marketing-attribution v1.0.0

name: marketing-attribution description: Marketing attribution that holds up: pick the right attribution model, enforce UTM naming discipline, verify incrementality with holdouts, and map spend to revenue. Use when asked what drove conversions or sales, which channel deserves budget, how to tag campaign links, or whether a channel is actually causing results.

Marketing Attribution

Purpose

Answer "what actually drove the revenue" without lying to yourself. Attribution models divide credit for past conversions; UTM discipline keeps the raw data clean enough to model at all; incrementality tests check whether a channel caused results or just showed up near them. This skill runs all three.

Workflow

1. Define the decision first. What will change based on this analysis? (budget shifts, keep/kill a channel, campaign tag rollout). Pick ONE decision. Attribution that answers everything answers nothing.
2. Check the data quality before modeling. Is every paid/owned link UTM-tagged with consistent casing and separators? Is there a shared naming doc? If not, start with references/utm-discipline.md and fix tagging before trusting any model output. Garbage tags in, garbage models out.
3. Pick a model that fits the buying cycle, not the default your dashboard ships. Rule of thumb: impulse/short cycle -> last-touch; considered B2C -> linear or time-decay; B2B -> position-based (U-shaped); mature volume -> data-driven; enterprise accounts -> account-level, not lead-level. Full selection matrix in references/attribution-models.md.
4. Compute per-channel attributed revenue and ROAS under the chosen model (and at least one second model for comparison). If the two models tell opposite stories about a channel, say so plainly: that is the finding, not a tie to break.
5. Verify with incrementality, not more modeling. Before defunding or scaling a channel, run a holdout: user-level randomization if the platform offers it (Meta/Google lift studies), geo holdout for everything else (4+ weeks, 10-15 matched markets, test spend vs holdout). Designs and math in references/incrementality.md.
6. Report revenue, not impressions. The deliverable is a per-channel table: spend, attributed revenue under the chosen model, attributed conversions (fractional allowed), cost per attributed conversion, attributed ROAS, and the incrementality verdict for any channel being scaled or cut.

Output Contract

• Every deliverable names the model used and the measurement window. If the window is too short for the channel (e.g. 7 days for SEO), flag it rather than shipping zeros.
• Fractional conversions are fine (0.4 of a conversion); round for humans, keep fractions in the math.
• Never present one model's numbers as "the truth." Attribution is a lens; incrementality is the check.
• Unverifiable numbers get cut or framed as estimates, never stated as fact.

Operating Rules

1. Ask which decision the analysis serves before touching data. No analysis without a decision.
2. Audit UTMs before modeling. A skill run that skips the tagging check is not done.
3. When models disagree, report the disagreement. Picking one silently is the failure mode.
4. A channel with great last-touch ROAS and no incrementality is capturing demand, not creating it. Say that in plain words.
5. Keep the heavy math in references and the CLI; SKILL.md stays the workflow.