TV incrementality

How to Geo Test TV When the Media Buy Is National

Geo test TV on a national buy with local spot heavy-ups, DMA targeted CTV and multi-cell designs beside paid social, plus spillover and synthetic control checks.

On this page

National linear TV cannot be held out by region, and measurement teams should stop designing tests that assume it can. A network or national cable schedule lands in every DMA at once, so the go-dark cell a standard geo holdout depends on is missing from the start. The design that works keeps the national buy running everywhere as the baseline, adds TV weight in chosen markets through local spot or DMA targeted connected TV, and reads what that extra weight produces against matched markets. Put paid social into the same design and the test also shows how the two channels behave together. Our explainer on what incrementality is covers the underlying question.

How do you design a geo test when TV is bought nationally?

A geo holdout test design for national TV and paid social incrementality starts with what can change by market. Paid social can: the Marketing API reference from Meta lists geo markets, regions, cities and zip codes among its targeting fields. A national schedule has no such switch, so the TV side of the test has to come from inventory sold market by market, of two kinds.

  • Local spot heavy-ups. Buy local broadcast or cable spots in a set of test DMAs on top of the national schedule. The geo experiment paper by Jon Vaver and Jim Koehler at Google models the treatment as a spend differential that is negative when campaigns are turned off and positive when spend increases, so a heavy-up fits the same analysis as a holdout. Google's newer library, Meridian GeoX, lists heavy-up beside holdback and go-dark among its designs, per its overview from Google.
  • DMA targeted CTV and addressable buys. When a streaming or addressable seller can deliver by DMA or zip code, you get a true go-dark or heavy-up cell for that slice of TV. Read the result as a proxy for linear, since audience, frequency and ad environment differ from the national spot.

Both levers measure the margin. A heavy-up estimates what the next block of TV weight returns on top of a national schedule that never stopped, the question behind any decision to add or trim TV. The average return of the national buy belongs to a marketing mix model, and a heavy-up readout is one of the better inputs for calibrating that model.

Running TV and paid social side by side

A multi-cell design lets TV and paid social cells share one control group. Google's GeoX overview describes "native multi-cell execution" against a common control, and Meta introduced multi-cell functions in GeoLift version 2.5. The multi-cell walkthrough from Meta is direct about the cost: "Our recommendation is to be very conservative with the number of cells in a single GeoLift study." For a panel of 40 cities it advises against more than 2 or 3 cells.

This hypothetical cell plan suits a brand running national cable all year beside Meta prospecting, ranked in the order we would fund it.

PriorityCellChange in the cell's DMAsQuestion it answersWhat to watch
RequiredControlNoneSupplies the markets for the synthetic controlRemove markets bordering any test cell
FirstTV heavy-upLocal spot on top of the national scheduleReturn on the next block of TV weightLocal inventory varies, so match on price as well as sales
SecondPaid social heavy-upSocial budgets raised in separate DMAsReturn on the next social dollar in the same weeksPlatform market lists may not follow DMA borders
Only with enough marketsTV plus socialBoth changes in a third set of DMAsWhether the pair beats the two cells added togetherMost markets needed, least power

Every figure in this readout is hypothetical. Over six weeks the TV cell adds $300,000 of local spot and the synthetic control shows $450,000 of incremental revenue, a marginal return of 1.5. The social cell adds $150,000 and shows $270,000, a return of 1.8. The combined cell, matched to the same share of sales, carries both changes and shows $900,000, against $720,000 if the channels simply added up. That extra $180,000 is the interaction the design exists to find, and it counts only when its own interval clears zero. In the walkthrough's own example the intervals were too wide to declare a winner at 90% confidence.

Spillover across DMA borders

In geographic experiments on national TV and paid social incrementality, spillover is the problem that hides best. A buyer who sees a test market's ads and purchases in a control market raises the control's sales and quietly pulls the measured lift toward zero. The Vaver and Koehler paper names the risk: "Ad serving inconsistency is a concern due to finite ad serving accuracy, as well as the possibility that consumers will travel across geo boundaries." Its remedy is the geo itself: "The location and size of the geos can be used to mitigate these issues." The authors add that geos as small as postal codes are not generally feasible, and they name the DMAs as one possible set of US geos.

GeoLift's guidance points to different units. Its best practices from Meta say "It is recommended to use the highest level of geographical granularity at which we can target Facebook campaigns in the study's region (Zip Codes, Cities, etc.)", and its commuting zones guide from Meta states that "Using Commuting Zones instead of cities or zipcodes in GeoLift will take the commuter effect into account for the geographical randomisation." Both are sound for a social-only test. Once a TV cell enters the design we overrule them: if your local spot is bought by DMA, the DMA is the smallest unit in which the TV treatment is clean, and every social cell has to sit on the same footprint.

That takes work on Meta. Its reference describes the geo_markets field as "Geomarkets using Comscore Markets," and we found no mention of Nielsen DMAs on that page. Building the social cell from the zip codes inside your TV DMAs keeps both channels changing in the same places, within the reference's limit of 50,000 zip codes.

Three habits keep the remaining leakage visible:

  • Drop control markets that border a test cell or share a metro area with one, and accept the lost power.
  • Assign online orders to markets by shipping address.
  • After the flight, compare the excluded border markets with the synthetic control; a rise during the test weeks points to leakage, and the main readout then reads as a floor.

Building the matched markets

Geo lift test methodology for a matched market test across TV and digital media comes down to how you build the comparison. The README from Meta opens with this line: "GeoLift is an end-to-end solution to measure Lift at a Geo-level using the latest developments in Synthetic Control Methods." A synthetic control weights untreated markets so their blend tracks the test markets before launch, and the gap that opens after launch is the lift estimate.

Google's researchers have worked both sides of the matching question. Vaver and Koehler favor random assignment because it guards against hidden differences between geos, and they report a gain from constraining it: "For example, we have found that grouping the geos by size prior to assignment can reduce the confidence interval of the ROAS measurement by 10%, or more." Synthetic Design, a NeurIPS 2021 paper in the research library of Google, picks test units and their weights together with a mixed-integer program, and its simulations on US Bureau of Labor Statistics data report better mean squared error and statistical power than simpler alternatives such as randomized trials.

TV pushes most teams toward matching, because you can only treat markets where local inventory exists at an affordable price. Let the design choose test markets from that buyable list, then hold it to the floors in GeoLift's best practices: "At minimum we recommend having 25 pre-treatment periods of 20 or more geo-units." The same page sets a flight of at least 15 days on daily data, or 4 to 6 weeks on weekly data.

Freeze everything else for the length of the test. The page says to "Keep all local marketing efforts to be constant across all markets that are taken into consideration across test and control", and to "Take into consideration any significant variations across national media such as TV, Print etc." In a TV test the national schedule has to hold its weight through the pre-period and the flight, or the synthetic control ends up fitting a moving baseline.

What to do with the readout

A heavy-up readout answers a marginal question, so put it to marginal decisions: whether the next block of local spot or social budget earns its cost. The paid social cell can also correct Meta's reported conversions through the multiplier method, and a TV cell repeated each quarter on fresh markets turns one test into a trend, on the cadence of a quarterly geo-holdout program.

Teams that get value from TV tests accept the narrower question. They test the TV they can move, run paid social in the same weeks against the same control, and leave the national average to the model. Our data analytics team designs and reads tests like this, and the Attribution Doctor is a further resource on the attribution side.

Sources

Frequently asked questions

Can you run a geo holdout on national linear TV?
Not as a go-dark test, because a national network or cable schedule reaches every market at once. The workable designs add TV where it can be bought by market: local spot heavy-ups on top of the national schedule, or DMA targeted connected TV and addressable buys that can be paused or raised in chosen markets. Both measure the return on added weight, and the average return of the national buy stays with the marketing mix model.
How many markets does a multi-cell TV and paid social test need?
More than most teams expect. GeoLift's multi-cell walkthrough recommends being very conservative with cells and, for a panel of 40 cities, advises against more than 2 or 3. Its best practices set a floor of 25 pre-treatment periods across 20 or more geo-units. A plan with a control, a TV cell, a social cell and a combined cell needs well beyond that once border markets are removed.
Should a TV geo test use DMAs or zip codes?
When local spot is bought by DMA, we use DMAs, because that is the smallest unit in which the TV treatment is clean. GeoLift's general advice favors the finest level you can target, and Google's Vaver and Koehler paper notes that geos as small as postal codes are not generally feasible. Build the paid social cells from the zip codes inside the same DMAs so both channels change in the same places.
How do you line up a Meta cell with TV DMAs?
Meta's Marketing API reference describes its geo_markets field as Comscore Markets, and we found no mention of Nielsen DMAs on that page, so market borders may not match your TV buy. Targeting the zip codes that sit inside your test DMAs keeps both channels on one footprint. The same reference lists a limit of 50,000 zip codes.

Free tools for this topic

CALCULATORROAS & Break-Even CalculatorKnow the ROAS you actually need before you scale.FREE TOOLCompetitor Ad ExplorerSee every ad your competitor is running right now.CALCULATORMedia Mix PlannerSplit any budget across channels with live projections.

Keep reading

GuidesHow to Run a Quarterly Geo-Holdout Test Your CFO Will Actually TrustRead →DataHow to Calibrate a Marketing Mix Model With Geo Lift Test ResultsRead →DataThe Multiplier Method for Calibrating Platform Attribution With Incrementality TestsRead →
CATALIST NEWSLETTER

Monthly dose of growth marketing.

Get marketing tips, narratives, guides, and playbooks delivered to your inbox.

Protected by reCAPTCHA — Google's Privacy Policy and Terms of Service apply.