Getting CTV attribution right earns the channel fair credit but credit is not proof. While a properly tuned MMP can tell you which installs CTV touched, it still cannot tell you which installs CTV caused. Those are different questions, and only the second one justifies the spend.
That second question is incrementality. Of the installs now credited to CTV, how many would not have happened anyway? Strip out the users who would have installed regardless, and what is left is the lift your ads actually drove.
We have run three CTV incrementality tests to answer it. One proved very little. One had the right method and the wrong inputs. One produced statistically significant lift, twice, in a single clean test. Here is what measuring CTV incrementality actually involves, the methods that work, and what those three tests taught us about getting a result you can trust.
Why is CTV the hardest channel to measure
CTV is one of the fastest-growing user acquisition channels and the hardest to measure honestly. Three things get in the way:
- No click to follow: Nobody taps a television, so there is no click tying an impression to an install.
- Cross-device and delayed: The ad plays on the TV, and the install happens later, on a phone. Your measurement has to bridge both devices and the gap in time.
- Last-touch hides it: Models that reward the final touch quietly strip CTV of credit it earned, unless attribution waterfall is carefully configured.
Before you can judge CTV's performance, you have to fix how it is measured. That starts with attribution and ends with incrementality.
Clear the attribution conflict first
The worry that CTV cannibalises your other channels is usually confusion with attribution settings. It is an attribution priority conflict. CTV view-through sits last in the MMP waterfall, below clicks and deterministic views, so it only claims an install when no higher tier can. It never competes on equal terms with your mobile channels unless you make it so with the equal attribution priority window. During this window of time, a CTV impression will have the same priority as a mobile click in the attribution waterfall.
We recommend a 6-hour equal attribution priority window plus a 24-hour view-through window, so real CTV installs are credited rather than pushed into mobile channels(Adikteev internal data, 2025). With attribution fixed, you can finally measure whether CTV is driving installs without the fear of cannibalising organic.
What incrementality actually measures
The method is a controlled experiment. Split a matched audience in two. One group sees your CTV ad. The other sees a placebo, usually a public service announcement (PSA). The two groups are otherwise identical, so any difference in their install rate is down to the ads.
The gap is the lift. It is the set of installs that would not have happened without your ads.
Incremental lift = (Exposed IPM minus Control IPM) / Exposed IPM, where IPM is installs per 1,000 impressions.
Two numbers then tell very different stories:
- Platform CPI tells you what you paid.
- Incremental CPI tells you what you actually gained, measured only against the installs that would not have happened without your ads.
How to measure it cleanly
CTV is a UA channel, and UA cannot split an audience the way retargeting can. So the clean method for CTV is a PSA holdout. Run your CTV ad to the test group and a public service announcement to the control group, then compare CPI and ROAS across the two campaigns. It is quick to set up and uses familiar campaign metrics through your MMP. The trade-off is the cost you incur on the control group, which sees a PSA rather than your ad. This is our core methodology.
Two other routes exist, each with an honest trade-off.
- Retargeting measures it differently: With RTG you show the test group your ad and the control group nothing, then compare the reactivation rate and ARPU. This is the basis of ITT 2.0. It is cost-efficient, transparent, aligned with business outcomes and works across platforms, but it needs a minimum audience size and a control group held out across channels.
- Third-party or in-house modelling: A specialist firm or your own BI team measures lift with a synthetic control group, time-series analysis or media mix modelling. It is independent and can use more advanced methods, but it is slower, more complex and carries tool or team costs.
The method follows the KPI, the audience size and how much independence you need.
Three tests with three lessons
This is the practical part. We ran three tests. Each one taught us something the next one used.
Test one taught us more than it proved
We ran a PSA holdout on a puzzle game and measured IPM-based lift. The result was plus 2 per cent. Technically positive, but practically meaningless. It was not a failure, though, because it showed us exactly what to fix.
Three things went wrong:
- The groups drifted apart mid-test: A high-risk IP blacklist, meaning an IP showing too many different device IDs, was applied partway through, so the exposed and control groups were no longer comparable.
- We had only four clean days of data: Far too short a window to read an incrementality signal with any confidence.
- We measured a narrower set of installs than the client did: We tracked attributed installs. The client also counted assisted conversions we could not see, so our picture was incomplete.
The four fixes that became our playbook
- Balance from day one: Apply the high-risk IP blacklist before launch, never mid-flight.
- Smooth the spend: Cap over-aggressive bidding so the budget spreads across as many CTV devices as possible.
- Match the frequency: We had run at roughly one impression per IP per day against a benchmark of two. Match the benchmark so the comparison is fair.
- Refine the targeting: Move budget towards higher-quality publishers and segments rather than buying across the whole of CTV supply.
Test two had the right method and the wrong inputs
For a social app, we tried something completely different. Instead of splitting users, we split the map, using a geo- and time-based design, difference-in-differences, with daily new users as the KPI.
Run CTV in a set of test regions, hold out a set of matched control regions, then compare how each group changes over two phases.
- Phase one-scale: Grow spend in both region groups together.
- Phase two, pause: Switch off the control regions and watch what changes.
Difference-in-differences then strips out seasonality. If the control regions drop by 200k and the test regions drop by 50k, the extra 150k is what CTV was really driving. The test's drop of 50k was just the season.
This was a strong method, but the trouble was the inputs.
- The KPI was a black box. The daily new users formula was never shared with us, so we could not optimise against it.
- The channel was not isolated. Several CTV vendors ran on the same test at the same time, so refining our own targeting could not move the overall number much.
The signal pointed to low uplift across all CTV channels, but we could not bank it. The lesson was simple. Agree on the metric and isolate the channel before you have spent anything.
Test three is the one that worked
For a social casino app, we applied everything we had learned. A PSA holdout, clean from day one, built from audience-builder segments and sliced by device count. We ran two campaigns on purpose.
- Campaign A, tight and high-value: Around one million users in a premium segment, filtered to IPs with only a few unique device IDs. The cleanest audience we could build.
- Campaign B, broad and modelled: Three to five million users generated by our CTV audience model. The test of whether incrementality survives scale.
Both were clean by design, with the blacklist applied from day one, balanced groups, a full clean run and frequency at benchmark.
Both produced statistically significant lift.
- Campaign A reached 0.09 IPM against a 0.06 control, on roughly 10,000 dollars of spend.
- Campaign B reached 0.10 IPM against a near-zero control, on roughly 11,600 dollars of spend.
The catch is the trade-off. Applying these incrementality-focused targeting strategies multiplied CPI by three.
What actually drove the lift
- Audience tightness: Narrow, low-device-count segments carried the incrementality. Broad reach diluted it.
- A high-intent vertical: Lower organic pull meant CTV-driven installs were genuinely additional, not installs that would have come anyway.
- Inventory quality: Authorised resellers and premium impressions converted into real, attributable lift.
Underneath all three sits one tension. Incrementality, CPI and reach pull against each other. Push for more incrementality, and you usually give back reach and pay a higher CPI. The skill is choosing where on that triangle each campaign should sit.
The lever matters more than the vertical
Your vertical does not determine whether CTV is incremental for you. It settles which lever unlocks it.
- Win with precision: For apps with heavy organic pull, like casual and puzzle games or top-chart titles, broad CTV gets swamped by organic. You win with tight high-value segments, low device counts, controlled frequency and premium inventory. Quality over scale.
- Scale is on the table: For leaner-organic, high-intent verticals, like casino and similar performance apps, broad modelled audiences can be genuinely incremental, so you can scale even before you tighten targeting.
The question is never whether CTV can be incremental for your app. It is which lever unlocks it.
How to run a test you can bank on
A result is only worth acting on if you can trust it. Five things make the difference.
- Clean holdouts from day one: Balance exposure and control before launch, and never change the rules mid-flight.
- Agree the KPI up front: Pin down the exact metric and its formula before any money is spent—no hidden KPI.
- Enough clean days and sample: Give the signal room to emerge. Fourteen days is our recommended minimum.
- Control inventory and frequency: Match quality and impression frequency across groups to ensure a fair comparison.
- Measure attributed and assisted installs: Count every install your ads touched, not only the last-view ones.
On every test, we disclose the holdout design, the sample size, the test duration and the significance. Ask any partner for the same.
Where to start
You do not need to solve all of this at once. Here are the three steps to begin with:
- Define: Pick the KPI and the holdout together, in writing.
- Design: Build a clean, balanced test from day one with the right audience.
- Scale: Measure incremental CPI, then grow the segments that prove out.
The bottom line
- CTV is incremental, under the right conditions: We have proven a statistically significant lift to show for it.
- CTV and mobile compete for the same installs: That is why MMP settings like an equal attribution priority window matter. They stop CTV from being disadvantaged before the test even starts.
- Scaling is the frontier: Incrementality, CPI and reach pull against each other, and the real challenge is finding and growing high-incrementality segments without giving back too much on CPI.
Attribution earns CTV fair credit. Incrementality proves the credit was deserved. Get both right, and you stop guessing what CTV is worth and start knowing.