Triangulation. click attribution + mix model + holdouts. When methods agree, the read is board-grade.
Quarterly read. Stable runs a daily causal model (BL · Baseline Lift) on the brand's pipes - no vendor required. When the brand also has Northbeam, Haus, Measured, or WorkMagic, we ingest those outputs and triangulate. Holdouts = causal standard. Mix model = brand-layer value pixels can't see. Triangulation = the read CFOs trust.
Synthetic incrementality on Stable's pipes alone. Causal proof without procuring a measurement vendor.
Synthetic-control / matched-market design (synthetic time-series). Lite geo holdouts. Time-based holdouts. Audience holdouts. Stable runs these on the brand's own data - geo + spend + outcome already in the pipes via Northbeam + GA4 + brand warehouse. No measurement-vendor procurement required.
Four methods side by side · TV lift % · with the range each method reports
How to read it. Each gray bar is one method's range; the dot is the number it lands on. The orange bar at the bottom is the combined read across all four methods. When the bars overlap, the methods agree and the read is board-grade. When they don't, the note below explains which method is the outlier and why. click attribution at +5% reads lower than the rest because click-attribution can't see the halo lift. Holdout and synthetic methods close that gap.
When the brand has them: triangulate against vendor reads
Combined read · weighted across 4 methods
disagreementMethodologyMethod averaging + path-share attribution + assumption stress tests
Method averaging. Each method (click attribution / mix model / holdout / synthetic incrementality) returns its own estimate + CI. Weights default holdout 0.35, synthetic 0.30, mix model 0.25, click attribution 0.10 - reflecting causal-rigor ordering. Weighted estimate + range propagation gives the "board-grade" number. Per-brand re-weighting requires methodology sign-off + log entry.
Path-share attribution. A fourth axiomatically-fair lens, distinct from mix model and click attribution, that scores each channel by its marginal contribution across observed conversion paths. Open-source library, well-cited in causal-attribution work.
Assumption stress tests. When methods disagree by >30%, the model runs three stress tests: placebo treatment (does a fake intervention show the same lift?), random common cause (does adding a noise variable change the answer?), and subset analysis (does the result hold on half the data?). Identifies which assumption is driving the disagreement (selection bias, confounder, or measurement error).
Why this is Stable's IP #2. Northbeam gives click attribution. Haus gives mix model. Stable gives synthetic incrementality. No vendor produces the combined read + the explicit disagreement diagnostic. The convergence engine is the moat.
The CFO-trust frame. "Mix model says +9%. click attribution says +7%. Holdout says +12%. Combined read lands at +9.83% with range +1% to +35%, and the methods disagree most on brand-search where click-attribution misses the halo." That sentence is what gets past a board audit. Single-vendor reads can't produce it.
Decision matrix · what the team does when each layer reads
| If this layer reads | The signal is | The action |
|---|---|---|
| L0 · delivery health fails | Spend below floor or video completion rate <90% | Fix delivery before reading anything else. |
| L2 · confidence layer disagrees | The two math approaches point opposite directions | Diagnose what's odd in the data. Don't act on either number. |
| L1 · site lift confirms | Brand search, direct, organic rise during flight | Tie the lift back to the CTV window in the readout. |
| L3 · marginal CAC has headroom | Below target with positive slope | Scale CTV spend. Re-read in 14 days. |
| L4 · ad-stock spikes early | CAC moves before day 10 | It's not the CTV change. Look at the rest of the mix. |
| L5 · LTV quality slips | One-time-to-subscriber rate declines while volume grows | Hold spend. Audit creative and targeting before scaling. |
| L7 · triangulation agrees | click attribution, mix model, holdout point the same direction | Take the read to the board with confidence. |
| All layers break together | Correlations fall, halo dark, marginal CAC over ceiling | Cut. Re-enter after creative refresh + 30-day pause. |
| Total Impact lens | DTC +23.1%. Top retail account +$421K · moderate confidence. 509M brand-layer TV impressions. | Report the Sep25-Feb26 readout. $156.3M total, +$3.33M trend-adj lift. |
What this dashboard explicitly does NOT do
- Not an incrementality test. Halo CAC is a parallel re-denomination, not a holdout.
- Not a mix model. We ingest Haus / Measured outputs; we don't build them.
- Not an click attribution replacement. Northbeam stays. WorkMagic stays.
- Not a media buying platform. UA places. Agency / in-house places. Stable measures.
- Not a single-source-of-truth. When methods disagree, diagnose data shape. Don't act on either.
- Not a real-time platform. Cadence is daily / weekly / monthly / quarterly. The cadence is the discipline.