Di Gary Mittman, amministratore delegato e cofondatore di KERV.ai
When someone watches an ad-supported streaming show, decisions about which ads to show and where to place them happen in near real time for live broadcasts and just before playback for on-demand content. The strongest placements connect the ad to what the viewer just watched, enabling ad adjacency: a sports drink after a game-winning play, a furniture ad after a home renovation reveal, or an offer for a product that appeared on screen. That connection makes the ad feel relevant to the moment, a massive benefit to brands looking to capture more user attention and pull users down the funnel.
Parts one and two of our Moment Match series focused on getting one decision right: understanding the moment, then activating against it quickly enough for live and strategically enough for VOD, to deliver the right ad.
But that decision doesn’t happen once. It happens billions of times a day, across publishers, platforms, devices, live broadcasts, streaming environments, and ad formats. And it has to happen at that volume without getting slower or less precise. Part three is about what that takes.
A model that classifies a scene perfectly is worthless if it can’t classify the next million scenes at the same speed, or if the answer arrives at the decisioning layer either too slowly or too far delayed from the moment of relevance.
At real scale, being slow and being wrong produce the exact same outcome.
Analysis Isn’t the Same as Orchestration
There’s a version of contextual AI that classifies the content but doesn’t have the means to align the ads closely with the relevant context it just produced.
Classifying content is necessary, but it’s just the first step.
To influence an actual advertising decision, a signal has to do more than describe the content. It has to be usable alongside everything else that determines whether an ad can and should run: campaign goals, available creative, budget pacing, brand suitability, publisher rules, inventory availability, marketplace terms, and performance signals.
And it all has to happen before the opportunity disappears.
Without a way to feed that decisioning layer, what you’ve built is a very good warehouse. But warehouses don’t win auctions.
That’s the distinction between moment targeting and moment orchestration. Targeting identifies an opportunity. Orchestration makes that opportunity usable by the SSPs, DSPs, and ad servers that decide, while there’s still time to act. That’s what turns a moment-level signal into true contextual ad adjacency.
Scale Changes the Problem
Processing more content is part of the challenge, but volume alone isn’t what makes this difficult.
Every new episode, live sporting event, FAST channel, online video, or streaming program creates more than additional content to classify. It creates new moments that have to be made actionable against thousands of campaigns, creative assets, objectives, budgets, publisher requirements, and transaction paths.
And those opportunities don’t arrive neatly in sequence. They happen simultaneously across an increasingly fragmented, premium video ecosystem.
Each opportunity expires quickly, on its own schedule. A live touchdown has to be acted on while the play is still fresh, not queued behind last night’s procedural drama. A pause ad is a different format altogether, and its window lasts only as long as the viewer stays paused. And a signal tied to a specific pod is worth nothing once that pod has run.
At scale, moment-level intelligence comes down to volume, timing, and coordination: continuously understanding what is happening, determining which opportunities matter to which advertisers, and making those opportunities actionable before their relevance expires.
Every Environment Has Its Own Clock
Part two closed on a point worth unpacking: every platform and content type runs on its own infrastructure, and each one sets different timing rules.
Live broadcast content has no library to analyze ahead of time. The scene has to be understood as it happens, and the signal has to arrive before the break it’s meant for. VOD allows deeper analysis in advance, but the signal still has to align to the timestamp of each pod and slot. FAST channels stream continuously, with their own insertion paths. A pause ad exists only as long as the viewer stays paused.
A system built for one of these doesn’t automatically work for the others. Scale means interoperably handling all of them at once, without a separate workflow for each.
What Orchestrating at Scale Requires
For that to work, several things have to be true at once.
New content has to be analyzed continuously, including live broadcast content as it happens. Moment-level signals have to reach the decisioning layer quickly enough to influence the impression they were generated for. Those signals have to be packaged so campaign objectives, creative readiness, budget, inventory, and marketplace terms can be matched against them together, rather than through disconnected workflows.
And the results have to flow back.
Scale shouldn’t mean making the same decision billions of times without seeing what happened. Every impression creates new information about the relationship between content, creative, placement, and performance. Reporting that connects those dimensions shows which moment-level combinations work, and gives advertisers and publishers a basis for better decisions next time.
Where Agentic Comes In
Agentic AI raises the stakes for moment-level intelligence. As agents take on more of media planning, negotiation, and optimization, they need to understand what makes an opportunity valuable: what’s happening in the content, when the relevant ad opportunity occurs, and which creative can turn that relevance into action. Campaign history can inform those decisions. Moment-level signals give agents the context to identify new opportunities.
Agentic buying should connect the advertiser’s objective to the moment, the available inventory, and the creative experience, including an actionable next step for the viewer. That requires signals that buying and selling systems can interpret and act on, with outcomes flowing back to improve future decisions. The opportunity is to make moment-level relevance part of how agents plan, buy, and optimize media from the start.
Scale Without Following the Viewer
Moment-level intelligence has another advantage at scale: it doesn’t depend on identifying and following an individual user across environments.
The intelligence comes from understanding the content and the advertising opportunity around it, not from reconstructing who is watching through third-party cookies, mobile advertising IDs, or personally identifiable information.
Relevance intuitively travels with the content, without requiring advertisers to identify the viewer first. As a result, it scales wherever viewers watch premium video.
From Billions of Signals to Better Decisions
This is the problem we’ve built KERV.ai’s Moment Match Engine™ to solve, across all three steps of this series: understanding the moment, orchestrating it, and activating it at scale.
We’re not simply classifying more video or generating more signals. We’re making those signals usable across the systems where advertising is actually planned, transacted, delivered, and measured, connecting moment-level intelligence to campaign and creative requirements quickly enough to influence the ad decision.
It also means closing the loop between moments and outcomes. Every impression creates another data point. Every campaign adds to the understanding of which moments, creative, and placements perform together. And that intelligence can inform the next decision.
At billions of moments, no marketer can manage those relationships manually. The system has to orchestrate them.
The Next Competitive Line in Moment Targeting
Show-level contextual analysis is already table stakes. The differentiator is moment-level intelligence that operates at the speed and volume premium video actually requires.
In part one, we offered questions to ask a contextual partner about the depth of their signals.
At scale, there are new ones:
- Does it perform the same across live broadcast, VOD, FAST, and pause inventory, or only in one environment?
- How quickly does a signal reach the decisioning layer after the content is analyzed, and does that hold at peak volume?
- Does it plug into the SSPs, DSPs, and ad servers already in use, or require a custom workflow?
- Does it show which moments drive results, or just report how many it found?
That’s the next bar for moment targeting.
At real, quantifiable scale, moment-level intelligence has to be as fast as the opportunity itself. Anything slower is a moment already missed.
Which brings the series full circle: not all contextual signals are created equal. At scale, the gap only widens.