An AI system can read a KPI target. It cannot know why your team is willing to miss it this month unless you tell it. This template gives the reporting workflow the business context needed to ask better questions without making the decision for you.
Start with a working document, not a perfect one
Make one shared document or repository. Give it one accountable owner: usually the person who owns media performance or the growth plan. Other people can contribute source material, but someone needs to approve the current version and make clear what is fact, what is a working hypothesis, and what is a policy.
This is not a replacement for data engineering, privacy review, event capture, or identity resolution. It is the operating context that sits beside the data. If your reports disagree because their definitions differ, begin with the reconciliation work. If you need the larger workflow, read how AI reporting should fit into DTC operations.
Copy this marketing context layer template
1. Purpose, scope, and owner
Decision this context supports: [For example: whether to maintain, increase, or reduce a channel budget.]
Accountable owner: [Name and role.]
Last approved: [Date.]
Review cadence: [Weekly, monthly, or before a major promotion.]
2. Primary outcome and metric definitions
North Star outcome: [Define the business outcome.]
Metrics used in the weekly report: [Name, definition, source, timezone, attribution window, exclusions.]
Known reporting differences: [For example: finance recognizes revenue differently from the ad platform.]
Do not write only “CPA under $X” or “ROAS above Y.” Define what the number represents and where it comes from. If a term has two valid definitions, record both and say which one governs this decision.
3. Campaign and offer roles
For each meaningful campaign or offer, document its job: workhorse acquisition, upper-funnel demand creation, test, funnel filler, loss leader, retention, or another defined role. State the normal guardrail and any intentional exception.
Example: “This offer may run below the normal CPA target during the launch window because it is designed to create a qualified customer pool for a later promotion. Do not recommend pausing it on CPA alone.”
4. Promotional and operating calendar
List planned promotions, launches, inventory limits, pricing changes, creative refreshes, site changes, major channel changes, and expected seasonality. Record the expected effect and the date range. A sharp change after a sale may be normal demand timing, not a reason to rewrite the media plan.
5. Customer, product, and market context
Summarize the customer segments that matter, the product benefits and tradeoffs, current product development, competitor context, keyword research, and customer feedback that changes how performance should be interpreted. Link to the source where possible. A context layer should point to evidence, not turn one person's opinion into permanent instruction.
6. Known data gaps and prohibited assumptions
State what the workflow cannot see or should not infer. Include missing event paths, checkout handoffs, identity limitations, reporting delays, consent effects, gaps in platform data, and any field the AI must not use. Do not place credentials or personal data in a general-purpose tool.
7. Approved hypotheses and prior learning
Write the hypothesis, the evidence behind it, its limits, and the next check. Include prior tests, model findings, or research that should inform the investigation. A test result belongs here with the conditions it covered—not as a permanent rule for every campaign and season.
8. Open disagreements
Record competing views as unresolved questions. For each one, name the people or teams involved, the evidence each view relies on, and what would help settle the difference. Do not present conflicting opinions as equally authoritative commands for the model.
9. Change log
Keep the date, change, contributor, approver, and reason for material updates. If a past assumption is retired, leave the history visible. Version history gives the team a way to ask why a recommendation changed instead of treating the latest prompt as timeless truth.
Three rules that keep the layer useful
Keep it close to live decisions
A 60-page brand archive is less useful than a maintained record of the decisions the team is making now. Start with the active campaigns, current offer, known caveats, and next budget question. Add depth when it changes an interpretation.
Let AI investigate, not execute
The system can assemble the report, identify a drop-off, and propose competing explanations. It should not unilaterally pause a campaign, change bids, or alter an offer. A human owner should decide whether the evidence calls for a test, a tracking fix, a watch period, or a real change.
Use the layer to improve the data work
Repeated caveats are useful signals. If the team keeps writing that orders disappear after a checkout handoff, the next job may be event capture—not another prompt. If the business needs to know whether a channel created incremental demand, move from observed patterns to a feasible incrementality test.