On this page
- What it is
- How it works
- Core method
- Market design
- Benchmarks (as-of simulation study — publication date unconfirmed)
- Recast simulation study (32,000 model fits)
- Key terms
- What practitioners report
- GeoLift vs. other approaches
- vs. CausalImpact (Google)
- vs. Meridian GeoX (Google)
- vs. commercial incrementality platforms (Haus, Measured, INCRMNTAL)
- vs. Media Mix Modeling (MMM)
- Limitations and structural risks
- Contradictions
GeoLift
GeoLift
GeoLift is Meta's open-source R library for measuring incremental advertising lift at a geographic market level using aggregated (not user-level) data. It provides an end-to-end workflow — market selection, power analysis, experiment design, and result analysis — and is used by ecommerce brands and media agencies as a free alternative or complement to commercial Incrementality Testing platforms.
What it is
- GeoLift is MIT-licensed and published under the facebookincubator GitHub organisation (github.com/facebookincubator/GeoLift); README status: "Active — stable, usable, with active development" (as-of 2026-10-14). (Primary — GitHub)
- The library requires R 4.0.0+ and the augsynth package, which must be installed from GitHub separately via the
remotespackage; it is not available on CRAN as a dependency. (Primary — GitHub) - Unlike Meta's proprietary in-platform Conversion Lift test (user-level randomisation embedded in Ads Manager), GeoLift operates at the market/region level and is platform-agnostic — it can measure incrementality for Meta, Google, TikTok, TV, or any channel. (Secondary — Triple Whale)
- The primary GeoLift user base is described in the README as brands accessing Meta Marketing Science Partners resources; the barrier to adoption is operational complexity, not cost — the software is free. (Secondary — GitHub, Jellyfish)
How it works
Core method
- GeoLift's effect estimator is the Augmented Synthetic Control Method (ASCM) of Ben-Michael, Feller & Rothstein (2021), implemented via the augsynth R package, using ridge-augmented Synthetic Control weights with the ridge penalty lambda chosen by leave-one-period-out cross-validation under the 1-SE rule. (Secondary — mlsynth replication docs)
- GeoLift fits treated markets as their mean (not their sum); using the sum shifts the treated series outside the donor convex hull, changing the p-value substantially (≈0.68 vs ≈0.01 in the documented example). (Secondary — mlsynth replication docs)
- Inference: conformal inference (Chernozhukov, Wüthrich & Zhu 2021) — the model is refit on all periods under the joint null, and the post-period test statistic is compared against permutations of the residual path to produce a p-value. (Secondary — mlsynth)
- Unit fixed effects are on by default; each unit's pre-period mean is subtracted before fitting so the donor pool matches trend shapes rather than levels. (Secondary — mlsynth)
- Outputs: Lift (%), incremental revenue, and iROAS (Incremental Return on Ad Spend) (incremental revenue ÷ ad spend). (Secondary — Triple Whale)
Market design
- GeoLift recommends DMAs (Designated Market Areas) as geographic units; US states work but provide fewer markets for matching; zip codes are rarely suitable due to volume noise. (Secondary — Triple Whale)
- Data requirements (Recast): ≥6 months of daily data or ≥1 year of weekly data; ≥20 geo units; a sales or conversion dependent variable. (Secondary — Recast)
- Practitioner guidance (Towards Data Science): ≥25 pre-treatment periods; ≥10 geos (ideally 20+); test duration ≥1 purchase cycle (≥15 days for daily, 4–6 weeks for weekly); pre-history of 4–5× test duration (ideally 52 weeks). (Secondary — Towards Data Science)
- GeoLift supports staggered rollouts and retrospective (quasi-experimental) analysis, not only prospective randomised designs. (Secondary — Towards Data Science)
- GeoLift includes a pre-launch power analysis via simulation, allowing practitioners to estimate the minimum detectable effect and select the optimal market combination before spending. (Primary — GitHub; Secondary — Triple Whale)
Benchmarks (as-of simulation study — publication date unconfirmed)
Recast simulation study (32,000 model fits)
Recast tested Meta GeoLift, Google CausalImpact, Google Matched Markets, and CausalPy across four scenarios. Scenario: 7.5% true lift, 15-day post-period:
| Tool | Coverage | False positive rate | False negative rate |
|---|---|---|---|
| Meta GeoLift | 92–95% (most calibrated) | 3–5% (lowest) | 89–96% (!!!) |
| Google CausalImpact | 70–72% (lowest) | 28–30% | 34–48% |
- Under a short pre-period (30 days), GeoLift missed the true 7.5% lift in 95.7% of simulation runs — near-total loss of statistical power. (Secondary — Recast simulation study)
- Recast self-caveat: results may partly reflect how the synthetic data was built; rankings could shift under real-world conditions. (Secondary — Recast)
- No tool reached 95% coverage and reliably detected real effects — each traded false positives against false negatives. (Secondary — Recast)
Key terms
| Term | Meaning |
|---|---|
| ASCM | Augmented Synthetic Control Method — GeoLift's core estimator |
| Conformal inference | GeoLift's inference method; compares post-period test stat against permutation residuals |
| DMA | Designated Market Area — recommended geographic unit for US GeoLift tests |
| MDE | Minimum Detectable Effect — estimated via GeoLift's pre-launch power simulation |
| Augsynth | R package implementing ASCM, required by GeoLift; installed from GitHub |
| False negative rate | Share of real effects GeoLift fails to detect (89–96% under short pre-periods) |
What practitioners report
- Jellyfish (media agency) positions GeoLift as one leg of a "triple alliance of measurement" alongside Media Mix Modeling (MMM) and Data-Driven Attribution (DDA), and has built internal tooling to speed up GeoLift design and analysis for clients. (Secondary — Jellyfish blog)
- Jellyfish states GeoLift requires "several weeks or months" to understand its many parameters fully, citing this as the primary implementation cost despite the software being free. (Secondary — Jellyfish blog)
- Brumate (insulated drinkware) ran a GeoLift test via the open-source library to measure omnichannel sales impact (DTC + Amazon + wholesale) of reach-and-frequency Meta campaigns; early read showed 1.5% conversion lift vs control after one week. (Secondary — Don't V*LOOKUP, December 2023)
- Recast launched "GeoLift by Recast" (2025) — a commercial managed-service wrapper on the open-source library targeting brands below the technical threshold for running GeoLift independently in R. (Secondary — Recast, 2025)
GeoLift vs. other approaches
vs. CausalImpact (Google)
- Google's CausalImpact is the most commonly compared open-source alternative; practitioner decision-making between the two is an active topic in measurement communities. (Secondary — Eppo blog)
- In Recast's simulation study, CausalImpact detected real lifts more often than GeoLift (34–48% FNR vs 89–96%) but at the cost of lower coverage (70–72%) and a 28–30% false positive rate and systematic overestimation of lift by 2–4 percentage points. (Secondary — Recast simulation study)
vs. Meridian GeoX (Google)
- Google's Meridian GeoX is an open-source geo experiment library (GA: 2025) that, like GeoLift, produces incrementality results usable as calibration priors for an MMM; in GeoX's case, results feed directly into Meridian. (Primary — Google Developers docs)
- Google claims GeoX reduces geo experiment costs by 31% for large advertisers and cuts design generation time by over 94% vs "other open-source solutions and legacy solutions" — but these are vendor-reported figures with no published sample size, named comparison set, or independent verification. (Secondary — PPC Land, 2025)
vs. commercial incrementality platforms (Haus, Measured, INCRMNTAL)
- SegmentStream's 2026 list of top incrementality tools does not include Meta GeoLift as a named tool, ranking Haus as "best for paid media teams running frequent geo lift tests" and Measured as best for large advertisers. (Secondary — SegmentStream, 2026; vendor source with competing product)
- Commercial platforms offer hosted interfaces, experiment libraries, and enterprise controls (audit trails, team access, support) that GeoLift lacks. (Secondary — Sellforte, 2026)
- GeoLift's free software cost is offset by data-science resourcing requirements; commercial tools bundle the analyst capability. (Secondary — multiple)
vs. Media Mix Modeling (MMM)
- GeoLift is designed to complement MMM: GeoLift results can serve as Bayesian priors or guardrails to recalibrate MMM models, providing causal validation that correlation-based MMM cannot produce on its own. (Secondary — Recast)
- MMM operates over longer strategic horizons as a continuous picture; GeoLift provides a causal snapshot at shorter test durations. (Secondary — Triple Whale)
Limitations and structural risks
- Geographic contamination: people travel, commute, and share ads online, so ad exposure can spill into control regions, violating the isolation assumption. (Secondary — Recast)
- Channel exclusion impossibility: organic search, podcasts, and PR cannot be switched off by region, making geo holdout tests structurally impossible for those channels. (Secondary — Recast)
- Budget reallocation error: if test-region budgets are excluded but total daily spend stays constant, platforms push that budget into control regions; Target CPA and Target ROAS bidding exacerbate this. (Secondary — SegmentStream)
- Seasonality bias: weights built on pre-test data may stop reflecting control behavior once a seasonal spike (e.g., Black Friday) arrives. (Secondary — SegmentStream)
- Region selection: including geographic outliers (e.g., California) or splitting by fixed proportions rather than behavioral similarity weakens the synthetic control. (Secondary — SegmentStream)
- Delayed effects: channels with lagged effects (TV, brand campaigns) may begin before a test or continue after, causing short tests to undervalue those channels. (Secondary — Recast)
- IP-based geolocation accuracy: much less accurate at region/city level than country level; VPNs and privacy rules degrade this further. (Secondary — Recast)
- False negative paradox: GeoLift's conservative calibration (low FPR) comes at the cost of very high FNR — practitioners need longer tests, more markets, or larger spend changes than most guides indicate. (Secondary — Recast simulation study)
- Version drift: GeoLift's published walkthrough ATT (155.556) differs from current augsynth output (156.81) due to library version drift not reflected in official documentation. (Secondary — mlsynth)
Contradictions
[!contradiction] Andava guide states Pearson correlation thresholds for market pair selection of "0.80 minimum" in one section and "above 0.90 only" in another — an internal inconsistency within a single practitioner guide [Andava]. No primary source resolves the correct threshold.