The MarketTwin™ engine

More than a twin. Five layers behind every number.

A twin compares each market with markets your ads didn’t touch. The engine also learns your calendar, picks the method that fits your data, tests itself on your history and grades every result.

Get my free location report Book a demo or try the live demo

A plain twin can mistake a storm or a school break for ad results. The baseline learns them first.

The five layers

What runs under every read.

Three layers run on every account. Two more add outside context as they switch on. Back-tests then check the whole stack on your own history.

01
What actually happened
Layer 1: POS sales
Weekly sales per location, straight from your POS. Add door-counter counts and the engine can tell “more people came in” from “each visit spent more”.
- Square
- Toast
- CSV upload
- Door counter · optional
Included
02
What would have happened anyway
Layer 2: Machine-learned baseline
A machine-learned model learns how each market moves with season, holidays, weather, school breaks and local data. It trains only on weeks and markets your ads didn’t touch.
- Holidays
- Weather
- School breaks
- Local events
- Local market data
Included
03
The right method for your data
Layer 3: Best-fit estimator
Augmented synthetic control, synthetic difference-in-differences, or a baseline-only model when comparison markets are scarce. The one with the lowest back-test error on your data wins.
- Augmented synthetic control
- Synthetic DiD
- Baseline-only
Included
04
What changed next door
Layer 4: Local market data
Competitor openings and closings, plus foot-traffic or card-spend trends for similar businesses nearby, used as outside controls. Upload your own or connect a provider you license.
- Competitor openings
- Nearby foot traffic
- Card-spend trends
Early access
05
Steadier reads on small samples
Layer 5: Benchmarks from similar brands
Give-to-get. Anonymised results pooled by category, channel and market size. Only groups of five or more brands, no brand over 25% of a benchmark, never prices or margins.
- Category
- Channel
- Market size
Opt-inTurns on per category once enough brands contribute

Then every read is checked

Included

Back-testedHundreds of fake and planted spend changes in your own history
Calibrated 95% rangeSized from the errors those back-tests found
Grade, evidence, next stepA to D, what was accounted for, and one recommended action

On every number

A grade, the evidence and one next step.

No bare numbers. Each result says how sure it is, what it accounted for and what to do next. Values below are examples.

CountliftTampa · Paid social · Weeks 12–20Example

Extra guests

+1,240

95% range +610 to +1,880

ConfidenceLikely

MethodSynthetic diff-in-diff, the lowest back-test error for this brand

What we accounted for

Two storm weeksLate August, all three Tampa locationsBaseline
Spring school breakHillsborough County schools outBaseline
A competitor opened 1 mile from WestshoreWeek 14Local data
Similar businesses nearby down 4%Same weeks, same areaLocal data

Door countsMore visits, spend per visit flat: the ads brought more people in.

Next stepKeep Tampa spend steady for four more weeks to firm up the read.

Example evidence panel for one result: the number, its 95% range, grade B, the method, what was accounted for and one next step.

Every result gets a grade

A · ProvenStrong evidence of lift. Safe to act on.
B · LikelyLift is probable but the range is still wide. Act, and keep watching.
C · Too earlyNot enough weeks or change yet to call it. Keep going.
D · Can’t measureThis setup can’t separate ads from noise. Plan a test instead.

CountliftTrack recordExample

Our 95% ranges contained the true answer X% of the time in N back-tests.

X and N are computed from your own back-tests and shown on your account.

How we prove the ranges

We test the engine on your history first.

Before a range reaches you, the engine has been checked against answers it already knows, in your own data.

Fake spend changesWe pick hundreds of past weeks where nothing changed and pretend spend did. The true lift there is zero, so any lift found is error.
Planted liftsWe add a known lift to your real history and check the engine finds it, at the right size.
Ranges sized from the missesRanges are calibrated from those errors (conformal inference), so a 95% range is built to hold the true answer 95% of the time.
Your record, on your accountEach account shows how often its own ranges held in its back-tests. Not an average from someone else’s data.

Each bar is one back-test’s 95% range. Green holds the true answer; amber misses it. Your account counts both.

What we don’t do

We don’t track phones. We measure your actual sales.

No Bluetooth or Wi-Fi sniffing

Phones randomise the IDs they broadcast, so sniffed counts drift. Your sales don’t.

No bought location data

We don’t buy device-level location data. The FTC has acted against its sale, and several states restrict it.

No personal data

Totals by location only. No names, emails, card numbers, device IDs or cookies.

See exactly what Countlift stores →

Methods & references

Published methods, not a black box.

Each layer builds on peer-reviewed or published work your analysts can read.

Augmented synthetic controlBen-Michael, Feller & Rothstein (2021). “The Augmented Synthetic Control Method.” Journal of the American Statistical Association.Estimator
Synthetic difference-in-differencesArkhangelsky, Athey, Hirshberg, Imbens & Wager (2021). “Synthetic Difference-in-Differences.” American Economic Review.Estimator
Conformal inference for synthetic controlsChernozhukov, Wüthrich & Zhu (2021). “An Exact and Robust Conformal Inference Method for Counterfactual and Synthetic Controls.” JASA.Calibrated ranges
Machine-learned baselines (CUPAC-style)Predictions from a model trained on untouched data used as a control, building on CUPED: Deng, Xu, Kohavi & Walker (2013), WSDM.Baseline
Hierarchical Bayesian poolingWang, Jin, Sun, Chan & Koehler (2017). “A Hierarchical Bayesian Approach to Improve Media Mix Models Using Category Data.” Google.Benchmarks (opt-in)

Read the method step by step →

Questions

What analysts ask about the engine.

Is MarketTwin still a twin?

Yes, and more. A twin, the blend of your markets that didn’t get the extra ads, is still the heart of most reads. Around it, a machine-learned baseline handles the calendar, the engine picks the best-fitting method for your data, and back-tests on your own history size the ranges and set the grade.

Do you track customers’ phones?

No. We don’t track phones. We measure your actual sales. Countlift never sniffs Bluetooth or Wi-Fi signals and never buys device location data: phones randomise their IDs, and regulators restrict that data. We read totals by location from your POS, plus door-counter counts if you add them.

What does the machine-learned baseline do?

It learns how each market normally moves with the calendar, holidays, weather, school breaks and local events, using only weeks and markets your ads didn’t touch. So a hot July or a spring break isn’t counted as ad lift. It’s a ridge regression with gradient-boosted trees on top when they predict held-out history better; markets it never trained on are used to check it.

What if we only have a few markets?

When there are too few comparison markets for a twin, the engine can use a baseline-only model. The result still gets a grade, the evidence and a next step. If it can’t separate ads from noise, you’ll see a D and a suggested holdout test.

Do we have to share data for benchmarks?

No. Benchmarks are opt-in and give-to-get. They pool anonymised results by category, channel and market size, only from groups of five or more brands, with no brand over 25% of a benchmark. Prices and margins are never shared. They switch on per category once enough brands contribute.

Do we need door counters or local market data?

No. Sales by location and ad spend are enough. A door counter adds a second outcome, so you can tell more visits from bigger spend per visit. Local market data is in early access: upload your own, or connect a provider you license.

See what your ads did, graded.

Free location report · No card · No customer data