Brand searches rose 9% where the CTV ran.
This is the report a CTV + companion display buyer gets: two reads of behaviour from their own Google data, plus an optional survey run in the same markets.
Forty markets, twenty pairs, one coin flip each
We rank the markets by size and pair each one with its neighbour on the list, so both markets in a pair have similar search and site traffic before the flight. A coin flip in each pair picks the market that gets CTV and companion display. The other market is the holdout and gets nothing from this buy. Everything else the brand runs (search, social, national TV) carries on as normal in both, so it cancels out.
- Test unit
- Markets. In the US, DMAs. In the UK, towns and cities, each built from the postcode districts that make it up (SW1A, NR29), so Google Ads, Google Analytics and the CTV geo-targeting all describe the same place. UK TV regions don't work: there are 14, and the test needs 40 markets.
- Media
- CTV for reach in the test markets. Companion display reaches the same households on their other screens, inside the same markets.
- Read 1
- Brand search and non-paid site visits by market, from the brand's own Google Ads and Google Analytics. Google Ads reports UK search down to postcode district; Google Analytics reports visits by town or city, which is why UK markets are whole towns. No pixels and no device matching.
- Read 2
- Optional. A panel partner surveys adults in the same test and holdout markets after the flight.
- Why markets
- Nobody's device or ID gets matched or handed over. The comparison is market against market, which works under GDPR and US state privacy laws without changing anything.
Brand search: CTV markets vs holdout markets
Weekly searches for the brand's name in each group of markets. The average of the eight weeks before the flight is set to 100. The flight is shaded red and the two weeks after it are shaded grey.
Fake launches before the flight, no ads running
- Fake launch 5 weeks before the flight: +0.2% (−3.4% to +4.0%)
- Fake launch 4 weeks before the flight: −0.5% (−4.2% to +3.4%)
- Fake launch 3 weeks before the flight: −0.3% (−4.8% to +4.5%)
- All close to zero, so the CTV markets weren't already pulling ahead before the flight.
20,000 random coin flips
Swapping test and holdout inside pairs at random shows what pure chance produces. 27 of 20,000 random flips came out as far from zero as the real assignment (p 0.001). Site visits: p 0.006.
Non-paid site visits: the visits no ad paid for
Non-paid visits are direct, organic search, organic social and referral: everything in Google Analytics that didn't come in through a paid click. Paid visits are left out on purpose, and the chart below shows why.
Fake launches before the flight, no ads running
- Fake launch 5 weeks before the flight: −1.0% (−5.6% to +3.8%)
- Fake launch 4 weeks before the flight: −1.2% (−5.3% to +3.1%)
- Fake launch 3 weeks before the flight: −1.4% (−5.4% to +2.8%)
- All close to zero, so the CTV markets weren't already pulling ahead before the flight.
Why paid visits stay out
Visits from paid display clicks rose +50.6% in CTV markets. That's the companion display doing its job, and the brand already sees those clicks in its ad reports. They were bought, so counting them would credit the ads with visits they paid for. Non-paid visits are the ones the campaign earned.
More CTV weight, more lift
Not every test market gets the same weight. Splitting the pairs by how much CTV the test market carried is a check that the lift comes from the ads. If it does, heavier markets should move more. Each half is only ten pairs, so the ranges are wide.
Heavier-weight pairs moved more on both reads: search +10.5% vs +7.5%, non-paid visits +9.7% vs +0.6%. A lift that has nothing to do with the ads wouldn't line up with the weight like this.
Brand lift survey in the same markets
After the flight a panel partner asks adults in the test and holdout markets the standard brand questions. The holdout markets give the baseline, including the people who say they remember an ad they could never have seen. Everyone in a market counts, whether the ads reached them or not, so these lifts are smaller than an exposed-vs-unexposed survey would show. In exchange, no device has to be matched to anyone.
1,500 completes in each group, 75 per market. Ad recall, consideration and purchase intent came back clearly above zero. Awareness read +2.2 pts with a range of −2.0 to +6.4 points, which can't rule out zero. The true lift we set was 4 points, so this is a miss on sample size, not on the ads: see Sizing.
Every range contains the true lift
Because this world is simulated, we know the right answers. A good method lands near them, and its 95% range contains them about 19 times in 20.
| Read | True lift we set | What the study read | 95% range | Contains the truth |
|---|---|---|---|---|
| Brand search | +10.0% | +9.0% | +3.9% to +14.3% | Yes |
| Non-paid site visits | +8.0% | +5.1% | +1.4% to +8.8% | Yes |
| Recall seeing an ad for the brand | +7.0 pts | +6.4 pts | +3.1 to +9.7 | Yes |
| Aided brand awareness | +4.0 pts | +2.2 pts | −2.0 to +6.4 | Yes Can’t rule out zero |
| Would consider the brand | +2.5 pts | +3.7 pts | +0.8 to +6.7 | Yes |
| Likely to buy in the next 3 months | +1.0 pts | +2.5 pts | +0.5 to +4.5 | Yes |
6 of 6 ranges contain the true lift. Some reads land under the truth (non-paid visits +5.1% against +8.0% set) and some over (consideration +3.7 pts against +2.5 pts). That spread is the noise in a 40-market test, and it's why every read comes with its range.
How big the test needs to be
We re-ran this study 300 times at each true lift, with fresh noise every time, and counted how often the read came back clearly above zero. A test should find the lift you hope for at least 80% of the time. Below that, a real effect can easily read as nothing.
Search and site visits, 40 markets, 6 weeks
Survey awareness, by sample size
| Markets in the test | Smallest search lift found 4 times in 5 | Smallest site-visit lift found 4 times in 5 |
|---|---|---|
| 20 markets (10 pairs) | +11% | +12% |
| 40 markets (20 pairs) this study | +8% | +7% |
| 60 markets (30 pairs) | +6% | +6% |
At the volumes in this sample. Before a real test we size it on the brand's own search and visit history, market by market, and tell them what lift the test can see before they commit.
From kickoff to the read
- Read-only access. Google Ads and Google Analytics, with at least 12 weeks of history, ideally a year. We read search and visits by market; we never change the account.
- Pick and pair the markets. We size each market from that history, drop the ones too small to read, pair the rest and flip the coins. In the UK we also check what share of the brand's search impressions Google places in a postcode district, since some only resolve to a region. The brand sees the list before anything runs.
- Keep the holdout clean. No CTV or display from this buy in holdout markets. Other media can run as normal, as long as it runs the same way in both groups. Tell us about any regional pushes.
- Run the flight. Six weeks of CTV and companion display in the test markets, with a weekly check that delivery stayed inside them.
- Field the survey (optional). The panel partner surveys both groups in the last week of the flight or just after it.
- Read it two weeks after the flight. Search, site visits and the survey, with fake launches, random reassignments and the weight check, in a report like this one.
What this design can't do
- Anything that happens in only some markets during the flight (a local news story, a competitor's regional push) can look like lift. Pairing many markets dilutes it, and the fake launches catch most of it.
- CTV geo-targeting isn't perfect at the edges, and people travel. Both leak a little exposure into holdout markets, which makes the read smaller than the truth, never bigger.
- A small brand with little search or site traffic per market can't be read this way at a sensible budget. We say so at the sizing step rather than after the flight.