Same campaign. Same quarter. Four measurement providers. Cross-platform reach came back anywhere from 7.4 million to 19.2 million households.
That spread comes from a Q3 2025 quick-service restaurant campaign in Warner Bros. Discovery's new report, The Unified Reach Paradox. The number that stopped me wasn't the spread itself. It was this: across campaigns measured by two or more providers, digital impression counts differed by about 12% on average. Digital reach differed by 92%.
The providers mostly agree on what was delivered. They disagree on who saw it. That gap lives in one place: the identity layer. And the industry is now building AI agents directly on top of it.
What WBD actually found
WBD's research team pressure-tested unified reach from three directions. They surveyed 154 agency and marketer decision-makers, reviewed the methodologies of Comscore, Innovid, iSpot, Samba and VideoAmp, and compared provider outputs across 50+ campaigns totaling 1.87 billion digital impressions.
The survey sets the stakes. 91% of buyers say different methodologies can produce significantly different reach for the same campaign. 79% say that without confidence in unified reach, they can't fully trust campaign outcomes.
The report names five methodological drivers. Here is how they rank by what they actually did to the numbers:
| Driver | What varied | Impact on reach |
|---|---|---|
| Identity resolution | One provider's match rates ran from 5% to 48% across campaigns | Leading source of variation |
| Platform coverage | No provider measured all 20+ digital platforms | Large, mostly on digital |
| Demographics and co-viewing | Adults 25–54 per household on linear ranged 0.37 to 0.61 | 65% spread on demo audiences |
| Impression qualifiers | Linear qualifiers ranged 0 to 5 seconds | Shifts reach vs. frequency |
| Universe estimates | Household universes differed by 32.5M | Smaller than expected |
The universe gap is the headline-friendly number. A 32-million-household difference in the denominator sounds decisive. It wasn't. The real divergence happened upstream, in how providers decided which exposures belong to the same person or household.
Why this doesn't surprise anyone who has run an ad server
I spent more than 15 years in adtech and video delivery, much of it around FreeWheel and Comcast. From the delivery side, an impression is the most concrete thing in the system. A creative was requested, a decision was made, a beacon fired. You can argue about viewability qualifiers, but the log line exists.
Reach is different. Reach is an inference built on top of those log lines. To count a household once, you have to decide that a CTV device ID, a mobile ad ID, an IP address and a set-top box belong together. Every provider makes that call with a different graph, different match keys and different confidence thresholds.
So the WBD numbers line up with how these systems are built:
- Impressions converge because they come from shared logs, and clean-room workflows have narrowed the gap further since WBD's 2023 analysis.
- Reach diverges because de-duplication is a modeling decision, not a counting one.
- Frequency moves inversely. Providers that match fewer exposures together report more reach and less frequency. Providers that match aggressively report the opposite. The WBD case study shows exactly that split.
None of this means anyone is wrong. It means reach is an output of an identity model, and you can't evaluate the output without seeing the model.
Meanwhile, the agents arrived
On October 6, 2026, Carat, NBCUniversal, FreeWheel and Newton Research announced custom AI buying agents for upfront inventory. Each agent is built for a specific advertiser's KPIs. The first one is live with an unnamed luxury retail brand.
The architecture is worth reading closely:
- Data: Dentsu audience intelligence plus NBCU's ID graph, unified through the FreeWheel Identity Network, combined with performance signals, search lift, trends and competitive data.
- Runtime: FreeWheel Buyer Cloud inside Dentsu's secure environment, with Newton's agentic analytics platform embedded in existing buyer workflows.
- Control: The agent surfaces recommendations. Humans keep strategy and activation.
- Placement: FreeWheel frames the key move as putting agentic AI inside the ad-serving infrastructure where media is actually activated.
That last point is the right instinct. Agents that sit next to the ad server, not three dashboards away from it, can close the loop between signal and action. It is the same reason MCP servers that connect models directly to operational systems beat copy-and-paste AI workflows.
But notice what the agent optimizes against: one identity graph.
The paradox, accelerated
Put the two pieces side by side. WBD shows that identity resolution is the biggest reason reach numbers disagree. Carat and NBCU show the industry is now handing optimization decisions to agents that run on a single identity graph.
Inside that graph, this is a real improvement. The agent sees consistent IDs, consistent exposure logs and fast feedback. It can rebalance an upfront commitment faster than any planning team.
Across graphs, it compounds the problem:
- The agent's reach is the graph's reach. If NBCU's graph matches aggressively, the agent will see lower reach and higher frequency than a third-party provider would. It will optimize toward a number other measurers may not reproduce.
- Agents make decisions, not reports. A disputed reach number in a post-campaign deck is a conversation. A disputed reach number feeding an agent's budget shift is an action already taken.
- Every network will want its own agent. WBD, Paramount, Disney and the rest have their own graphs. A brand running agents on each gets several confident, internally consistent and mutually incompatible views of the same audience.
The WBD report warns that the highest reach figure isn't automatically the most accurate. Agents remove the human pause where someone would have asked which figure to believe.
What to do about it
WBD's own recommendations point the right way: standardize the rules rather than the methodology, make transparency non-negotiable, and get direct access to impression logs. Agentic buying raises the bar on all three.
For buyers deploying agents:
- Ask which identity graph the agent optimizes against, and what its match rates look like on your campaigns.
- Keep an independent reach read on the same campaign, from a provider outside the selling network's graph.
- Track the gap between the agent's reach and the independent read over time. A stable gap is a calibration. A widening gap is a warning.
- Require agents to log the identity assumptions behind every recommendation, not just the recommendation.
For those of us building agent infrastructure:
- Treat identity as a declared input, not a hidden dependency. An agent's tool calls should say which graph and match keys produced each audience number.
- Build for multiple measurement sources from day one. An agent that can only read one provider's reach can't notice when it's wrong.
- Use open interfaces like MCP to expose measurement and identity services, so the same agent can query several graphs and reconcile them rather than trusting one.
The identity layer is the whole game
For years, unified reach disagreements were a reporting problem. Analysts reconciled decks after the campaign ran. Agentic buying turns them into an operating problem, because the number now drives the next action in real time.
The providers that win won't be the ones reporting the biggest reach. They'll be the ones whose identity assumptions are visible enough for an agent, and a human, to check. Build for that, and agents make measurement better. Skip it, and we automate the paradox.