Key Takeaways:
- Ask what the machine is optimizing for, not whether to test it. ML-driven DSPs chase whatever signal you give them, which may not match your actual business goal.
- CTV is not a simple extension of mobile programmatic. DSPs built for app installs bring the wrong measurement models to brand recall and delayed retail purchases.
- Evaluate data provenance and default optimization targets first. First-party signal beats modeled audiences, and every platform’s default should match your real KPI.
- Demand incrementality methodology, not just reach or ROAS. If a DSP leads with impressions instead of proof of incremental impact, that’s a red flag.
- Scrutinize agency partner programs for misaligned incentives. Preferred pricing or rebates can shape recommendations independent of actual performance.
When a machine learning digital signal processing (DSP) moves into CTV and retail media, the right question is not “should we test it?” The right question is: “Do we understand what the machine is actually optimizing for, and does that match our business goal?” Most brands never ask that second question. That gap is where media dollars disappear.
The DSP landscape is shifting fast. Platforms built on mobile and app-install performance are pushing into connected TV, retail media networks, and beyond. It makes sense for them. It does not automatically make sense for you.
What “Machine Learning-First” Actually Means for Your Campaigns
“Machine learning-first” is becoming a badge every ad tech vendor pins to its chest. But the term means something specific, and brands deserve a plain-language definition.
A machine learning DSP uses algorithmic models to make real-time bidding and targeting decisions without manual rules. Instead of a human saying “bid $4.50 on this audience segment,” the system processes signals continuously and sets its own bids. The promise is efficiency at scale. The reality is that the machine optimizes relentlessly toward whatever signal you give it.
Give it click-through rate, it will chase clicks. Give it purchase events, it will chase purchases. Give it reach and frequency targets, it will chase reach and frequency. The machine does not know your brand. It knows your objective function.
This matters enormously for CPG and DTC brands because your actual goal is rarely captured cleanly in a single signal. You want new buyers, not just any buyers. You want incremental sales, not just sales that would have happened anyway. You want brand equity alongside conversion. None of that fits neatly into a single optimization target.
Why Expanding Into CTV Creates Real Risk for Brands That Aren’t Ready
CTV is not a straightforward extension of mobile programmatic. The inventory quality variance is significant, measurement standards are still maturing, and the creative requirements are completely different.
A DSP that has spent years optimizing app installs on mobile has built models around short attention spans, thumb-stopping creative, and last-touch attribution. CTV asks different questions:
- Did this drive brand recall?
- Did it introduce the product to a genuinely new household?
- Did it contribute to a purchase that happened four days later at retail?
We have written before about which CTV metrics actually matter for performance, and the short answer is that reach alone tells you almost nothing. The same logic applies to evaluating a new DSP entering the space. Ask what their measurement methodology is before you ask about their CPMs.
For brands already managing retail media alongside CTV, the evaluation gets even more complex. We unpacked that in our piece on retail media strategy beyond sponsored listings. The TL;DR: channel expansion without measurement alignment creates chaos, not efficiency.
How to Evaluate an Emerging DSP Without Getting Sold a Story
Here is the framework we use at Junction 37 when a client asks us whether to test a new or expanding DSP.
1. Ask about data provenance.
Where does their audience data come from? Is it first-party signal from a closed ecosystem, or is it modeled from third-party sources? First-party signal is more durable. Modeled audiences are often just expensive guesses.
2. Ask what they optimize toward by default.
Every platform has a default. Make sure it matches your actual KPI, not a proxy metric that sounds good in a case study.
3. Ask for incrementality methodology.
Any DSP worth testing should have a clear answer on how to measure whether their placements drove outcomes that would not have happened otherwise. If they lead with reach or impressions, that is your answer.
4. Ask who manages the campaign.
Some DSPs selling through agency programs are structured so a junior account team from the platform side is making optimization decisions. You want your agency, not the vendor, holding the strategic controls.
5. Ask for category-relevant case studies.
A CPG brand selling at grocery retail has different success criteria than an app publisher. Ask for proof that is relevant to your situation, not the flashiest name they can drop.
The Agency Partner Program Model Deserves Scrutiny Too
When ad tech platforms launch agency partner programs, they are often building distribution networks as much as anything else. That’s fine. But it means the incentives are worth examining.
An agency that gets preferred pricing, volume rebates, or co-marketing benefits from a DSP platform has a reason to recommend that platform beyond pure performance. At Junction 37, we have one incentive: your results. That is a structural difference, not a marketing line. Read more about how programmatic partnerships affect your media ROI.
We are not anti-innovation. We test new platforms regularly on behalf of CPG and DTC clients. But we test with a control, a clear measurement plan, and a genuine willingness to walk away from something that does not work.
FAQ
What is a machine learning DSP and how is it different from a traditional DSP?
A machine learning DSP uses algorithmic models to make bidding and targeting decisions in real time without manual rules. A traditional DSP relies more heavily on human-configured audience segments and bid rules. The difference matters because ML-driven systems optimize aggressively toward whatever signal you set, which can be a strength or a problem depending on how clearly you define your goal.
Should CPG brands test emerging DSPs expanding into CTV?
Yes, but only with a structured plan. That means a clear objective, an incrementality measurement approach, a defined budget ceiling, and a comparison baseline. Testing without those guardrails is just spending money with extra steps.
How do I know if a DSP’s machine learning is actually working for my category?
Ask for case studies specific to your vertical and your retail channel. Ask to see incrementality results, not just ROAS or reach. If they cannot provide category-relevant proof, that is meaningful information.
What is the biggest mistake brands make when evaluating a new programmatic partner?
Prioritizing CPM efficiency over measurement quality. A low CPM from a platform with weak attribution is not a deal. It is an untracked expense. Always start with “how will we know if this worked” before you ask “how much does it cost.”
—
Ready to build a DSP evaluation process that actually protects your media budget? Talk to the Junction 37 team.
Chris Pyne, Founder, Junction 37 – 30+ Years in Performance Media