When Should a Game Studio Add a Mobile DSP? 2026 Demand-Side Platform Pricing, an Evaluation Checklist, RFP Questions, and a Holdout Pilot
A DSP (demand-side platform) is an advertiser-side buying tool that lets you purchase app ad inventory scattered across many ad exchanges from one place, through real-time bidding. Unlike platforms such as Google or Meta, which own their inventory and their algorithms, a DSP evaluates and bids on open inventory from thousands of apps one impression at a time. For a game studio, adding a DSP is therefore less like switching on one more channel and more like buying control over which placements you pay for and at what price, together with the responsibility of running that control.
The problem is that the cost structure behind this decision is hard to see. The Association of National Advertisers (ANA) programmatic supply chain transparency study, released in December 2023, analyzed log-level data from 21 major advertisers and found that only 36 cents of every dollar entering a DSP effectively reached the consumer. In the quarterly benchmark the ANA has published since, the share of spend going to impressions that were measurable, viewable, fraud-free, and not made-for-advertising (MFA) reached a record 45.1% in Q2 2026, but the share taken by transaction costs rose to 27.2% over the same period. These figures describe programmatic spend by large advertisers in general and do not map one-to-one onto mobile game UA, but the lesson that inventory quality can improve while fees still need managing separately applies to game studios just as much.
This article is for game studios already running UA on Google, Meta, and ad networks who want to decide, as of November 2026, whether to add a mobile DSP. It covers when to add one, how to read DSP pricing, the difference between self-serve and managed, an evaluation checklist and RFP questions, signals that should make you hold off, and a pilot design that includes a holdout. How programmatic advertising works as a system is covered in Programmatic Advertising for Mobile Game Growth: How It Works and Where It Fits. This article takes that foundation as given and focuses on the buying decision: which DSP, on what terms, and how to verify it.
What is a DSP, and when should a game studio add one?
A game studio should add a DSP when its core channels can no longer buy more at the same efficiency and, at the same time, it has the budget, event data, and measurement in place to sustain a DSP's learning and verification. If only one of these is true, a DSP usually ends up as an expensive experiment.
A DSP does three things. First, it connects to multiple exchanges and SSPs and bids on other apps' ad placements impression by impression. Second, it uses machine learning models to decide how much to pay for which impression, based on the goal the studio sets (installs, a specific in-game event, or ROAS). Third, it lets you run both new user acquisition (UA) and retargeting, which brings back users who already installed, in the same tool. Compared with large self-serve platforms, a DSP shows you more about placements and bid logs, but more of the responsibility for inventory quality and fraud shifts to the studio.
When judging timing, ask "is our team ready to run a DSP now?" before "is DSP good?" The table below lists readiness criteria commonly used in practice.
Check | Ready to consider | Too early |
|---|---|---|
Limits of core channels | Scaling budget on Google, Meta, and similar platforms now visibly worsens CPI or ROAS | Core channel budgets have not yet been scaled far enough |
Test budget | A separate budget that can absorb several weeks of learning and the cost of a holdout | A budget that must hit target CPI within the first two weeks |
Event data | Post-install events (tutorial complete, level reached, purchase) are collected and posted back reliably through the MMP | No events defined beyond installs, or event values change often |
Measurement | Cohort ROAS and retention can be compared across channels on the same basis; incrementality testing experience or a plan | Network dashboard numbers are taken at face value as performance |
Creative supply | Banner, interstitial, video, and playable formats can be refreshed regularly | A handful of vertical social videos is all there is |
People | Someone who can review placement reports and adjust blocks and bids, or room for a managed contract | The UA manager already runs several channels alone |
The item most often underestimated in this table is event data. A DSP's machine learning learns from signals about what happened after the install. If those signals arrive late through the MMP or their definitions keep shifting, the bidding model ends up moving on install counts alone. The DSP then converges on finding cheap installs, which is also exactly where fraudulent traffic and low-quality inventory hide best. Where DSPs sit among all platform types is compared in How to Choose a Mobile Advertising Platform for Your Game in 2026: Evaluation Criteria, RFP Questions, and Pilot Design.
How is mobile DSP pricing structured?
Mobile DSP pricing falls broadly into a CPM-based structure that adds a fee on top of media cost and a structure that bills per outcome such as an install or an event. Either way, what determines your real cost is less the unit price than where the fees sit and how large they are.
Start with the CPM-based structure. The DSP bids per thousand impressions, and the studio pays the cleared media cost plus the DSP's fee (or a gross amount from which the fee is deducted). The share the DSP keeps is called the take rate. Most DSPs do not publish an official take rate. As a reference point, The Trade Desk, a publicly listed DSP focused on brand advertising, does not disclose its rate directly, but PPC Land derived roughly 21.6% from the company's 2025 reported results ($2.896 billion in revenue on $13.4 billion of gross platform spend). Mobile UA DSPs may structure their fees differently, but a sense that a share of this size can disappear somewhere you cannot see is a useful starting point for negotiating.
Buying on CPM and optimizing for CPM are different things. Most mobile DSPs bid in CPM units but let you set the goal that drives the bid as CPI, CPA (cost per specific event), or ROAS. If you set the goal to "cost per D7 paying user," for example, the DSP's model bids a higher CPM on impressions likely to produce that event and skips impressions unlikely to. You are billed on impressions, but you judge performance on events.
Outcome-based billing means the DSP takes on the impression cost risk and charges the studio a fixed price per install or per event. It makes budgets easier to forecast, but it also gives the DSP an incentive to choose placements where it can make a margin within the fixed price, so transparency about where installs came from often drops.
Pricing structure | Billing unit | Risk the studio carries | Check first |
|---|---|---|---|
CPM bidding + fee (rate disclosed) | Media cost per thousand impressions + a set percentage fee | Inefficiency during learning, inventory quality | Whether the fee is calculated on media cost or on the gross amount; whether data, fraud protection, and creative tools cost extra |
CPM bidding + fee (rate undisclosed, gross contract) | The DSP splits the studio's gross spend between media cost and margin | You cannot see how much actually went to media | Whether log-level reports show the clearing price |
CPM bidding + CPI, CPA, or ROAS goal optimization | Billed on impressions, goal set on events | Optimization breaks down if event signal quality is low | Minimum event volume per goal; relearning period when the goal changes |
CPI or CPA outcome-based billing | Fixed price per install or event | Price is fixed, but placement transparency can be low | Whether the apps and exchanges that drove installs are disclosed; whether installs flagged as fraud are excluded from billing |
Minimum spend and commitments | Monthly or quarterly minimum spend, annual commitments | Spending obligations regardless of performance | Whether minimums apply during the pilot; early termination terms |
Minimum spends are rarely published and vary by contract. Self-serve DSPs often have no minimum or a low one, but the studio runs the campaigns itself; managed DSPs often require a monthly minimum or quarterly commitment in exchange for providing operators. When discussing a pilot, make sure the same document answers "does a minimum spend apply to the pilot?" and "does the rate change when we move from the pilot to a full contract?"
Putting numbers on the fee impact makes the decision easier. The table below is an illustrative calculation: a studio spends $30,000 a month on a DSP, media prices and conversion rates are held constant, and only the fee share changes.
Fee share (assumed) | Amount reaching media | Installs at a media-level CPI of $3 | Effective CPI for the studio |
|---|---|---|---|
10% | $27,000 | 9,000 | About $3.33 |
20% | $24,000 | 8,000 | $3.75 |
30% | $21,000 | 7,000 | About $4.29 |
Buying the same inventory at the same efficiency, moving the fee share from 10% to 30% raises the effective CPI by about 29%. That is why DSPs should be compared not on the CPI in their dashboards but on "total amount the studio paid ÷ installs or events counted by the MMP." How the choice of optimization event changes bidding outcomes is covered in Mobile Game Ad Bidding in 2026: How to Choose, Set, and Scale the Right Strategy.
Self-serve or managed: which should you choose?
Self-serve means the studio runs bidding, targeting, and placement blocking itself; managed means the DSP's operations team runs them on the studio's behalf. The deciding question is who can review placement reports every week and act on them.
Self-serve offers control and learning. Your team sees directly which apps and exchanges perform and which creatives work on which placements, and keeps that knowledge. In return, blocklist management, bid adjustments, creative rotation, and fraud signal checks all become internal work. Managed reduces the operating load and lets you borrow experience the DSP has built on other game campaigns, but the reasoning behind operating decisions is less visible, and there is a risk that the metric the operations team optimizes diverges from the one the studio cares about.
Criterion | Self-serve | Managed |
|---|---|---|
Who operates | The studio's UA team | The DSP's operations team (the studio approves goals and budget) |
Minimum spend | Often none or low | Monthly minimums or commitments are common |
Transparency | Direct access to dashboards and logs | Report scope depends on the contract |
Internal capability needed | Programmatic operating experience, people for weekly placement reviews | Goal and event definitions, ability to verify results |
Good fit for | Teams running multiple titles with programmatic experience | First-time DSP adopters, small UA teams |
Watch out for | Heavy internal workload during learning | The operations team's metrics (installs, clicks) may differ from the studio's goals (ROAS, retention) |
In practice, many studios use a hybrid: start managed, learn placement and creative patterns over three to six months, then move part of the spend to self-serve. Either way, writing into the contract "what the operations team optimizes for" and "that the studio can receive log-level data at any time" keeps your data continuous even if you change the operating model later.
How should a game studio evaluate a DSP?
A practical DSP evaluation looks at six axes: inventory and supply path, machine learning signals on Android, fraud protection, creative formats, MMP integration and reporting, and cost transparency. The "AI optimization" that sales decks emphasize only means something once these six are in place.
First, inventory and supply path. What matters is not how many exchanges the DSP connects to but how much usable inventory exists for your game's genre and countries. When the same placement is resold through multiple paths, bids duplicate and fees stack. IAB Tech Lab's app-ads.txt is a standard that lets app developers publicly declare which companies are authorized to sell their inventory, and version 1.1 in August 2022 added fields to show owner and manager relationships. Read together with sellers.json, which discloses seller information, and the OpenRTB SupplyChain object, which records the transaction path, it lets you check whether a DSP buys only from authorized sellers and how many resale hops sit in between. At the evaluation stage, simply ask whether inventory not authorized through app-ads.txt is blocked by default.
Second, machine learning signals on Android. As of November 2026, Android ad measurement still rests on the advertising ID (GAID) and the Play Install Referrer. On October 17, 2025, Google announced it would retire most Privacy Sandbox technologies, and the list included the Attribution Reporting API on Android, Protected Audience, Protected App Signals, SDK Runtime, and Topics. The path that was expected to replace the advertising ID is gone, so for now the optimization quality of Android DSPs depends on advertising ID based learning and the quality of MMP postbacks. Under Google Play policy, however, apps targeting Android 13 or higher must declare the AD_ID permission, and when a user deletes the advertising ID, a string of zeros is returned instead. That is why you should ask a DSP how it bids on impressions without an advertising ID and how it feeds them into learning. iOS runs on a separate framework of SKAdNetwork and AdAttributionKit, so the same DSP may operate very differently there; we simply flag that here.
Third, fraud protection. Programmatic mobile app inventory carries a high share of invalid traffic. Pixalate analyzed more than 106 billion global programmatic impressions in Q3 2025, mostly buy-side open auction traffic, and found invalid traffic (IVT) rates of 33% for mobile apps, 21% for web, and 19% for CTV. These numbers describe the state of inventory across the industry and say nothing about any individual DSP, but whether a DSP uses its own pre-bid filters, integrates with third-party verification, and excludes installs the MMP flags as fraud from billing must be confirmed before you sign.
Fourth, creative formats. DSP performance in game UA depends heavily on the range of creative formats. Check supported formats such as banner, interstitial, video (rewarded and non-rewarded), playable, and native, the share of inventory by format, playable file size limits, and end card support. Starting with nothing but vertical social videos sharply limits the inventory a DSP can buy.
Fifth, MMP integration and reporting. As an official MMP integration partner, a DSP passes click and impression data to the MMP and learns from install and in-app event postbacks it receives back. Check whether postbacks include event values (revenue amounts), whether reports can be downloaded by app bundle, exchange, and creative, and whether the DSP can explain how much and why its dashboard differs from the MMP. Why every dashboard shows a different install count is covered in Why Every Dashboard Shows a Different Install Count: A Game Marketer's Guide to Mobile Attribution in 2026.
Sixth, cost transparency. The ANA benchmark cited above reported that in Q2 2026 the lower-performing advertiser group paid an average CPM of $6.55, but recalculated on quality impressions only, its CPM was $13.80; the higher-performing group paid $7.90 and $11.94 respectively. The group with the lower headline CPM was actually paying more per useful impression. When evaluating a DSP, judge it not by dashboard CPM but by whether it separates clearing prices from fees, and by what you actually paid for the one event you care about.
What questions belong in a DSP RFP?
The point of a DSP RFP is not to collect feature lists but to get answers in a common format on the six evaluation axes above: is it included by default, does it cost extra, and can it be written into the contract?
The questions below are written so you can send them to vendors as is. Once answers come in, transcribe each item as "included / extra cost / not supported / evaded" and compare the candidate DSPs in one table.
Area | Question | What it reveals |
|---|---|---|
Pricing | Is the fee calculated on media cost or on the gross amount? Do data, fraud protection, and creative tools cost extra? | Effective take rate |
Pricing | Do minimum spends or commitments apply to the pilot? Does the rate change when moving to a full contract? | Pilot cost and later terms |
Transparency | Can we receive log-level data separating app bundle, exchange, clearing price, and fee? Can we still receive data after the contract ends? | Whether costs can be verified |
Inventory | What is the monthly biddable impression volume and the top placement types in our target countries and genre? Is inventory not authorized through app-ads.txt blocked by default? | Real reach and supply path |
Optimization | Which events can be used as optimization goals, and what is the minimum event volume per goal? How long is relearning when the goal changes? | Whether and how fast learning can happen |
Android signals | How do you bid on impressions without an advertising ID? What is your measurement and targeting roadmap after the Privacy Sandbox retirement? | Response to signal loss |
Fraud protection | What is the scope of pre-bid filtering and third-party verification? Are installs flagged as fraud by the MMP excluded from billing? | Who bears the cost of fraud |
Creative | Which formats are supported, what is the inventory share by format, and what is the playable size limit? Do you provide creative-level reports? | Whether creative operations are workable |
Measurement | Which MMPs are you officially integrated with? Do you receive event revenue values via postback? How do you explain differences between your dashboard and the MMP? | The basis for comparing performance |
Incrementality | Do you support incrementality methods such as holdouts or ghost bidding? Who pays for the control group? | Whether net lift can be verified |
Operations | If managed, which metric does the operations team optimize, and what goes into the weekly report? | Goal alignment |
When comparing answers, trust answers with conditions over a one-word "supported." An answer that states scope and conditions, such as "log-level data is provided under enterprise contracts" or "installs flagged as fraud are credited if requested within 30 days of the MMP verdict," can be moved into a contract; an unconditional "yes" becomes a source of disputes over interpretation later.
What signals should make you hold off on a contract?
If a DSP refuses to provide log-level data, will not commit to excluding fraud-flagged installs from billing, or wants pilot performance judged only on its own dashboard, it is safer to hold off. All three conditions take away the studio's means of verifying performance later.
Other red flags commonly seen in practice:
Signal | Why it is a problem | Response |
|---|---|---|
The fee structure is explained only as "handled within the gross amount" | You cannot see how much actually went to media | Require reports that separate clearing price and fee as a contract term |
No app bundle level reporting | Low-quality placements and MFA apps cannot be filtered out | Secure at least a list of top placements and blocking controls |
Very low CPI from week one | Fraudulent traffic or incentivized installs may be mixed in | Check MMP fraud reports alongside D1 and D7 retention |
Click-to-install ratio abnormally low or high | Possible click spamming or click injection | Check the click-to-install time (CTIT) distribution |
Insisting on installs or click-through rate as the success metric | Misaligned with the studio's goals (ROAS, retention) | Fix pilot success criteria in writing before signing |
Requiring a long-term commitment as a pilot condition | Spending obligation before verification | Separate the pilot from the full contract |
Answering "not needed" to incrementality requests | Risk of claiming organic or other-channel results | Design the holdout yourself or consider other candidates |
Numbers that look too good in the first week deserve particular caution. A DSP bids broadly early in learning, so under normal conditions early efficiency is lower and improves as data accumulates. If you see the opposite pattern, check placement mix and fraud reports first.
How do you design a DSP pilot with a holdout?
The purpose of a DSP pilot is to confirm not "the performance the DSP reports" but "the performance that would not have happened without the DSP," and that requires fixing success criteria, a control group, duration, and stop conditions before you start.
DSPs often target the same users as other channels, so attribution numbers alone can credit the DSP with users who would have installed through another path anyway. The problem is especially large for retargeting, so comparing against a holdout group, a random subset of target users withheld from ads, is the default. For new user acquisition, user-level holdouts are often impractical, so studios either use geo-lift, turning the DSP on only in some of a set of similar countries or regions, or, if the DSP supports it, ghost bidding, which records impressions that won the auction but were not shown as the control. Experiment design details, sample sizes, and how to read the result are covered in How to Design an Incrementality Test for Mobile Game UA: Holdouts, Geo-Lift, Sample Size, and Reading the Result.
Phase | Duration (example) | Tasks | Pass criteria (example) |
|---|---|---|---|
Preparation | 1 to 2 weeks before start | Confirm MMP integration and event postbacks (with revenue values), block inventory not authorized through app-ads.txt, prepare creative formats, choose the holdout method (user, geo, or ghost bidding), document success criteria | Postbacks received on test installs; control design finalized |
Learning | 2 to 3 weeks | Start broad with CPM bidding and a CPI or early event goal (such as tutorial complete); review placement and creative reports weekly | Fraud-flag rate and D1 retention in line with existing channels |
Optimization | 3 to 4 weeks | If event volume is sufficient, switch to a deeper event (purchase, level reached) or a ROAS goal; block underperforming placements | Effective cost per goal event within a predefined range of existing channels |
Readout | 1 to 2 weeks after optimization | Calculate incremental installs and revenue versus the holdout, compare D7 and D30 cohort ROAS, calculate effective cost including fees | Incremental effect confirmed statistically; cost per incremental outcome below target |
Decision | Right after readout | Decide to scale, maintain, or stop; negotiate full contract terms | Renegotiate rate and minimum spend based on pilot results |
Write down stop conditions in advance too. For example: "stop if the MMP fraud-flag rate exceeds twice that of existing channels for two consecutive weeks," or "if goal event volume is below the threshold at the end of learning, move the goal one event shallower and extend by two weeks." Without rules like these, pilots tend to drag on with the promise that "a bit more learning will fix it."
Reading a pilot also requires a reference channel to compare against. Playio is an Android-based community of 5 million gamers who love games, and it uses AI to analyze genre preferences, play history, and in-game behavior data to connect games with gamers whose tastes fit. Where a DSP finds users impression by impression across open inventory, Playio is closer to designing the install and the play that follows within a pool of gamers already interested in games. Pricing defaults to CPI, and CPE based on in-game actions such as tutorial completion, reaching a level, or hitting a playtime milestone is also supported, so if you use the same MMP event you chose as the DSP pilot's success criterion as the performance basis, you can compare cohorts from both channels on the same yardstick. If a DSP pilot is coming up, share your target countries and genre and the one in-game event you will use as the success criterion, and we can review a small campaign setup that can serve as a reference cohort over the same period.
More details are available here. (https://playioadsen.oopy.io/bizdeck)
Key takeaways
A DSP is an advertiser-side buying tool that bids impression by impression on app inventory across many exchanges, and a game studio should add one when its core channels have reached their limits and it has budget for learning and a holdout, reliable event postbacks, and a way to compare cohorts on the same basis. As of November 2026, mobile DSP pricing splits into CPM bidding with a fee on top and CPI or CPA outcome-based billing, and since most DSPs do not disclose their rates, log-level reports separating clearing price and fee, whether minimum spends apply to the pilot, and billing exclusion for fraud-flagged installs should be contract terms. In the ANA benchmark, the share of spend reaching quality impressions was 45.1% in Q2 2026 while transaction costs took 27.2%, and Pixalate measured a 33% invalid traffic rate for programmatic mobile app inventory in Q3 2025. Evaluate on six axes: inventory and supply path (app-ads.txt, sellers.json), Android machine learning signals (the advertising ID and the roadmap after the Privacy Sandbox retirement), fraud protection, creative formats, MMP integration and reporting, and cost transparency; and judge performance not by dashboard CPI but by "total paid ÷ events counted by the MMP" and incremental lift versus a holdout. Choose between self-serve and managed based on whether someone can review placement reports every week, and set the pilot's preparation, learning, optimization, and readout phases and its stop conditions in advance, and adding a DSP becomes a predictable decision.
For inquiries about Playio's advertising solutions, reach out at: [email protected]
Sources (for data verification; whether to display them at publication to be decided in review):
ANA, Programmatic Media Supply Chain Transparency Study (released December 2023; log-level data from 21 major advertisers; only 36 cents of every dollar entering a DSP effectively reaches the consumer; $22 billion in available efficiency gains): https://ana.net/programmatic
Fiducia, ANA Programmatic Transparency Benchmark Q2 2026: Working Media Efficiency Reaches a Record 45%, 2026-08-06 (TrueAdSpend 45.1%, up 1.8 points from 43.3% in Q1; share reaching measurable, viewable, fraud-free, non-MFA impressions; transaction costs rose to 27.2%; lower performers CPM $6.55 and TrueCPM $13.80, higher performers CPM $7.90 and TrueCPM $11.94): https://www.fiducia.eco/post/ana-programmatic-transparency-benchmark-q2-2026-working-media-efficiency-reaches-a-record-45
PPC Land, Take rate, 2026-09-04 (take rate defined as fee retained ÷ gross spend transacted; The Trade Desk's roughly 21.6% for 2025 is PPC Land's calculation from $2.896 billion revenue on $13.4 billion gross platform spend, not a company-stated rate): https://ppc.land/take-rate/
Pixalate, Q3 2025 Global Ad Fraud Benchmarks Report, 2025-11-07 (more than 106 billion global programmatic impressions, mostly buy-side open auction traffic; IVT rates of 33% mobile app, 21% web, 19% CTV; IVT defined per the MRC): https://www.pixalate.com/blog/q3-2025-global-ad-fraud-benchmarks-report
Privacy Sandbox (Google), Update on plans for Privacy Sandbox technologies, 2025-10-17 (retired: Attribution Reporting API on Chrome and Android, Protected Audience on Chrome and Android, Protected App Signals, SDK Runtime, Topics on Chrome and Android, among others; timelines to follow Chrome and Android processes): https://privacysandbox.google.com/blog/update-on-plans-for-privacy-sandbox-technologies
Play Console Help, Advertising ID (apps targeting Android 13 or higher must declare the AD_ID permission; a deleted advertising ID returns a string of zeros; use the advertising ID for advertising purposes, App Set ID for analytics and fraud prevention): https://support.google.com/googleplay/android-developer/answer/6048248?hl=en
IAB Tech Lab, ads.txt / app-ads.txt (Authorized Digital Sellers; app-ads.txt extends the standard to apps distributed through app stores; version 1.1 in August 2022 added OWNERDOMAIN and MANAGERDOMAIN; supply chain validation specs alongside sellers.json and the OpenRTB SupplyChain object): https://iabtechlab.com/ads-txt/
Own calculations: installs and effective CPI at 10%, 20%, and 30% fee shares assuming $30,000 monthly spend and a $3 media-level CPI; pilot phase durations and pass criteria (illustrative)