Ask yourself where conversion data goes after a campaign runs. It flows back to the advertiser, the measurement vendor and the DSP. It does not flow back to the SSP.
That single fact explains the last 15 years of programmatic power dynamics better than any discussion of algorithms or auction mechanics. The DSP owns outcome prediction because it owns the feedback loop. The SSP supplies inventory and context, executes the transaction and learns nothing about whether it worked.
This was a tolerable arrangement when programmatic was a volume game. It is an increasingly bad one as the market moves toward outcome-based buying. And agentic trading is the first structural development in programmatic history that has any chance of changing it – not because protocols will open the architecture (they will), but because genuine buy-side and sell-side collaboration now has a mechanism.
For the first time, neither side has to solve the outcome-prediction problem alone.
Why neither side can do this alone
The DSP’s intelligence advantage is real, but it’s incomplete in a way that’s rarely acknowledged. DSPs have deep outcome data – conversion signals, attribution, measurement feedback – but almost no visibility into what was actually true on the supply side at the moment before an ad ran. They bid into a context they can’t fully see and optimize against outcomes that arrive later and imprecisely.
SSPs have the opposite problem. They’re sitting on contextual and behavioral signals that are genuinely predictive – session depth, content adjacency, on-device interaction sequences, SDK-level behavioral data – but they’ve had no way to know which of those signals actually correlate with outcomes downstream. They’ve been generating inputs into a model they’re never allowed to read.
The result is that today’s outcome-based buying is really just DSP-based buying. A buyer asks a DSP to solve the problem with whatever it can see, and the supply side is present at the transaction but absent from the intelligence.
What a shared brain actually looks like
Now let’s make this real with a concrete example.
An SSP with SDK-level data knows at the moment of an auction that a user just completed a purchase inside a shopping app, is mid-session on a finance content site and has been engaging with product comparison content for the past four minutes. That signal cluster – real purchase intent, in-session, actively researching – is the kind of pre-impression context that genuinely predicts conversion.
But under the current model, these signals either don’t make it into the bidstream or arrive as a thin taxonomic label that a DSP’s model has no reason to weigh heavily, because it’s never been tied to outcomes the DSP can verify.
In an agentic model, those signals travel differently. A sell-side agent can surface that context with specificity: Here’s what we know about this impression, here’s the data we’re drawing on and here’s our estimate of purchase intent. A buy-side agent can evaluate that against its outcome history – “We’ve seen 40% higher conversion rates when this type of contextual cluster is present,” for example – and price accordingly.
Neither agent could reach that answer alone. The SSP doesn’t know outcomes. The DSP doesn’t know the supply-side context at this resolution. Collaboration closes both gaps simultaneously, and the result is something that doesn’t exist today: A bid price that reflects predicted value from both sides of the transaction.
The data that makes it real and the vendors that matter
Not every SSP can participate in this. The enrichment layer is the differentiator.
Publisher-passed content signals are table stakes. The higher-value inputs are the ones built from direct SDK integrations, including interaction sequences, session behavior, device context and return-visit patterns. Survey-validated intent data – while smaller in scale – adds a layer of stated purchase intent that neither behavioral modeling nor contextual inference can replicate.
The SSPs and data enrichment vendors that have built proprietary access to these signals through publisher partnerships and on-device models will be the ones whose sell-side agents actually have something useful to contribute to outcome prediction.
The ones that haven’t will still be pipes. They’ll execute what a buy-side agent decides at prices the buy-side agent sets without any ability to participate in the intelligence that determines those prices.
The price discovery problem that agentic collaboration solves for advertisers
The downstream consequence for advertisers of moving value prediction beyond the DSP is significant and underappreciated.
When outcome prediction is entirely owned by the DSP, floor prices and inventory value are set in a vacuum. An SSP can’t justify a higher price for a genuinely high-intent impression because it can’t prove intent in terms the buy side is able to trust. Meanwhile, advertisers can’t buy on outcomes with real confidence because the only outcome model they have access to is the DSP’s, which is trained without supply-side context.
Agentic collaboration changes the price discovery mechanism. If a sell-side agent can contribute a credible, data-backed intent estimate – and a buy-side agent can verify it against its own outcome history – then high-value impressions can be priced as such. From there, buyers can commit to outcome-based deals with genuine confidence and the market stops leaving arbitrage on the table that currently flows to whoever is least transparent about what they know.
Who captures the opportunity depends on timing
The buy-side agents now being deployed are forming priors about which sell-side signals are useful. They’re learning from early agentic campaigns to identify which supply contexts correlate with outcomes and which don’t.
SSPs that show up to those early conversations with enriched, outcome-correlated signals will shape what the market considers valuable for the next several years. Those that wait will be told what their inventory is worth by a model they had no hand in training.
DSPs have held a monopoly on programmatic intelligence not because they built the better algorithms, but because they sat at the end of the feedback loop. Agentic trading moves the feedback loop. The question for the supply side is whether it moves with that shift or watches the new intelligence layer get built entirely by the buy side yet again.
For more articles featuring David Simon, click here.
