Lens Marketing Attribution

As of: 2026-05-17

marketing-attribution

Mode skill. Toggled on the Analyst when the question is "which channel gets credit."

When to use

Use when the user asks "which channel is working," "what should we cut," "is paid social pulling its weight," "what's our ROAS by channel," or hands you a dashboard showing each channel claiming the same conversion. Use before any budget reallocation. If the question is "why are conversions down," route back to funnel-diagnosis — attribution sits downstream of a working funnel.

Trigger phrases:

• "What's working?"
• "Should we cut paid social?"
• "Meta says 200 conversions, GA4 says 60. Which is right?"
• "How do we measure brand?"
• "What's the real ROAS?"

Procedure

1. Count conversions in the window. Below 100 conversions, refuse to assign attribution. You have signal — proof a channel produces something — not attribution, which needs volume to distinguish contribution from noise. State the number, the floor, and the date the floor hits at current run-rate.

2. Pick the model that matches the question. Four common models, each answers a different question:

• Last-click — credits the final touch. Honest about closing channels (branded search, direct, retargeting). Lies about everything upstream. Use when asking "what closes."
• First-click — credits the channel that opened the journey. Honest about discovery channels (organic, paid social, podcasts, PR). Lies about retention and closing. Use when asking "what introduces us."
• Linear — splits credit evenly across touchpoints. Honest about multi-touch journeys. Lies about which touch did the work. Use when asking "how many touches before purchase."
• Time-decay — weights recent touches heavier. Honest about momentum. Lies about top-of-funnel patience. Use when the cycle is days, not months.

3. Name what the chosen model under-credits. Every model misses something. State it explicitly. Last-click systematically under-funds brand, organic, and discovery. First-click under-funds retention and conversion optimization. Linear and time-decay sit between, missing less of either extreme but adding their own distortions.

4. Read each platform's number as the platform's case for itself. Meta counts any user who saw or clicked a Meta ad within 7 days. Google Ads does the same. Shopify defaults to last-non-direct-click. GA4 uses data-driven attribution within observed sessions. These numbers sum to more than 100% of real conversions. Triangulate against one source-of-truth — the order or signup record with first-touch and last-touch UTMs captured at creation.

5. For small-budget operators, run MMM-lite. Full Marketing Mix Modeling needs a statistician. The lite version: hold every channel constant for two weeks, turn one off, measure the topline. Repeat per channel. Slow, low-tech, harder to fool than any dashboard. Works above the 100-conversion floor.

6. Stamp it. Write model used, window, total conversions, per-channel credit, and what the model under-credits to TEAM_MEMORY.md under ## Analyst. Channel-budget decisions read from this, not from platform dashboards.

Decision rules

• No attribution under 100 conversions. Below the floor, you are reading noise. Report direction, not credit.
• Use two models, not one. Run last-click and first-click in parallel. The gap is where under-credited work lives.
• Never trust a platform's self-reported credit. Every pixel claims credit it doesn't have alone. The sum across platforms always exceeds the real conversion count.
• One window, one model, one source-of-truth. Mixing produces numbers that look comparable but aren't.
• Cut on triangulated evidence, not single-platform ROAS. A channel that looks bad on last-click and good on first-click does discovery, not closing. Stage it; don't kill it.

Anti-patterns

• Treating Meta's reported conversions as Meta's conversions. The pixel reports influence, not causality.
• Cutting a channel based on a 30-day last-click report when the sales cycle is 90 days. The decision precedes the data window.
• Switching models mid-quarter and reporting the swing as a real change. Methodology shifts are not performance shifts.
• Building a custom model to make the existing budget look right. Confirmation bias dressed as analysis.
• Running holdout tests on a channel during its seasonal peak. The test measures the season.

Before / after

Brief: "Meta says 240 conversions last month at $35 CPA. Looks great. Double the budget?"

Before:

Yes — $35 CPA well under our $90 target. Scale Meta 2x next month.

After:

Window: 30 days. Total conversions in order record: 310. Meta-reported: 240 (claims 77%). Google-reported: 180 (claims 58%). Sum across platforms: 560, well above 310 — every platform over-claims. Reading order-record UTMs: last-click Meta 95 (31%), last-click organic 110 (35%), last-click direct/branded 75 (24%). First-click Meta 160 (52%), first-click organic 85 (27%), first-click podcast 35 (11%). Diagnosis: Meta does discovery; organic and branded search close. Pure last-click under-credits Meta by ~65 conversions; pure first-click over-credits it. Recommendation: hold Meta budget one cycle, run a 2-week MMM-lite holdout to confirm causal contribution before scaling. Re-measure 2026-05-31.