How AI Powers Connected TV Advertising: Targeting, Creative, and Incrementality Benchmarks for 2026

The budget question around ai digital advertising in connected TV is no longer whether streaming can be measured at all. It can. The harder question is whether it should take money from search, social, or PMax when those channels already have operating rhythms a growth team understands.

That question is suddenly uncomfortable because CTV has produced one of the strongest performance signals in digital advertising. In Stella’s incrementality benchmarks, Tatari CTV delivered a 3.30x median incremental ROAS across 225 controlled geo experiments run from August 2024 through December 2025, ahead of Google PMax at 2.98x, Meta at 2.92x, and TikTok at 0.94x.[1] That is not a vague engagement claim. It is a controlled incrementality result.

It is also not a blank check. The same benchmark shows CTV with the widest outcome spread, with a coefficient of variation of 0.85.[1] For a media buyer, that caveat is not legal copy at the bottom of the slide. It changes how the channel should be planned, tested, and defended after the campaign runs.

CTV is big enough to belong in the performance budget conversation

The scale argument is no longer the interesting part, but it matters enough to clear quickly. US CTV ad spend is projected at $37.95 billion in 2026, programmatic biddable inventory is estimated at 47% of the CTV market after 40% year-over-year growth, and streaming eclipsed cable plus broadcast for the first time in May 2025, according to figures cited in Equativ’s 2026 CTV guide.[2]

Buyer behavior is moving with that scale. Equativ cites 68% of advertisers now considering CTV a must-buy and 58% expecting interactive capabilities as standard over linear TV.[2] Roku Advertising also projects that up to 50% of streaming advertisers in 2026 will fund CTV spend increases by pulling from search and social budgets.[2] That last number should be treated as directional, not as proof that budget migration has already happened. It is a forecast from a company with a clear stake in the outcome.

Still, the direction is hard to ignore. CTV is no longer sitting outside the performance plan as a soft awareness line item. It is being compared against the same dollars that would otherwise go to lower-funnel channels. That comparison is where the channel becomes both promising and dangerous.

The incrementality benchmark is strong. The variance is the story.

A 3.30x median iROAS deserves attention because median performance is harder to dismiss than a cherry-picked case study. Stella’s benchmark is based on 225 controlled geo tests, and the channel ranking puts Tatari CTV above Google PMax and Meta on median incremental return.[1] If CTV were still only a reach vehicle with nicer reporting, that result would be unlikely.

But median iROAS answers only one question: what happened in the middle of this measured set. It does not tell a marketer how likely their own campaign is to land near that middle. The coefficient of variation matters because it describes dispersion around the average. CTV’s CV of 0.85 was the highest among the measured channels in the benchmark.[1]

That means the same channel can look brilliant for one advertiser and expensive for another. A top-performing campaign can make CTV look like the cleanest new growth lever in the plan. A weak one can make it look like the team bought television with digital vocabulary. The planning implication is straightforward: CTV should not be funded as though the median is guaranteed.

The study’s own sample also narrows what can be concluded. Stella describes a set of measurement-sophisticated DTC advertisers using a single platform, not a random sample of every advertiser trying streaming for the first time.[1] The benchmark also recommends discounting results by 15–20% for conservative planning.[1] That does not erase the signal. It prevents the number from becoming a sales guarantee.

This is where CTV differs from channels with more mature performance habits. Search has obvious intent capture. Meta has years of creative testing muscle memory and platform-native feedback loops. PMax has its own opacity problems, but many teams already know how it behaves under marginal budget changes. CTV can outperform them, but it asks for more discipline before the test starts.

What AI actually changes inside the CTV stack

AI does not make CTV a performance channel by magic. It works only if it improves at least one of four things: audience confidence, creative relevance, delivery efficiency, or measurement confidence. Anything else is packaging.

Audience targeting: logged-in graphs, IP signals, and clean rooms

CTV targeting leans more heavily on deterministic signals than many open-web display buys: logged-in user graphs, IP addresses, and data clean rooms are central pieces of the targeting infrastructure described in Equativ’s guide.[2] That matters because household-level streaming environments often do not behave like click-based channels. The advertiser is trying to decide whether the household is likely to contain the right buyer, not whether one person just typed a query.

AI sits on top of those signals by scoring audience fit, resolving identity where permitted, building lookalike or propensity segments, and suppressing households that are unlikely to move. The practical question is not whether the model sounds advanced. It is whether the advertiser has enough audience density for the model to find reachable households without overpaying for thin segments.

This is also where governance belongs in the discussion. AI-targeted CTV can be powerful precisely because household, content, and identity signals are becoming more usable. Teams that need a deeper risk lens should treat targeting governance as part of the channel evaluation, not as a compliance appendix; the governance issues are broader than CTV and are covered in The AI-Targeted Advertising Trap.

Contextual intelligence: matching the spot to the viewing moment

Contextual AI in CTV analyzes on-screen metadata and surrounding content signals to place ads in environments that better match sentiment, category, or viewing context.[2] This is not the same as keyword matching on a webpage. The viewer is leaning back, the screen is shared, and the ad often has no clickable path. Creative-context fit has to do more work.

For performance teams, contextual intelligence is useful when it reduces wasted exposure or improves comprehension. If it merely gives a prettier explanation for where an ad ran, it belongs in the reporting appendix.

Creative optimization: attention is useful, but it is not profit

The creative data is encouraging, especially because it moves closer to observable behavior than generic engagement. Equativ and TVision’s commissioned research found that AI-powered interactive CTV ad formats drove a 14% increase in viewer attention and a 52% total lift in attention scores compared with standard spots.[2] The same source reports that shoppable overlays produced a 53% increase in comprehension of advertised products, and that 76% of consumers would scan a QR code in a TV ad if the offer were relevant.[2]

Those are useful signals, not final outcomes. Attention can make a campaign more likely to work. Product comprehension can reduce friction. QR willingness can reveal a path from the television to a measurable action. None of those automatically becomes incremental sales.

The operational lesson is still important: CTV creative cannot be treated as a repurposed social cutdown or a brand film with a QR code pasted on at the end. AI can help vary offers, sequence messages, adapt overlays, and learn which calls to action produce more downstream response. But a model needs creative options to choose from. If the team brings one generic thirty-second spot and no landing-page logic, AI has little room to improve the economics.

Programmatic buying and supply-path optimization

The buying layer is where AI can quietly protect performance. Programmatic CTV has inventory fragmentation, platform differences, duplicate supply paths, and quality variation. Equativ describes agentic AI being deployed for autonomous media planning and supply-path optimization.[2] If that reduces waste, improves reach against the intended audience, or avoids inefficient supply routes, it has a real budget argument.

The buyer should still ask what the system is optimizing toward. Completion rate, attention, reach, household frequency, CPA proxy, and incremental lift are different objectives. A platform can be very efficient at buying the wrong outcome.

Measurement: CTV needs incrementality more than last-click

CTV breaks the habits that make many performance dashboards feel comforting. Viewers usually do not click the television. A household may see an ad on a streaming app, search later on a phone, visit through direct traffic, or convert through a branded paid search ad. If the team relies only on click-through attribution, CTV will either look weaker than it is or get credit through indirect paths that are hard to trust.

That is why geo testing, holdouts, modeled incrementality, and cross-platform measurement matter. Equativ cites 72% of buyers prioritizing cross-platform measurement, and B2B measurement has a related dark-funnel problem, with 38% of pipeline arriving without attributable touchpoints.[2] For CTV, those problems are not edge cases. They are the measurement environment.

Teams without the measurement infrastructure to run and interpret these tests should fix that before they scale spend. For broader stack design, How to Build an AI Marketing Analytics Stack is the more relevant starting point than another vendor demo. The same caution applies to ROI claims generally; the measurement gap around AI systems is addressed in The AI Analytics ROI Gap.

Where CTV is most likely to earn budget

The campaigns most suited to AI-powered CTV share a few traits. They have a reachable audience that can be identified with enough confidence. They can support video creative beyond one generic asset. They have enough conversion volume or business signal to read incrementality. They are not expecting the channel to behave like paid search.

• Audience density: the target segment must be large enough for household-level targeting without creating a tiny, expensive pool.
• Creative readiness: the team should have multiple messages, offers, or overlays to test, not one broad brand spot.
• Measurement setup: the advertiser should be able to run geo tests, holdouts, or another incrementality method before making a scale decision.
• Budget patience: the test needs enough spend and time to detect lift without being judged like a same-day click campaign.
• Channel interaction awareness: the team should expect CTV to influence search, direct, retail, and social paths rather than live neatly inside one attribution column.

That last point is the one most likely to cause internal friction. If CTV works, it may show up partly as more branded search, stronger retargeting pools, higher direct traffic, or improved conversion efficiency elsewhere. That does not make it unmeasurable. It means the test design has to look for total incremental effect, not only channel-native actions.

The B2B claims are interesting, but should be handled carefully

B2B CTV is becoming part of the conversation because account-based targeting, logged-in viewing environments, and intent data make streaming feel less like a pure consumer reach channel. Specificity Inc. claims that 73% of B2B organizations had integrated CTV into performance marketing stacks by mid-2026, using account-based targeting and human-verified intent data.[3] The same source cites premium B2B CTV CPMs ranging from $25 to $65.[3]

Those numbers are worth noting, not overbuilding around. They come from a single agency source, not an independently verified market census. For a B2B growth team, the more useful takeaway is qualitative: CTV can participate in account-based programs when the buying committee is reachable and the measurement plan accounts for long, partially unattributed journeys.

A B2B advertiser evaluating CTV should ask whether the target-account universe is large enough, whether household exposure is a credible proxy for business influence, and whether sales-cycle movement can be tested against a holdout. If the answer is no, the channel may still have brand value, but it should not be sold internally as a clean performance engine.

A practical go/no-go framework for 2026 CTV investment

The right decision is not “move budget into CTV” or “stay in search and social.” It is whether the next marginal dollar has a better testable return in CTV than in the channels already absorbing spend. The Stella benchmark makes CTV too strong to ignore. The variance makes it too risky to fund casually.

A sensible first test does not need to prove that CTV is the new center of the media plan. It needs to answer whether the channel can produce incremental lift under the advertiser’s own audience, creative, offer, and measurement conditions. If it cannot, the team has learned something before overcommitting. If it can, the next budget conversation becomes much easier.

Advertisers comparing vendors should also separate platform capability from operating model. A strong DSP, clean-room integration, or creative optimization product does not guarantee the team using it will design a clean test. For broader vendor categorization, see AI Advertising Companies: A Practical Category Guide for Paid Media Teams. If an agency will run the program, How to Evaluate an AI Advertising Agency is the more useful filter.

AI-powered CTV has earned a place in the 2026 performance testing plan for advertisers with the right conditions. It has not earned the right to be treated as universally scalable performance media. The channel’s promise is real precisely because the incrementality signal is real; the discipline is necessary because the variance is real too.

References

1. 2025 DTC Digital Advertising Incrementality Benchmarks, Stella.
2. CTV Advertising 2026 Guide, Equativ.
3. B2B CTV Advertising Case Studies: Driving High-Intent Performance in 2026, Specificity Inc.