Thesis: Attribution isn’t a single truth; it’s a decision aid. The right “story” depends on your stage, traffic mix, and the decisions you need to make this week vs this quarter. Pick the model that minimizes regret for your current constraints—and evolve it as you grow.
1) A simple lens: Three axes that decide your model
- Decision horizon: Days–weeks (tactical) vs months–quarters (strategic/finance).
- Data maturity: From ad‑platform pixels only → 1st‑party events in a warehouse → modeled/aggregate (MMM).
- Spend risk: <$50k/mo (cheap exploration) → $50–250k/mo (material) → $250k+/mo (must quantify ROI with error bars).
Rule of thumb: Short horizon + low maturity → click‑path heuristics. Longer horizon + higher maturity/spend → experiments + MMM.
2) Stage playbooks
A) Pre‑PMF / Early Traction (0–$50k/mo)
Decisions: Which messages and channels show promise? What should we do next week?
Useful story: Directional and fast: last‑click for activation; first‑touch for discovery.
What to implement
- UTM discipline (see §7): every link tagged; insist on
utm_contentfor creative. - Source‑of‑truth events: signup/lead, qualified (MQL/SQL), revenue. Use server‑side tagging where possible.
- Assisted views: position‑based (U‑shape 40/40/20) in GA/warehouse to avoid last‑click myopia.
- “How did you hear about us?” free‑text survey on thank‑you; hand‑classify weekly.
- Cheap incrementality checks: geo/off‑geo toggles for paid social; content off/on tests.
Guardrails
- 7–14‑day lookback; dedupe by user; separate brand vs non‑brand search.
- Don’t annualize thin data; run 2–3 week sprints and archive learnings.
- Keep a kill list: channels/tactics to pause unless new evidence appears.
Outputs
- A weekly “next bet” list (3–5 items).
- A 1‑pager of early signals by channel and message.
B) PMF & Growth ($50–250k/mo)
Decisions: Budget shifts across channels/creatives, lifecycle and sales‑assist priorities.
Useful story: Hybrid. Heuristic MTA for ops + lift tests to calibrate + early MMM‑lite for pacing.
What to implement
- Position‑based or time‑decay MTA in your warehouse (user‑level paths → weighted credit).
- Sales‑assisted attribution: tie CRM stages to campaigns; score revenue influence, not just clicks.
- Lift & holdouts: platform conversion‑lift on Meta; geo experiments for YouTube/CTV/OOH.
- Creative taxonomy: map copy & angle to outcomes (win themes).
- MMM‑lite: weekly spend vs outcomes with simple Bayesian regression across channels (aggregate), used for forecasting, not micro‑ops.
Guardrails
- Fix eligibility (who could see a touch), deduplicate conversions, and standardize windows per channel.
- Publish error bars on MMM‑lite; it should disagree constructively with MTA—use both.
- Make budget shifts in 5–15% increments aligned to risk.
Outputs
- A monthly allocation memo with rationale + error bars.
- Creative leaderboard (themes, not only ads) tied to pipeline metrics.
C) Scale ($250k+/mo, multi‑market)
Decisions: Board‑level ROI, market expansion, brand vs performance split, saturation curves.
Useful story: Experiments + MMM as the backbone. Treat MTA as an operational view.
What to implement
- Continuous geo‑experiments (switchbacks, staggered start) for TV/CTV/OOH and paid social.
- Production MMM (Bayesian with priors): weekly, with carryover/adstock, diminishing returns, and seasonality; integrate costs and margins.
- Budget optimizer using MMM response curves + guardrails (min spend, max lift risk).
- Causal creative testing: ghost‑ads or audience split for big ideas.
- Unified reporting: finance‑grade ROI and CAC payback with CIs.
Guardrails
- Instrument SRM checks on experiments; pre‑register power/MDE.
- Recalibrate MMM quarterly with fresh experiments; don’t let it drift.
- Maintain a brand tracker and non‑click outcomes (search demand, direct, surveys).
Outputs
- Quarterly ROI read with confidence intervals and recommended reallocations.
- Scenario planning: “+$100k to YouTube vs +$100k to Meta vs +$100k to SEO.”
3) Model chooser: decisions each story is good for
| Model | Good decisions | Weakness | When to use |
|---|---|---|---|
| Last‑click | UX fixes, checkout/landing diagnostics | Under‑credits upper funnel | Early ops, micro‑conversion tuning |
| First‑touch | Discovery channels, content audit | Ignores nurturing | Early & PMF, editorial/media planning |
| Position‑based (U/W‑shape) | Ops velocity; assisted paths | Arbitrary weights | PMF; combine with lift tests |
| Time‑decay | Recency‑sensitive journeys | Penalizes long cycles | PMF where cycles are short |
| Rules‑based MTA (warehouse) | Cross‑channel ops, creative scoring | Click‑bias, cookie loss | PMF+ as a daily operating view |
| Experiment/lift (geo, ghost‑ads) | True incrementality by channel | Needs scale/power | Growth & scale; calibrate others |
| MMM (aggregate) | Mix, budget, market expansion | Coarse, needs discipline | Growth & scale for CFO/Board |
Meta‑rule: Use at least two stories: one for daily operations (MTA heuristic) and one for causal allocation (experiments/MMM). Resolve disagreements with new tests.
4) Privacy & signal loss (and how to adapt)
- ATT/ITP/cookie deprecation → expect fewer deterministic user paths.
- Shift to 1st‑party data: server‑side tagging, conversion APIs, login events, consent management.
- Modeled conversions: accept platform‑modeled lift as input; validate with your own holdouts.
- Granularity trade‑offs: more privacy means more reliance on experiments and aggregate models.
5) Incrementality experiments (crash course)
Design
- Choose unit (geo, user, time). Randomize cleanly; avoid spillovers.
- Size for MDE you’d act on (e.g., +8% CPA improvement) with 80–90% power.
- Pre‑register metrics, duration, and stopping rule.
Common designs
- Geo split/switchback (on/off in matched markets).
- Ghost ads (PSA placeholders) where available.
- Brand search holdout to estimate cannibalization.
- CTV/OOH on‑off with matched control regions.
Analysis
- Report point estimate + CI; convert to marginal ROAS with costs.
- Roll out with a 10–20% holdout to catch novelty effects.
6) Measurement architecture
- Event contract: canonical names, units, and owners (e.g.,
signup,qualified_lead,sql,won_deal). - UTM taxonomy (see below).
- IDs: first‑party user ID; click IDs where legal; campaign/creative IDs.
- Pipelines: server‑side GTM/Segment → warehouse (BigQuery/Snowflake) → BI.
- Dashboards: ops (daily) vs finance (monthly with error bars).
- Governance: QA checks (SRM, event volume shifts), versioned schema, and data SLAs.
UTM naming (copy‑paste)
utm_source=platform
utm_medium={paid_social|paid_search|email|referral|organic_social}
utm_campaign=yyyymm_topic_or_offer
utm_content=creativeAngle_adFormat_vX
utm_term={keyword or audience}
7) 90‑day roadmap (practical)
Days 1–14 — Instrumentation: UTM hygiene; server‑side events; map CRM stages; build last/first/assisted views; HYH survey live.
Days 15–45 — Heuristic MTA + creative taxonomy; run 1–2 geo lift tests (Meta/YouTube).
Days 46–75 — MMM‑lite pilot for channel‑level forecasting; publish error bars; reconcile with MTA; shift budgets 10–15% where both agree.
Days 76–90 — Decide: invest in production MMM + experiment calendar, or continue hybrid with quarterly lift tests.
Cadence
- Weekly ops review (creative, CPA, SQL rate).
- Monthly allocation review (MMM/MMM‑lite + experiments).
- Quarterly calibration (re‑run lift; refresh priors).
8) Templates & copy blocks
Allocation memo (excerpt)
- Recommendation: +$40k to YouTube consideration, –$25k brand search.
- Evidence: Geo‑lift +9% incremental leads; MMM curve shows improving marginal ROAS up to +$60k.
- Risk/mitigation: Creative fatigue → rotate two new angles; cap frequency at 2.
- Decision window: Next 30 days; reevaluate with new lift.
Survey (HYHBU)
“How did you hear about us?” with free‑text, plus optional chips: Search, LinkedIn, YouTube, Podcast, Friend/Colleague, Event, Other. Review and re‑map weekly.
Dashboard must‑haves
- Ops: CPA, SQL rate, win rate by channel/creative; assisted paths; creative themes.
- Finance: marginal ROAS with CIs; payback; channel response curves.
9) FAQ (for stakeholders)
- Why do models disagree? Different questions: clicks vs incrementality vs mix. Use disagreements to design new tests.
- Do we need MMM now? Only when spend is material and decisions are budget‑allocation, not just ops.
- Can we trust platform numbers? Treat them as one view; verify with holdouts.
- What about brand effects? Track search demand, direct, and HYHBU; model long carryover in MMM.
10) SEO kit
- Title (≤60): Attribution Is a Story: Choose the Right One for Your Stage
- Meta (≤160): A pragmatic guide to stage‑appropriate attribution—heuristics, lift tests, and MMM—so you can make decisions you can defend.
- Slug:
/attribution-as-a-story - Keywords: marketing attribution models, mta vs mmm, incrementality testing, geo experiments, time‑decay attribution, position‑based attribution, how did you hear about us survey, marketing mix modeling, marginal roas, budget allocation
Bottom line: Pick the story that fits your decisions today (heuristics for ops, experiments for truth, MMM for finance). Run them in parallel, reconcile with tests, and evolve your attribution as your spend, markets, and risks grow.
Add comment