The MarketTwin™ method
Compare every market that got more media with a twin built from markets that didn’t. The gap is what the ads did. Written for marketers, checkable by your CFO.
The problem
Click attribution was built for online checkouts. For locations, most customers never click, and the ones who do get claimed by every platform at once.
A billboard, a radio spot or a video seen on the couch leaves no trail to the front desk.
Each platform counts every customer it touched. Add the reports up and you often get more sales than you made.
A regular who saw an ad still counts. The real question is who came because of it.
The method, step by step
Each step is a documented module in the engine, described in plain terms.
Sales by location (guests or transactions, revenue if you have it) and ad spend by market and channel. Daily rows roll up into weeks.
Totals only. No customer records.
Markets whose spend rose become media markets. Steady-spend markets become twin candidates. For a planned test, you pick both.
Synthetic control: the weighted blend of steady-spend markets that best matched the media market’s sales before media started. Weights are never negative and sum to 1.
Augmented synthetic control and synthetic diff-in-diff both run; the lower back-test error on your data wins.
The twin runs on through the measured weeks as the counterfactual. Actual minus twin is incremental guests and revenue.
Sized from back-tests on your own history: hundreds of fake spend changes where the true lift is zero, plus planted lifts of known size. A market only counts as a win when the low end clears zero.
Calibrated with conformal inference, then graded A to D.
Each location inherits its market’s counterfactual by its share of pre-period guests. New stores stay out, so an opening never looks like ad performance.
Market lift is explained by channel spend: guests per dollar, never negative, shrunk toward the brand average. It is labelled modeled. Weak channels get a holdout test.
New locations are compared with a ramp built from your own openings without launch media, adjusted for seasonality. The 16-week gap is what launch media added.
Spend in markets with no measurable lift moves to the lowest-cost proven market. Returns diminish, so the next guest there costs more than the average one.
How we keep it honest
A tool that always finds lift isn’t measuring. These guardrails run on every read.
The engine falls back to a baseline-only model, the result carries an honest grade, and you get a prompt to plan a holdout test.
Each result shows its grade, what was accounted for, the back-test record on your data and one next step.
See the five layers of the engine →Weekly sales history before the measured period.
A road closure, new competitor or storm hits one market but not its twins.
A machine-learned baseline absorbs weather, holidays, school breaks and local events. Other shocks show up as a weaker pre-period fit and a wider range.
Media in one market reaches shoppers in a nearby twin, shrinking the measured gap.
Exclude neighboring markets from the twin pool, or plan a test with holdouts far apart.
Compare methods
MarketTwin reads spend changes you already made, and tells you when a geo test is the better next step.
| Method | In-store sales | Location detail | Speed to read | Cost | Needs a data team | Watch out for |
|---|---|---|---|---|---|---|
| Platform attribution | Online clicks; modeled store visits on some | Where store-visit reporting exists | Real time | Included with ad spend | No | Each platform grades itself; credit is not cause |
| Marketing mix model | Yes, total sales | Usually national or regional | Months to build; quarterly refresh | High: consultants or in-house team | Yes | Best for long-run, all-channel planning |
| Geo lift test | Yes, the gold standard | Market level | Weeks per test, planned ahead | Holdout markets give up media | Usually | One question per test |
| Countlift MarketTwin | Yes, from your sales totals | Every market and location | Reads past spend changes; refreshes weekly | Software subscription | No | Needs steady-spend markets; channels modeled |
strong partial or with caveats weak
Validated in code
Methodology FAQ
Same family of method. A geo lift test plans holdouts in advance. MarketTwin reads the spend changes you already made, so your history gives an answer now. When a read is weak or directional, Countlift helps you design a holdout test to confirm it.
Because you can check it. Each twin is a named mix of your own markets, with weights that add up to 1 and its pre-period fit shown. The placebo test shows how much “lift” appears by chance in markets with no change.
Free location report · No card · No customer data