How to Choose a Mobile DSP: The Questions That Matter

Most DSP evaluation checklists are useless. They ask about supported formats, targeting options, geographic coverage, and reporting dashboards. Every platform on your shortlist answers yes to all of it. You finish with four vendors who look identical on paper, and you decide based on rapport.

The questions that separate platforms are the awkward ones. Here are the ones we'd want asked.

Why standard DSP evaluations don't work

A feature checklist compares capabilities. Capabilities converged years ago. What still varies enormously between platforms is economics, transparency, and what happens to your money on a bad week — and none of that appears on a feature grid.

Ask about mechanics and incentives instead of features. A vendor's answer to "what happens when this doesn't work" tells you more than any deck.

What is your fee structure per hop?

Start here, because it determines what every other answer means.

You want the take rate stated as a percentage of your spend, and you want to know how many intermediaries sit between your budget and the publisher. Supply that reaches you through a reseller that bought it from another network, which sat on top of the publisher's own supply-side platform, has passed through three sets of hands taking margin before anyone shows an ad. Your effective CPM bears no relation to the rate you agreed to.

Ask for it in writing. A platform that describes its pricing as "we're competitive," or answers with a blended number,de7 is telling you the detail is unfavorable.

The follow-up matters more: does the take rate change by channel, format, or supply source? Often it does, which means the platform has an undisclosed financial preference about where your budget goes.

Will you give me log-level data?

This is the single most revealing question on the list, and plenty of platforms will not answer it well.

Log-level data means impression-level records: what was bought, from which supply source, and at what price. With it, you can audit the supply path, spot duplicated inventory, and verify that the CPM you were charged is the CPM that cleared.

Some platforms provide it. Some provide it on request for larger accounts. Some don't provide it at all and explain that their reporting is sufficient. That last answer isn't automatically disqualifying, but it should change your price expectations, because you're paying for a service you cannot verify.

Ask specifically: at what account size, in what format, at what frequency, and at what additional cost.

What does your model optimize toward, and what happens when I change it?

Every DSP will tell you its machine learning optimizes toward your goal. The useful question is what happens mechanically when that goal changes.

If you move from installs to a post-install event, does the model retrain from scratch, and how long does it take? How much event volume does it need before its bids are better than random? Ask for the number. A platform that needs 200 conversions a week to learn is a different proposition from one that needs 20, and if your event fires 30 times a week, the first one will never work for you no matter how good the technology is.

Then ask what the model does while it's learning. Some platforms bid conservatively and underdeliver. Others spend aggressively to gather data. You're paying for that exploration either way, and you should know which pattern to expect before you judge week one.

What can you actually see on iOS?

Ask this even if the answer sounds technical, because a vague response here predicts trouble later.

Apple's AdAttributionKit supports three conversion windows, covering days 0 to 2, days 3 to 7, and days 8 to 35. It sends the fine-grained conversion value only in the first postback and returns up to three postbacks in total, but only if the advertised app updates its conversion value within each window. Postbacks arrive after a random delay of 24 to 48 hours for the first and 24 to 144 hours for the others. Installs that fall into the lowest data tier return a single postback with no conversion value.

So no platform sees your iOS performance clearly in week one, and any that implies otherwise is describing modeled output as though it were measurement. What you want to hear is a specific account of which signals the platform uses, what it models, and how confident it is in each. "Our AI handles it" is not an answer.

What happens on a bad week, and who absorbs it?

Most programmatic media is bought on impressions, which means the outcome risk sits with you: the impressions arrive whether or not they do anything. That's normal and not a criticism. It's still worth asking whether the platform offers any structure that shifts that risk elsewhere.

Outcome-priced buying exists in places — offerwall inventory can be bought per completed action rather than per impression, which moves that risk to the supply side, while rewarded video prices on completed views. Not every channel supports it, and it costs a premium per event, but if a platform has never mentioned the option, ask why.

The broader question underneath: has this vendor ever told a client to spend less? The ones who have will tell you the story. The ones who haven't will change the subject.

Will you run an incrementality test against your own performance?

This is the question that ends evaluations, and we include it because we know it's uncomfortable.

A geo holdout suppresses the channel in matched markets, runs it in others, and compares total conversions with organic included. It doesn't care what the dashboard says, because it never asks which touchpoint deserves credit. That is what separates conversions the platform caused from conversions that would have happened anyway. Other designs reach the same question, including PSA tests, ghost ads, and user-level holdouts. Still, a geo test is usually the one an advertiser can run without needing the platform's cooperation.

A vendor confident in its performance will help you design one. A vendor that resists, or explains why incrementality doesn't apply to its channel, has told you something.

Be fair about the cost, though. A holdout means going dark in real markets while it runs, and not optimizing during it. How long it needs to run depends on your conversion volume, and the honest answer comes from a power calculation on your own history rather than a number a vendor quotes you. Don't ask for one during a pilot too small to produce a readable result.

How do you handle fraud, and who pays for it?

Two separate questions, and most vendors answer only the first.

The detection story matters less than the commercial one. When invalid traffic is identified after the fact, who bears the cost? Is there a clawback window? How long is it? Does it cover the full amount or a percentage? Get the answers before you sign, because the aftermath of an incident is a bad time to discover the terms.

What we'd tell you to do with the answers

Score the awkwardness. The platforms that answer the uncomfortable questions plainly, including where the answer isn't flattering, are usually the ones worth shortlisting. Vagueness on fees, log-level data, or incrementality is rarely accidental.

Then test small before you commit. One channel, one clearly defined goal event, enough budget to produce a readable result, and a fixed window you agree in advance not to interfere with.

We're a mobile performance DSP running in-app, CTV, and rewarded, so we're an interested party here rather than a neutral one. These are the questions we'd expect a buyer to put to us, and we've written elsewhere about the platform and policy changes reshaping in-app measurement this year, several of which sit behind the iOS question above. If you're running an evaluation and want a straight answer to any of the above, ask us.

FAQ

What should you ask a DSP before signing?

What the fee structure is per hop, whether log-level data is available and at what account size, how much event volume the optimization model needs, what the platform can actually see on iOS, and whether it will support an incrementality test against its own results.

What is log-level data, and why does it matter?

Impression-level records showing what was bought, from which supply source, and at what price. They let you audit the supply path and verify the price you were charged. Without them, you're paying for performance you can't independently confirm.

How long should a DSP test run?

Longer than most plans allow. On iOS, the third conversion window closes 35 days after a user's first launch, and its postback arrives up to 144 hours after that, so the measurement tail runs roughly six weeks past your last install. Budget for the flight plus that tail, and agreeing not to interfere while the test runs matters as much as the duration.

Is a cheaper take rate better?

Not by itself. A lower stated rate on a longer supply path can cost more in total than a higher rate on a direct one. Compare effective cost per outcome after all hops, not the headline percentage.