or resolves every identity.
6. Attribution and platform reports
Bring the current platform claims and attribution views to the review, especially where their combined conversions exceed actual business outcomes. They are valuable for framing the problem, inspecting observable paths, and finding hypotheses. They are not additive sales totals and should not be used as the model's commercial outcome by default.
7. Incrementality-test results
A relevant, well-designed lift test can anchor a specific causal claim under the conditions it tested. Record the intervention, population, timing, spend scale, outcome, duration, and uncertainty. An MMM can put that evidence in the broader portfolio context; it should not treat a test score as a permanent, universal channel truth.
The MMM handoff checklist
For each source that enters the work, keep a short source card. It makes the model easier to audit and the next refresh less dependent on the person who built the first version.
- What does this source measure? State the exact outcome, activity, or condition.
- What is its grain and history? Note daily, weekly, geographic, or other usable detail; start and end dates; and known gaps.
- Who owns it? Name the person or system responsible for the source and definition.
- How is it transformed or grouped? Record channel mappings, exclusions, aggregations, and definition changes.
- What could make it misleading? Capture tracking gaps, taxonomy changes, incomplete coverage, or a reason it may move with another variable.
- What decision does it inform? If no one can name one, it may be context for investigation rather than a model input.
What if a source is missing?
Missing history does not automatically end the project. It may change the scope: aggregate related tactics, narrow the decision, begin a clean calendar going forward, or use a focused test for the claim that cannot wait. What should not happen is quietly filling a material gap with an assumption and presenting the result as settled.
For a broader readiness decision, read Do You Need an MMM Yet?. If a specific channel claim needs causal calibration, start with the incrementality-testing guide.