As foundation models converge, the real advantage in agentic ad buying shifts to how richly publishers can describe their own inventory.
Ask any vendor in the agentic landscape what makes their agent smart, and you’ll hear about models, reasoning and autonomy. That sounds great, but it’s the wrong end of the problem.
Foundation models are converging in capability and collapsing in price. Intelligence is commoditized. Within a planning cycle or two, every model provider will have access to roughly the same intelligence.
What will not converge is the information each agent has to reason with. The question then becomes: Where does that information come from? Agentic strategy has to start with data strategy, but the thin programmatic layer isn’t enough for agents to truly use. Without rich data, there’s nothing for the buyer agent to react to.
The future of agentic buying requires major changes in how the entire ecosystem thinks about data integrity. The market is discovering, project by project, that agentic buying is bottlenecked not on buyer intelligence, but on how richly the supply side can describe what it is selling.
Also Read: Why AI-Informed Marketing Needs a New Set of Rules
The Programmatic Paradigm
Part of the problem is that an agent buying media today inherits the signals programmatic has always run on: domain, URL, category taxonomy, and a bid request’s thin description of a page. These signals were designed for millisecond auctions executed by rules, not for systems capable of nuanced judgment.
Handing a capable agent shallow data is like hiring a brilliant analyst and giving them a single spreadsheet column. The reasoning is wasted because there is nothing to reason about. This is why so many agentic products in the market feel like demos: The agent is real, but its decisions are no better than the rules it replaced, because the inputs are identical.
Change is afoot, and it’s coming from the sell side. Seller agents, publisher operations platforms and SSPs are the ones building the agent-mediated direct buying vector. So what would inventory data adequate for agentic buying actually contain?
7 Dimensions of Relevant Data
- Provenance. In an agent-to-agent market, an unverifiable impression is an unpriceable one. There must be proof of where an impression came from and the path it traveled to the buyer.
- Viewability and rendering. A measured confirmation that the ad actually displayed correctly on screen.
- Ad environment. The quality and context of the page surrounding the ad: clutter, ad density, white space, and the difference between a placement and wallpaper.
- Attention and outcome. Evidence of engagement with the ad, and its ability to drive a result. This is the dimension that moves inventory data from descriptive to commercial.
- Context. Classification of the content itself, at the level of what is happening on screen, not what section of the site it lives in.
- Existence and standard brand safety. Confirmation that the impression was real, not fraudulent, and ran in a safe environment.
- Salience and authority. This is whether the content is timely in its relevance and comes from a source with genuine authority. Salience is the dimension the current stack is structurally blind to, because it changes hour by hour and lives partly outside any single publisher’s walls, in social signals about what audiences are responding to right now.
Also Read: Cookies Are Dying. Advertisers Are Fighting Over Scraps
Not Just Quality
If you’ve been around long enough, you’ve likely read plenty of proposed checklists for ensuring inventory quality. This is different, because we’re talking about upending the current flow of data in the programmatic ecosystem.
To begin with, all of these dimensions must be connected. Knowing an asset’s context is useful. Knowing its context, its provenance, its derivative versions across formats, and how each version monetized, as one linked record, is what lets an agent make an actual decision. The industry needs a reporting model keyed to the editorial object rather than the ad unit, to turn content data into commercial data.
On top of that, the dimensions must be queryable in advance, not auditable after the fact. For example, URL-level context may say “automotive,” while scene-level context says “car chase.” An advertiser who wants one does not necessarily want the other, and today’s bidstream cannot tell them apart.
With scene-level indexing across a publisher footprint, car-chase-scene inventory can be identified, forecasted and packaged before a dollar is committed. The buyer’s agent is no longer scanning the bidstream hoping to recognize quality as it flies past. It is interrogating a catalog.
This is the standard against which agentic solutions should be measured, on both sides of the transaction. A buyer agent without multidimensional data is an old optimizer with a new interface. A seller without it has nothing for a smart agent to find. The first question in agentic advertising is not “How good is the agent?” It is “What does the agent know?”
