Phase 1 (Weeks 0‑2): Instrument & pre‑test

• Pick one north‑star outcome (e.g., qualified lead, first purchase) and freeze conversion definitions.
• Pull 12–16 weeks of historical, weekly data by geo (sales, traffic, store density, seasonality).
• Establish pre‑test baselines for geos—Google’s own guidance suggests 4–8 weeks of pre‑test when running geo experiments.⁶

Phase 2 (Weeks 3‑6): Run a geo‑holdout on your highest‑spend channel

• Randomize 8–12 DMAs into test/control and hold spend flat elsewhere.
• Use GeoLift (synthetic control) to select balanced markets, run power analysis, and compute lift with jackknife+ or conformal intervals.⁷ ⁸
• Expect wider CIs with few geos; underpowered tests will swing from +2% to +50% “lift” just on noise.⁹
• If your stack supports it, Google’s Conversion Lift (geography) can complement the open‑source workflow, though it tends to require higher budgets than user‑level lift.¹⁰

Deliverables: incremental conversions, iROAS, and 90% CI for Channel A, plus a sanity check that test and control reconverge after the test window.¹¹

Phase 3 (Weeks 5‑8): Layer a people‑level lift where feasible

• On Meta or YouTube, run platform Conversion Lift (sales outcome) or Brand Lift (creative outcome). Lift tests typically report at 90% confidence and handle user‑level randomization.¹² ¹³ ¹⁴
• If you must keep auction dynamics identical across cells, a PSA/placebo control can work (showing controls a charity PSA). But be careful: platform optimizers often treat PSAs differently, biasing reach and wrecking comparability; ghost‑ad style controls avoid that pitfall.³ ¹⁵ ¹⁶ ¹⁷

Deliverables: channel‑specific incremental conversions and cost‑per‑incremental with CIs; for video, brand‑metric direction by creative.

Phase 4 (Weeks 6‑10): Triangulate with MMM—but calibrate it

• Fit a lightweight MMM (Meta’s Robyn or Google’s Meridian) on weekly aggregates with adstock and saturation to get cross‑channel response curves.¹⁸ ¹⁹ ²⁰ ²¹
• Calibrate MMM using your experimental ground truth (Robyn includes an MAPE.LIFT objective for this; Meridian explicitly supports integration of lift results).²² ²¹
• Output mROAS by spend level to plan the next quarter.

Deliverables: calibrated MMM with credible intervals and a budget reallocation recommendation consistent with your lift results.

What results to expect (and how to read them)

Lift by channel. Expect a spectrum. For high‑intent brand search, large incumbents sometimes see near‑zero short‑term lift (users would have arrived anyway), while other advertisers show strong incrementality—Google’s meta‑analysis averages ~89% incremental clicks when search is paused.¹ ² Read the local result, not the average.

Confidence intervals. Most platforms report 90% CIs for lift; your internal standard should be 90–95% depending on risk tolerance. A CI like “+12% ± 6%” is actionable; “+26% ± 24%” is not.¹² ¹⁴ ⁹

Creative effects. Don’t ignore creative. Controlled Brand Lift studies routinely show winners and losers among assets; picking the right asset can meaningfully change downstream conversion lift.¹³

Geo‑holdouts, PSA tests, MMM: when to use what

A CFO‑ready readout (what to put on one slide)

• iROAS by channel with 90% CI (from lift).
• Cost per incremental outcome vs. last‑click CPA.
• Creative winner (if video) from Brand Lift.
• Calibrated MMM reallocation: “Move 15–25% from Channel B to Channel A at current spend levels” with expected range.
• Next test queued (e.g., holdout on retargeting, PSA‑controlled CTV).

Common failure modes (and how to avoid them)

• Underpowered tests. Too few geos or too short a test produce wide CIs; run power analysis first.⁸ ⁹
• Biased controls. PSA ads can distort who you reach; Google has shown how optimizers treat PSAs differently. Prefer platform lift or ghost‑ad logic.¹⁵ ¹⁶
• Cookie contamination. Running overlapping platform tests can pollute each other’s control cells; stagger or split by geo.¹⁸
• Un‑calibrated MMM. Treat MMM as triangulation, not truth—calibrate to experiments or don’t use it for budget moves.²² ²¹
  
  

Bottom line: Last‑click is a story; incrementality is evidence. In 90 days, you can replace the former with the latter—and move the budget with conviction.

Endnotes

1. Thomas Blake, Chris Nosko, Steven Tadelis, “Consumer Heterogeneity and Paid Search Effectiveness: A Large‑Scale Field Experiment,” Econometrica / NBER working paper (eBay brand vs. non‑brand effectiveness). (Haas School of Business, NBER)
   
   
2. David X. Chan et al., “Incremental Clicks Impact of Search Advertising” (Search Ads Pause meta‑analysis; ~89% incremental clicks on average). (Google Business, Google Research)
   
   
3. Garrett Johnson, Randall Lewis, Elmar Nubbemeyer, “Ghost Ads: A Revolution in Measuring Ad Effectiveness,” Journal of Marketing Research (ghost‑ad methodology for causal lift at auction scale). (SAGE Journals)
   
   
4. Jon Vaver, Jim Koehler, “Measuring Ad Effectiveness Using Geo Experiments,” Google Research (geo randomization across markets). (Google Research, Google Research)
   
   
5. GeoLift documentation (synthetic control methods; open‑source geo experimentation workflow). (Facebook Incubator)
   
   
6. Unofficial Google Data Science blog, “Estimating causal effects using geo experiments” (structure with 4–8 weeks pre‑test). (Unofficial Google Data Science)
   
   
7. GeoLift Walkthrough (market selection, power calculations, execution). (Facebook Incubator)
   
   
8. GeoLift methodology for confidence intervals (conformal and jackknife+). (Facebook Incubator)
   
   
9. Polar Analytics, “Why Most Geo Tests Fail” (interpretation of wide CIs; power constraints with small geo counts). (Polar Analytics)
   
   
10. Google Ads Help, “Conversion Lift based on geography” (budget considerations; aggregated by non‑overlapping regions). (Google Help)
    
    
11. Expedia Group Tech, “Measuring marketing success: incrementality & geo‑testing” (post‑test reconvergence and CI reporting). (Medium)
    
    
12. Meta for Developers, “Conversion Lift Measurement” and Meta Business Help, “About confidence in your tests” (lift methodology and 90%+ confidence guidance). (Facebook Developers, Facebook)
    
    
13. Google Support (DV360) “Understand your Brand Lift measurement data” and YouTube Brand Lift case guidance (creative study design and thresholds). (Google Help)
    
    
14. Google Ads Help, “Lift measurement statuses and metrics” (reporting with 90% confidence intervals). (Google Help)
    
    
15. Think with Google, “A Revolution in Measuring Ad Effectiveness” (how PSA controls can bias auctions; why ghost‑ad logic is superior). (Google Business)
    
    
16. Remerge, “Incrementality Tests 101: ITT, PSA and Ghost Ads” (PSA pros/cons; cost and bias trade‑offs). (Remerge)
    
    
17. Viant, “Lift & Incrementality in Programmatic Advertising” and “Why Your Control Group Could Be Ruining Incrementality” (PSA/ghost‑ad overview; control comparability). (Viant Technology LLC)
    
    
18. Meta’s Robyn MMM (open source; adstock, saturation; documentation). (Facebook Experimental)
    
    
19. Google’s Meridian MMM (open source; launched 2025; integration path). (Google for Developers, Google Business)
    
    
20. Google’s LightweightMMM (Bayesian MMM foundations). (GitHub, LightweightMMM Documentation)
    
    
21. eMarketer coverage of Meridian’s aims (triangulation with experiments and attribution). (EMARKETER)
    
22. Robyn calibration to ground truth (MAPE.LIFT objective to minimize gap to experimental lift). (Facebook Experimental)