Demand-Side Platforms (DSPs)
What a DSP actually does
A Demand-Side Platform is the software an advertiser (or their agency/trading desk) uses to participate in programmatic auctions at scale. Concretely, a DSP takes a human-defined campaign brief — budget, target audience, flight dates, bid strategy, creative assets — and translates it into an automated system that evaluates every single incoming bid request across potentially millions of impressions a day, decides which ones to bid on, calculates what to bid, and submits that bid within the auction's millisecond-scale window.
No human is involved in any individual auction. The human sets the strategy once; the DSP executes it continuously, learning and adjusting bid prices in real time based on which impressions are actually converting into the outcomes the campaign is optimizing for.
Audience targeting across inventory the DSP doesn't own
The core value proposition of a DSP is reaching one defined audience consistently across inventory from thousands of unrelated publishers and apps, rather than rebuilding that audience separately inside each platform's own walled-garden ad manager. A DSP does this by layering several targeting data sources on top of the raw bid request data: first-party data the advertiser uploads (e.g. a CRM list, hashed and matched), third-party data segments licensed from data providers, contextual signals read directly from the bid request (page category, keywords, app genre), and increasingly, the DSP's own modeled audiences built from its aggregate view of behavior across its entire client base.
This is also where privacy regulation has hit hardest: third-party cookie deprecation and the decline of mobile ad IDs (IDFA restrictions, Android's Privacy Sandbox) have forced DSPs to shift weight toward contextual targeting, clean-room data collaborations, and privacy-preserving cohort-style targeting (the aims behind Google's Privacy Sandbox and the IAB's various identity-alternative initiatives) rather than individual-level identifiers.
Advertiser sets strategy once:
Budget: $50,000 flight
Audience: CRM match + lookalike + contextual (finance content)
Bid strategy: Target CPA $40
Creative: 3 banner sizes + 1 video
DSP then, per impression (millions/day):
1. Receive bid request
2. Match against audience criteria → qualify / skip
3. If qualify: predict conversion probability
4. Calculate bid = predicted value × pacing adjustment
5. Submit bid (or skip if floor too high / budget paced out)
6. If won: serve matched creative
7. Feed outcome data back into the prediction modelBid strategy and pacing
A DSP doesn't just bid the maximum it's told to on every qualifying impression — that would exhaust a daily or flight-level budget within minutes on high-volume campaigns. Pacing algorithms spread spend across the intended time window, dynamically throttling bid aggressiveness up or down so the budget lasts the full flight rather than front-loading entirely on the cheapest early impressions.
Bid strategies typically map to the optimization goal: Target CPA bidding adjusts bid price per impression based on predicted conversion probability, bidding more for impressions the model believes are more likely to convert; Target ROAS does the same but weighted by predicted order value, not just conversion likelihood; Viewability-weighted bidding (common in brand/awareness campaigns) prioritizes impressions likely to actually be seen (above the fold, sufficient time in view) over raw impression volume. Advertisers choosing the wrong strategy for their goal — e.g. running viewability-weighted bidding on a performance campaign — frequently see spend concentrated on the wrong impressions for what they're actually trying to achieve.
DSP buying vs direct platform buying (e.g. Meta/Google Ads)
| Aspect | DSP (programmatic) | Direct platform (Meta Ads, Google Ads) |
|---|---|---|
| Inventory access | Open web, apps, CTV across many publishers | Limited to that platform's own owned-and-operated inventory |
| Targeting data | First-party upload + third-party + contextual | Platform's own first-party user data, generally richer per-user |
| Setup complexity | Higher — brand safety lists, floor management, creative specs per format | Lower — platform handles most inventory curation |
| Audience consistency across properties | High — one audience definition, many publishers | Low — rebuilt separately per platform |
| Typical fit | Scale, retargeting across the open web, CTV/video reach | High-intent search, in-feed social, platform-native formats |
Walled gardens are technically DSPs too, just closed ones
Choosing between DSPs: what actually differs
Major DSPs (The Trade Desk, DV360, Amazon DSP, Yahoo DSP, and others) differ less in raw auction mechanics — they're all speaking OpenRTB to largely overlapping exchange inventory — and more in: proprietary data access (Amazon DSP's retail purchase-intent data is a genuine moat no competitor can replicate; DV360 gets privileged access to Google's own inventory signals), the sophistication of bid shading and prediction modeling, measurement and attribution integrations, and minimum spend commitments / self-serve vs managed-service access tiers.
For advertisers already spending heavily inside Amazon's retail ecosystem, Amazon DSP's access to actual purchase behavior as a targeting signal is frequently decisive regardless of other feature comparisons — no independent DSP can replicate first-party retail purchase data it was never given access to.
DSP reported metrics and advertiser-side measurement rarely match exactly
What's next
DSP-driven campaigns frequently exist alongside retargeting strategy built on pixel data rather than cold prospecting — see Retargeting & Remarketing Strategy for how funnel-stage-specific targeting is layered on top of the raw bidding mechanics covered here.
Next: Retargeting & Remarketing Strategy →
I build these systems professionally.
Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.