# Decision map: which evidence answers which question Attribution allocates credit across observed interactions. An incrementality experiment estimates the effect of a defined intervention against a counterfactual. Marketing mix modelling (MMM) estimates channel effects from aggregate data over time under explicit assumptions. Self-reported attribution records what buyers remember. They answer different questions; choose by the decision, not by the tool the team already owns. ## Contents - Step 1: write the decision and the alternative - Questions attribution can answer - Questions attribution cannot answer - When an incrementality test is the right evidence - Feasibility screen when no test design skill is installed - When MMM is the right evidence - Self-reported attribution - Attributed versus incremental: the arithmetic that changes decisions - Triangulation rules ## Step 1: write the decision and the alternative "Which channel works?" cannot be measured. Rewrite it until it names an action, a size, a date and an alternative: - "Should we add $10,000 a month to prospecting on paid social next quarter, versus keeping spend flat?" - "What happens to new paid accounts if we stop brand search retargeting for six weeks, versus running it as now?" The alternative matters. A test that switches a channel off estimates the effect of switching it off under the test conditions. It does not tell you the return on the next dollar above today's budget. ## Questions attribution can answer Use attribution, labelled with platform, model, window and date basis, for: - Which campaigns precede measured outcomes, and how that pattern changes over time. - Where measurable demand first enters the journey (first observed view). - The last measurable route into conversion (last non-direct view). - Whether a campaign is meeting its in-platform target under its own configured model. - Operational diagnosis: a campaign that suddenly sends visits but no verified outcomes points to audience, landing page or tracking problems. No MMM is needed to find a broken signup flow. - Where a tracked journey loses people. ## Questions attribution cannot answer Route these elsewhere, and say so in the brief: - "How many sales would we lose without this channel?" Attribution sees who converted after an ad, not who would have converted anyway. Retargeting audiences already contain people with demonstrated intent, so strong attributed results mix ad effect and selection. - "What is the return on the next $10,000?" Average attributed return is not marginal return. Response curves flatten as spend grows. - "How much did TV, podcasts, out-of-home or word of mouth contribute?" These leave few or no observable clicks. - "What should the cross-channel budget split be next year?" Recurring allocation across channels is an MMM question, ideally calibrated with experiments. - "Is this iOS install campaign profitable at user level?" SKAdNetwork and AdAttributionKit postbacks are aggregate and privacy-thresholded; see `mobile-limits.md`. Decision rule used by this skill: if the preferred action moves budget between channels or cuts a channel, and it is costly to reverse, attribution is hypothesis evidence only. The brief must name the causal evidence still needed or state that the team accepts the risk. Costly to reverse means either condition holds: - The change is at least 20% of the channel's monthly spend. - The decision owner says the downside if the decision is wrong exceeds their stated tolerance. If neither is known, ask the owner for the tolerance (a revenue or new-customer amount over a stated period) and record it in the brief. A 10% trim with a known tolerance that the worst case stays inside can run on attribution evidence plus a monitoring plan. ## When an incrementality test is the right evidence Use a test when the decision is about one defined intervention and the business can hold a control. - Randomised user-level or account-level tests separate the intervention from pre-existing differences. Analyse by assignment, not by who happened to see an ad; exposure is not random. - Geo tests assign non-overlapping regions to treatment and control. They need enough comparable regions, outcomes measurable by region, control over targeting and limited spillover. - Platform lift studies: Google Ads Conversion Lift offers user-based and geography-based studies and is not available to every account (check eligibility live with the account team). Meta's Ads MCP server has conversion lift and A/B test tools; creating a study changes the account and needs approval. - A creative A/B test measures creative A against creative B. If both groups see ads, it says nothing about the programme against no programme. - Plan sample size and minimum detectable effect before launch. If the feasible design can only detect a 40% lift and the decision turns on 5%, the test does not fit the decision. Worked uncertainty (fictional, simple randomised design, normal approximation): - 10,000 people per group, 600 versus 500 purchasers (6% versus 5%). Absolute difference 1 percentage point, relative lift 20%, about 100 extra purchasers. 95% interval roughly 0.37 to 1.63 points, so about 37 to 163 extra purchasers. - 1,000 per group, 60 versus 50. Same 20% lift, interval roughly minus 1.00 to plus 3.00 points. The test cannot rule out zero or a loss. Hand off test design to the incrementality-test-design skill. If it is not installed, run the screen below and put its answers in the brief. ## Feasibility screen when no test design skill is installed Answer each question with evidence. Any "no" means the test does not fit the decision yet; say which and what would change it. - Can the campaign hold different conditions in the chosen regions or groups for the whole period (geo targeting, audience exclusions)? - Can business outcomes be assigned to regions or groups consistently, online and offline, from the ledger? - Are promotions, distribution changes and sales-team actions balanced across groups or explicitly handled? - Is spillover small enough: customers crossing regions, national media, word of mouth? - Can the business tolerate the holdout, cut or extra spend for the full period, plus the measurement tail? - History: is there daily outcome history by region of at least 3 times the planned test length (28-day test: 84 days; 42-day test: 126 days), and longer when seasonality is strong? - Response metric: is the response an absolute count or gross revenue from the ledger, never attributed conversions, ROAS, or net revenue after refunds that can go negative? - Detectable effect: estimate the minimum detectable effect from historical outcomes by region or group. If it is larger than the effect the decision turns on, the test cannot settle the decision. - Measurement end: run the readout through the test end date plus the 90th percentile of observed touch-to-outcome lag, so late outcomes from test-period exposure are counted. Vendor tools to name in the brief (running any of them changes live campaigns, so it needs approval): - Google Ads Conversion Lift: user-based or geography-based studies; availability varies by account, so confirm eligibility. - Meta conversion lift and A/B tests: the lift and A/B test tools on Meta's Ads MCP server, or Meta's own experiment tools; confirm the account can run one. - Meridian GeoX: Google's geo experiment design method in the Meridian docs, with stratified assignment, A/A simulation checks and power expressed as minimum detectable effect. Its pre-test data rules (read 2026-10-20): daily rows only (weekly is not supported), one row per date and location; raw, unfiltered conversions, not attributed conversions; absolute, non-negative metrics such as gross revenue or conversion counts, never ROAS or net revenue after refunds; a spend series is required for go-dark and heavy-up designs but not for holdback cells; at least 3 times the test length of history, longer with strong seasonality. ## When MMM is the right evidence Use MMM when the decision is recurring cross-channel allocation, channels include ones with little click data, and there is enough history with real variation in spend. - Google's Meridian guidance (checked 2026-10-17): at least two years of weekly data for geo-level models, three years for national models, and at least three years if only monthly data exists. Required inputs include a KPI summable across geography and time, spend per paid channel, one exposure metric per paid channel (impressions, clicks or spend), and control variables. These are planning guidelines, not a guarantee of a useful model. - Spend that rises because the team expected demand to rise will be credited to the channel by a naive model. Collect the reasons budgets changed, not only the spend. - Controls are chosen for the causal question. Controlling for a mediator such as site visits can remove part of the advertising effect being estimated. - Experiments can calibrate MMM priors, but a test of a 20% spend increase is not the same comparison as switching a channel off. - Data sources: platform spend and impressions exports (Meta's Insights API has a `breakdowns=mmm` CSV export at ad set level with impressions and estimated spend), finance outcomes by week and region, promotions and price changes. - Google publishes Meridian agent skills (Apache-2.0) for model building, budget optimisation and scenario planning. They assume a data scientist owns the model. Hand off readiness to the mmm-readiness-check skill first. If it is not installed, check the history rule above, list which of outcome by week and region, spend and one exposure metric per paid channel, controls, and reasons for budget changes exist, and name the owner of the model before any Meridian work. ## Self-reported attribution Use it to find influences that leave no click: podcasts, communities, peers, events, AI assistants, offline sales conversations. - One optional open-text question ("How did you hear about us?") on a high-intent form or at signup. A required dropdown produces the first plausible option, not the truth. - Store the raw answer unchanged with a timestamp. Code it into categories later with a written codebook, and keep the code version. - Report it as its own view beside software attribution, by the same outcome and period. Do not overwrite source fields with it and do not add its counts to attributed counts. - Read disagreement as the finding. If 30% of qualified opportunities name a podcast and click data shows direct or organic for most of them, that is a hypothesis about where demand is created, worth a test. - It has recall and response bias and a non-response rate. Report the response rate next to every self-reported share. ## Attributed versus incremental: the arithmetic that changes decisions Fictional numbers: - $5,000 of spend, $20,000 of attributed revenue: attributed ROAS 4.0. This does not say the spend created $20,000. - A credible test estimates $8,000 of additional revenue: incremental ROAS 1.6. - At a 60% contribution margin before media, $8,000 contributes $4,800 against $5,000 of media: a $200 shortfall before other costs. Label every ROAS as attributed or incremental, with the platform, model and window, and as a period view (this period's value over this period's spend) or a cohort view (value from the customers this spend acquired, through a stated date). Hand ROI and payback maths to the marketing-roi-and-payback skill; if it is not installed, stop at these labels and do not compute ROI or payback here. ## Triangulation rules - Align outcome, population, period and comparison before comparing methods. If they differ, the numbers answer different questions. - Where attribution and an experiment disagree, check measurement failures and uncertainty first, then use the evidence that matches the decision. Do not force them to match. - Where two attribution views disagree (organic search leads first observed, paid brand search leads last non-direct), investigate paths and customer evidence. Do not split budget evenly between the two charts. - If the decision flips under plausible assumptions, recommend a smaller reversible step or better evidence.