All answers

The Prevailing
Answers for
Business model:
Business model, pick up to two

How do we calibrate a marketing mix model or response curve with our lift and geo test results, and how often should we recalibrate it?

The PrevailingPublished Oct 28, 2026

Our answer

Feed each finished lift or geo test into your mix model as a prior that keeps its interval. Weight several tests by relevance, recency, precision and spend. Refresh on your budget cycle, and recalibrate early after big changes.

Confidence
Medium
Strongest proof
1 controlled test
5 Meta rules, 2 datasets, 4 practitioners, 12 opinions or claims
Sources
15 sources
24 statements

New here? Take the 1-minute tour

Why this is the answer

3 reasons and our judgment

No single formula turns a test into a prior (Google). Meridian adds uncertainty for spend, duration and age, with a 52-week half-life by default. Google suggests refreshing quarterly, annually or on your budget cycle.

Several tests should be weighted, not averaged. Sellforte, a marketing mix modeling firm, weights them by relevance, recency, statistical confidence and spend. It adds that a calibrated model need not copy the test's point estimate.

A lift-calibrated curve predicted a peak season well in one case. Lifepro planned peak Meta spend on such a curve and landed within about 2.2% of the predicted ROAS, in a vendor-run test (WorkMagic).

Our judgment

  • Enter each of your own finished tests as a prior that keeps its interval. Use tests that match the model's channel, market and KPI, not other brands' benchmarks, because borrowed priors can shape a weak-data model more than marketers realize. Apply each to the channel, KPI, dates and spend it covered, and let the model fit the rest. Google says the translation adds uncertainty beyond the standard error, Sellforte wants the sources made coherent, not copied, and Meta's Robyn treats experiments as ground truth.
  • Use holdouts against zero spend for a channel's average return and tests at two or more spend levels for its curve. Register only the holdouts in a tool such as Meridian that expects a zero-spend baseline. This matches the two-level tests in Finding where marginal returns fall off, and Robyn calibrates saturation by ground truth whenever possible.
  • Weight several tests by relevance, recency, precision and spend rather than averaging them or keeping only the latest. Let older tests lose weight instead of deleting them, as Google's CalibrationBuilder does with spend and duration adjustments.
  • Refresh at least as often as you make budget decisions, and fold each finished test in at the next refresh. Don't refresh faster than you can act on or check the result, and check each refresh against a new test or a holdout forecast. LiftLab ties refreshes to new evidence, and Sellforte starts the workflow whenever a test ends.
  • Recalibrate early after a big change in budget, channel mix, pricing or where customers buy, or when the model's incremental return rises while contribution margin doesn't (Common Thread Collective). Test next where the model is least certain or disagrees with your tests, and size budget moves by the step and retest rule in Acting on a finished lift or geo test, not by the curve alone. Sellforte keeps the prior wide when tests disagree, Michael Kaminsky of Recast warns two model runs can name different best channels, and the Lifepro curve was checked against the result.

Our judgment, built from Google, Sellforte, Meta’s Robyn, Sonia Chung, PyMC Labs, Dew, Padilla and Shchetkina, Circana, LiftLab, Haus, Common Thread Collective, Michael Kaminsky and WorkMagic: our reading of these sources, not a finding from any one of them.

What the sources say

24 statements from 15 sources
  • MetaNov 12, 2026Supports

    Meta, summing up talks with finance and marketing leaders, says to agree with finance on the outcome before a test and to start with a high-spend channel such as branded search, where last-touch and lift likely differ.

    “Several leaders recommended starting with one high-spend channel where last-touch attribution and incremental impact are likely to tell different stories.”

    Read it on facebook.com

  • AdExchangerOct 19, 2026Supports

    Boathouse's strategy chief says that for one client an off-the-shelf mix model credited paid media with about 75% of new customers, while a custom model on the same data said under 1%. Check models with lift tests.

    “A result below 1% deserves as much scrutiny as a result of 75%.”

    Read it on adexchanger.com

  • Meta's open-source Robyn MMM says models are usually refreshed twice a year or less, and its frequent-refresh feature is not yet thoroughly tested. It says adstock and saturation should be calibrated.

    “this feature is not yet tested thoroughly and might provide instable refresh results.”

    Read it on facebookexperimental.github.io

  • Robyn (Meta open-source MMM)Oct 15, 2026Supports

    Meta's Robyn mix model docs say ad returns diminish, so budgets should follow the next dollar's response, and models should be calibrated with lift or geo tests. A relayed whitepaper puts uncalibrated models 25% off.

    “uncalibrated models show 25% average difference to the ground truth”

    Read it on facebookexperimental.github.io

  • HausOct 14, 2026Supports

    Haus says about half of YouTube tests are built to detect lifts under 4%, and brands that missed a benchmark on a first test beat it on the retest 44% of the time, so one test is not a verdict.

    “one test is a read, not a verdict.”

    Read it on haus.io

  • PyMC Labs (PyMC-Marketing)Oct 12, 2026Supports

    PyMC-Marketing's documentation says lift-test results can constrain a media mix model's response curves and improve marginal return estimates. Improvement is not guaranteed, and its example uses simulated data.

    “Calibrating the MMM with valid lift estimates can change its channel responses, but improvement in every parameter or mROAS estimate is not guaranteed.”

    Read it on pymc-marketing.readthedocs.io

  • GrünsOct 12, 2026Supports

    Gruns' growth head matches the tool to the decision: incrementality tests, a mix model and marginal return for big budget calls, and attribution and platform data for daily moves.

    “I'm also not letting attribution decide whether Meta deserves another $500K.”

    Read it on operatorscontent.com

  • GoogleSep 24, 2026Supports

    Google says there is no single formula for turning an experiment into a mix model prior: combine evidence and allow extra uncertainty, since lift tests and the model measure different things.

    “There is no single formula to translate an experiment result into a prior.”

    Read it on developers.google.com

  • GoogleSep 23, 2026Supports

    Google's Meridian docs say its CalibrationBuilder turns lift experiments into ROI priors, discounting older tests with a fixed 52-week half-life. It suits tests against zero spend, not reduced spend.

    “Recency calculations use a fixed half-life of 52 weeks (1 year).”

    Read it on developers.google.com

  • GoogleSep 11, 2026Supports

    Google's Meridian flags channels whose results look uncertain, biased or implausible (a score below 67.5) as priorities for experiments, putting higher-spend channels first.

    “Spend-weighting also prioritizes calibration for higher-spend channels due to their greater business importance.”

    Read it on developers.google.com

  • SellforteSep 2, 2026Supports

    Sellforte, which sells MMM software, says its tool pulls a brand's Meta Conversion Lift results in daily and updates the model's priors. It claims higher lift-based incremental ROI than GA, with no data shown.

    Read it on sellforte.com

  • LiftLabAug 28, 2026Supports

    LiftLab, which sells mix models, says each experiment should feed back into the model, and the channel with the most uncertainty should be tested next. No data.

    “A fast MMM should refresh because new evidence changed the model, not because time passed.”

    Read it on liftlab.com

  • SellforteAug 21, 2026Supports

    Sellforte, an MMM vendor, says not to average a Meta conversion lift result and a geo test result. Translate both to the model's KPI and period, combine the distributions, and weight by relevance and precision.

    “the team should not average the two headline iROAS figures”

    Read it on sellforte.com

  • SellforteAug 18, 2026Supports

    Sellforte, which sells mix models and testing, says to weight multiple lift tests by relevance, recency, confidence and spend when calibrating a mix model, leaning on the best test rather than averaging.

    “The calibration should instead draw most heavily on the highest-quality test.”

    Read it on sellforte.com

  • HausAug 14, 2026Supports

    In this episode HexClad's Connor Rolain suggests a mature Meta-led brand start with tests inside Meta and save holdouts for new channels; Haus's Olivia Kory says a Meta holdout helps calibrate a mix model.

    Read it on youtube.com

  • SellforteAug 11, 2026Supports

    Sellforte says a calibrated mix model can legitimately show a different incremental ROAS than a lift study, and a large gap should prompt a calibration audit. Its numbers are hypothetical.

    “Calibration should make the two sources coherent. It should not force a broad model output to copy one experimental point estimate in every period.”

    Read it on sellforte.com

  • CircanaJul 29, 2026Supports

    Circana, which sells mix modeling, says there is no universally correct cadence for refreshing it. Match it to budget decisions, and refresh after a major campaign, pricing change or media-mix shift. No data.

    “There is no universally correct cadence, only the one that fits your decisions, your category, and your data.”

    Read it on circana.com

  • RecastJul 28, 2026Caveat

    Michael Kaminsky, whose firm Recast sells mix models, says most businesses do not need one and it can mislead them; he cites studies where last-touch was about 85 percent as good as experiments in simpler mixes.

    “most businesses do not need an MMM. In fact, an MMM will be net negative for them because it will actively mislead them.”

    Read it on mdm.co

  • experiMENTALJul 15, 2026Supports

    Sundar, formerly Uber's marketing science head, says refresh a mix model every six months, or quarterly for promotion-heavy ecommerce, and check its channel curves with real budget-shift tests.

    “feed them back into the model to make it smarter”

    Read it on experimental.beehiiv.com

  • Common Thread CollectiveJul 2, 2026Supports

    CTC says even a well-calibrated incremental ROAS should answer to revenue and contribution margin; if iROAS improves while margin is flat or falling, recheck the measurement.

    “Even a well-calibrated incremental return on ad spend should always be subordinate to the realities of revenue and contribution margin”

    Read it on commonthreadco.com

  • GoogleJun 30, 2026Supports

    Google's Meridian guidance says refresh a mix model as often as you like, such as quarterly or annually, by adding new data to the old and keeping a 2-3 year window. Fitting new data alone is generally not recommended.

    “Model refreshes can be done as frequently as you would like.”

    Read it on developers.google.com

  • AdExchangerApr 1, 2026Disagrees

    AdExchanger reports Jones Road Beauty feeds its own lift tests into Haus's causal mix model. Google branded search was not incremental at about $3,000 a day, and Meta was incremental but overspent.

    Read it on adexchanger.com

  • WorkMagicJan 29, 2026Supports

    WorkMagic, which sells mix modeling, says Lifepro calibrated its model with lift tests, raised Meta spend about 200% at BFCM and landed within about 2.2% of predicted ROAS; 57.9% of incremental orders came on Amazon.

    “Lifepro scaled BFCM spend by 200%, with actual ROAS landing within ~2.2% of the predicted outcome”

    Read it on workmagic.io

  • This academic paper says standard mix data often cannot tell diminishing returns from effects that change over time, which point to different budget splits. It proposes experiments that vary spend.

    “nonlinear and time-varying effects are often not identifiable from standard marketing mix data”

    Read it on arxiv.org

In the full answer

For readers who sign in
  • What to do this week3 steps
  • The numbers behind it3 figures from cited sources
  • Where each source standsA chart placing each source on the question
  • The strongest disagreementThe best case against the answer, from a cited source
  • How sure we areWhat its confidence rests on, and what would change the answer

Read next

More in Measurement