A demand-side platform (DSP) is software advertisers use to buy digital ad impressions from supported publishers and supply sources, often through automated auctions. It is useful when a media plan needs centralized buying, controls, and reporting across inventory sources the platform actually supports.
For example, a campaign seeking display, streaming video, and audio placements across several eligible sources may use a DSP to manage bids and reporting in one place. A campaign confined to one social network, search engine, or retail media environment may be simpler to run in that platform’s native buying tool.
The practical question is not whether a DSP has the longest feature list. It is whether the platform provides a capability the campaign needs, whether the account can access it in the target region, and whether the added fees and workflow are justified.
What is a DSP?
A DSP is the advertiser-side buying tool in programmatic advertising. The advertiser defines objectives and constraints. The DSP evaluates eligible impressions and submits bids or uses a supported deal. Publisher-side supply makes inventory available, and campaign reporting informs the next decision.
A DSP is not the publisher-side supply-side platform (SSP), which helps publishers offer inventory. It is also not an ad server, which delivers and measures ads, or a guarantee of access to every publisher, social network, search platform, or placement. Amazon’s DSP guide describes the general model as automated purchasing across supported digital advertising supply.
Google’s Display & Video 360 is an example of a buying platform. Campaign Manager 360 supports ad serving, attribution, reporting, verification, and integrations. It is not a complete DSP substitute.
How programmatic buying works
A campaign owner configures the objective, audience, geography, dates, budget, creative, inventory, bidding approach, and conversion measurement. The DSP then considers eligible opportunities. In real-time bidding (RTB), an impression may be offered in an auction that runs in milliseconds. Depending on the platform and inventory, buying may also use private marketplace deals or programmatic guaranteed arrangements. Auction mechanics and available deal types vary.
- Configure: The advertiser sets approved campaign inputs, including budget, dates, targeting, creative, and measurement.
- Evaluate: The DSP checks whether an impression or deal is eligible under those settings.
- Bid or transact: The DSP submits a bid or uses the configured deal. A bid does not guarantee a win, delivery level, or business outcome.
- Deliver and measure: The ad is served if the transaction succeeds. The platform reports results according to its definitions and attribution settings.
- Review: The campaign owner compares performance with the objective and decides whether approved settings should change.
Formats, channels, and inventory depend on the DSP, account, region, and supply source. Google says Display & Video 360 supports reservation, programmatic, and programmatic guaranteed campaigns across formats including display, video, TV, and audio. That description does not promise that every format or placement is available to every advertiser.
When to use a DSP, and when not to
Consider a DSP when the media plan has a specific need for multiple supported inventory sources, cross-channel campaign management, private deals, centralized controls, or consolidated reporting. Start with a native advertising platform when most spend and conversion value are within one closed network and there is no defined incremental capability that justifies another platform’s fees and workflow.
Before comparing vendors, list the required channels, countries, inventory sources, audience types, conversion events, reporting dimensions, and service model. Confirm that the specific account can access the required inventory and controls. Compare total cost, including media, platform, data, verification, and service charges. Centralized buying alone does not prove savings.
A DSP is justified by a campaign capability you can name and verify, not simply by the promise of a single dashboard. If the required inventory, deal, control, or reporting need is already met in a native platform, the extra operating layer may not be worthwhile.
Choose a platform by operating need, not feature count
Vendor capability pages describe what a company says its product can do. They are not independent evidence of performance, savings, accuracy, or universal availability. Use the examples below to frame a requirements discussion, then confirm current access, terms, and product behavior for the advertiser’s account.
| Platform example | Documented focus | Useful evaluation question |
|---|---|---|
| Display & Video 360 | Display, video, TV, audio, reservation, and programmatic buying options. | Are the required formats, inventory, and buying methods available in the target account and region? |
| Knorex XPO | Vendor-described cross-channel capabilities and KAIROS machine-learning optimization. | Which channels use eligible inventory or a platform-specific connection, and what can the owner control? |
| Jampp and Smadex | Mobile user acquisition and retargeting positioning, with product-described optimization and reporting capabilities. | Does the campaign need app installs, re-engagement, or in-app outcomes, and can the vendor report the required events? |
| MediaMath APIs | Documentation for campaign, creative, reporting, audience, and log-level functions. | Are credentials, permissions, quotas, report grain, and revision behavior confirmed before building an import? |
| Amazon Ads Agent | Amazon’s unified platform, with programmatic capabilities available through DVA+ and advanced settings. | Is the feature available to this advertiser in this region, and under which workflow and service terms? |
Knorex describes KAIROS as a machine-learning engine for allocating campaign budgets toward selected KPIs. That supports a capability claim, not a guaranteed efficiency or conversion improvement. Jampp and Smadex are positioned around mobile acquisition and retargeting, so they should be evaluated against app-install, re-engagement, and in-app-event requirements rather than treated as interchangeable general-purpose platforms.
MediaMath’s API documentation covers campaign, creative, reporting, audience, marketplace, and log-level functions. Access requires an account, credentials, and appropriate permissions. Documentation of an API category does not establish that every advertiser has access to every endpoint or that a direct CRM connector exists.
A Basis vendor presentation reported more than 25,000 audience segments across more than 30 data providers. This is a dated vendor-reported figure, not a guarantee of current availability or account eligibility. Verify current audience options and permissions directly. Audience data should be evaluated for provenance, consent, freshness, and suitability, not just quantity.
Amazon’s current materials describe Amazon Ads Agent as the unified platform, with programmatic capabilities available through DVA+ and advanced settings. Advertisers do not necessarily need to sell on Amazon, but account eligibility, regional availability, and service terms should be confirmed. Managed-service minimums vary by country and are time-sensitive. Do not treat a legacy quoted threshold as universal.
Set campaign guardrails before enabling optimization
Keep deterministic constraints explicit: approved budget ceilings, dates, geography exclusions, frequency policies, sensitive-category exclusions, consent requirements, creative approval, landing-page checks, and attribution settings. Model-assisted optimization may adjust eligible bids, rank approved audiences, or allocate budget among approved line items, but only within the authority the campaign owner has granted.
The following is an illustrative campaign specification, not a built-in DSP template. It includes an advertiser account, landing-page status, and attribution window so the activation contract is testable:
{
"advertiser_account_id": "authorized_account_identifier",
"campaign_name": "Spring prospecting",
"objective": "qualified_site_visits",
"budget": 12000,
"currency": "USD",
"start_at": "2026-11-01",
"end_at": "2026-11-30",
"timezone": "America/New_York",
"geography": ["US"],
"audience_source": "approved_first_party_segment",
"conversion_event": "qualified_visit",
"attribution_window": "owner_defined",
"creative_approval_status": "pending",
"landing_page_status": "pending",
"consent_status": "verified",
"human_approval_status": "pending"
}Parsing belongs before activation: convert brief details into allowed fields, validate dates, currency, timezone, budget, and inventory, and reject unknown values rather than guessing. AI can optionally organize a brief into a draft and flag missing or conflicting inputs. It must not authorize spend, infer consent, approve creative, or activate a campaign.
Reject activation if a required field or approval is missing, the date range is invalid, the destination fails validation, consent or audience provenance is unresolved, or requested inventory is unavailable. Measure operational success by checking that the campaign matches its approved specification and that delivery and reporting reconcile.
Measure performance without mixing unlike data
Define what one row represents before importing or comparing results. An impression record is not a campaign-day summary, and a report retrieval is not a new impression. Clicks and conversions are separate event types. Campaign-day totals are aggregates. Keep these grains distinct so a replayed file or revised report does not inflate totals.
For every imported metric, retain the source vendor and account, report definition, request or file reference, retrieval time, reporting period, timezone, currency, grain, attribution model, and ingestion run or revision. These are proposed provenance fields, not a common vendor schema. When a vendor exposes a documented event identifier, use it for event-level deduplication. If it does not, preserve the raw record and use a carefully scoped fingerprint without presenting that fingerprint as a vendor ID.
For an aggregate report, a possible row key is vendor, account, campaign, line item, creative, report date, timezone, currency, attribution model, report definition, and source revision. One row under that key represents the aggregate for those dimensions and that report definition. It does not represent an individual impression or a CRM contact or deal event.
Store the raw response first, validate its schema and grain, then write normalized data with a database-enforced unique constraint and transactional upsert. A read-then-insert check alone can race when imports run concurrently. For revised partitions, record the new source revision and supersede or replace the earlier aggregate rather than adding the revised total to it. MediaMath’s log-level documentation notes that recent impression partitions may be rewritten, so its ingestion process needs replay or reconciliation.
Do not assume an API, export, CRM connector, or permission exists because a vendor describes a related capability. Confirm the authorized access path and data use before designing the integration. A CRM should receive advertising data only when its purpose, access permissions, update grain, and system-of-record role are defined. Teams planning that boundary can review CRM systems consulting.
Cross-platform comparisons also need aligned reporting grain, currency, timezone, attribution window, and revision status. Attributed conversions describe a platform’s measurement method; they do not by themselves establish incremental lift.
Practical operating patterns
The following patterns separate documented platform capabilities from proposed implementation design. They are useful for evaluating readiness without implying that every vendor supplies the same modules, connector, or workflow.
| Trigger | AI job | Validation | Action and fallback |
|---|---|---|---|
| New campaign brief | Optionally extract proposed fields and flag conflicts. | Check account, dates, budget, consent, creative, destination, and inventory. | Create a controlled DSP draft. Keep activation with the campaign owner; reject unresolved fields. |
| Scheduled MediaMath report retrieval | No AI required. Anomaly summaries may be generated after reconciliation. | Check credentials, quotas, schema, grain, identifiers, and rewritten partitions. | Store the raw response and transactionally upsert normalized data. Route authorization or schema failures to an exception queue. |
| Cross-platform report arrival | AI may propose a field mapping for review. | Confirm vendor, account, report definition, date, timezone, currency, attribution, and grain. | Write the source-specific normalized record. Preserve raw input and supersede revised aggregates. |
| Amazon Ads Agent evaluation | No separate AI step is required. | Confirm DVA+ or advanced-settings access, region, inventory, permissions, measurement, and service terms. | Use the confirmed Amazon Ads Agent workflow. Pause if the required feature is unavailable or terms are unclear. |
These patterns do not establish a universal DSP-to-CRM connection. A production design still needs documented access, permissioned data use, rate-limit handling, pagination, report latency, and deletion procedures for each participating platform.
A practical DSP evaluation and launch checklist
Use this as a go or no-go review before committing spend or connecting reporting:
- The target account can access the required inventory, formats, and deal types in the campaign region.
- The advertiser understands media, platform, data, verification, and service charges.
- Audience provenance, permissions, consent, and deletion handling have an accountable owner.
- Budget limits, dates, exclusions, frequency rules, creative approval, and landing-page checks are explicit.
- Reporting access, grain, attribution definitions, timezone, currency, latency, and revision behavior are known.
- API or export permissions, credentials, quotas, and the reporting destination are confirmed before integration work begins.
- A named person owns launch approval, exceptions, and reconciliation.
If a required capability is unavailable, the cost model is unclear, or reporting cannot be reconciled at a defined grain, pause the launch or use a better-fitting native platform. A DSP is an operating choice, not a prerequisite for programmatic advertising.