The Dynamic Traffic Engine (DTE) is a framework that lets a demand-side platform (DSP), the software advertisers use to buy impressions, tell a supply-side platform (SSP), the software publishers use to sell them, which bid requests it does not want before those requests are sent. Amazon Ads built it for its own DSP, announced it in beta on November 25, 2024, and donated it to the IAB Tech Lab on April 15, 2026, where it now lives as code under an Apache 2.0 licence. It exists because exchanges forward far more auction invitations than any buyer can use, and every unwanted request costs both sides bandwidth and computing power.

How the filter works

The system has three parts, according to its GitHub manual. DTE Cloud is storage hosted by the buyer, holding configuration and signal files. The evaluator library, supplied in Java and Go, runs inside the seller's ad-serving stack and checks each outgoing request against those files. DTE Filtering is the enforcement step: the SSP drops or forwards the request on the buyer's behalf. The reference implementation stores files in Amazon Web Services S3 buckets across four regions, though other cloud storage can be adapted.

When a page or app asks for an ad, the SSP builds an OpenRTB bid request, the JSON message format maintained by the IAB Tech Lab, and hands it to the evaluator before sending. A model configuration file tells the evaluator which features to extract, each defined as a JSON path into the request: whether an app or site object is present, whether the impression carries a video object, the publisher ID, the country in device.geo.country, the slot dimensions, the position field and the device type. Named transformations normalise the values. In the example configuration, positions 1, 4 and 7 (above the fold, header and full screen in the IAB's AdCOM lists) collapse to "a", position 3 to "b", and everything else to "u".

The output is a pipe-delimited tuple. The manual's example, app|banner|537075271|MEX|320x50|a|4, describes an in-app banner from publisher 537075271, shown in Mexico at 320 by 50 pixels in a prominent slot on device type 4, which AdCOM defines as a phone. The evaluator matches that string against the buyer's hourly signal file and returns a filterDecision of 0.0 (filter) or 1.0 (forward) for each impression object in under a millisecond. Signal files take between 1 and 10 megabytes of memory. The library polls storage every five minutes, so a buyer can change what it receives without anyone shipping code.

In a LowValue model the listed tuples are traffic to drop and anything unlisted passes. In a HighValue model the list is traffic to keep. The deals model added in version 2.0 is of the second kind, keyed on deal IDs.

An experiment configuration stops the buyer going blind. Each request receives an allocation ID between 0 and 4095, and the published example splits traffic 80 to 20 between a treatment group, which is filtered, and a control group flagged learning=1, which is always forwarded. Both values travel in extension fields, so the buyer can compare filtered and unfiltered traffic and retrain. The design fails open: if no fresh model can be fetched, the library reuses the last one for up to 24 hours and then stops filtering; if evaluation throws an error, the request goes through. The manual tells sellers not to evaluate programmatic guaranteed deals, connected TV requests or native requests at all.

The manual estimates one to two weeks of integration work, or about a month for a custom evaluator, plus automated weekly reports per data centre on filter rate, fill rate, bid rate and spend per million requests.

From Amazon beta to open source

Sellers filtered traffic long before DTE, but by guessing. SSPs modelled each DSP's recent bidding, typically hour by hour, according to TripleLift, which said programmatic supply grew 2.3 times between 2020 and 2023 and outpaced demand budgets. Buyers answered with caps on queries per second (QPS), the number of requests they accept from each exchange. DTE reversed the direction of the model: the buyer writes it and the seller executes it.

Amazon's November 2024 announcement named Yieldmo and TripleLift as live partners, with Index Exchange to follow. The specification's own history begins with version 1.1, a closed beta, on January 14, 2025. TripleLift disclosed its participation on February 4, 2025, replacing predictive models with Amazon's signals. Version 2.0, dated July 1, 2025, brought the model type field, the deals model and a reference Dockerfile for container deployments; Equativ said the same month that it had helped containerise the library. Version 2.1, on September 24, 2025, added optional revenue per million ad requests (RPMA) files for sellers, excluded from filter decisions. Version 2.3, on February 5, 2026, rewrote the manual in buyer-neutral language.

On April 15, 2026 the IAB Tech Lab took the project into its Open Source Initiative. Pieter de Zwart, director of engineering at Amazon Ads, described the change as "shifting from guessing to knowing". Initial proposals would place DTE within the Agentic Advertising Management Protocols, the umbrella the IAB Tech Lab named on February 26, 2026, using the Agentic RTB Framework opened for comment on November 13, 2025 to update DTE components in real time. Amazon archived its own repository on July 7, 2026 and now directs contributions to the IAB Tech Lab copy.

Why buyers and sellers care

Request volume drives cost whether or not anything sells. Publishers including Chegg and the Daily Mail began throttling duplicate requests in 2025, and PubMatic started billing publishers for volume above their allocations in April 2026. Duplication compounds the cost: a DataBeat report found 46% of publisher domains behind established SSPs reachable by more than one route. DTE lets a buyer shed traffic it will never bid on without lowering a QPS cap, which cuts good and bad supply alike. Anthony Katsur, chief executive of the IAB Tech Lab, called it a "smart traffic filter", according to AdExchanger.

Every reported result is self-supplied. Amazon said October 2024 tests cut unwanted requests from one SSP by more than 40%, and Yieldmo reported a 7% average gain in fill rate and spend per million requests. PubMatic said on March 30, 2026 that its global integration lifted eCPM by up to 10%, citing its own December 2025 and January 2026 data. On July 21, 2026, Amazon said OpenX saw RPMA rise 41.4% and programmatic spend grow 11.0% within three months, with cost per action (CPA) down 3.4%, while conceding that CPA reflects every Amazon DSP optimisation feature and that seasonal effects are included.

For media buyers, a campaign on Amazon DSP sees a bidstream pre-sorted by Amazon's own view of value. That ties DTE to Amazon's Certified Supply Exchange programme, which Magnite joined as one of the first partners in June 2024and Microsoft Monetize joined on October 7, 2025. Amazon lists traffic optimisation among the programme's eight certification areas.

Limits and disputes

The first objection is structural: the signals express one buyer's wishes, and the seller enforces them. Publishers never see the files, and one tuple can remove an entire publisher, country and slot-size combination. Coverage of the OpenX integration noted that shaping systems can discard valuable inventory while letting weak impressions through. The control group limits that error but does not remove it.

Neutrality is incomplete. Despite version 2.3's generalised wording, the specification's extension namespace is still "amazontest", the reporting template asks for total bid requests sent to Amazon, and the reference cloud is AWS. PPC Land has found no public announcement, as of September 18, 2026, of any other DSP publishing DTE signals, which leaves a shared framework with a single producer.

Coverage has gaps. Connected TV sits outside the manual's scope, yet streaming is where request volumes spike; StackAdapt reported a 71% QPS reduction by bidding on whole ad pods through Index Exchange, a different technique. Code maturity is another question: on September 14, 2026, an issue in the Tech Lab repository reported a data race in the Go library that could crash the host bidding process during a configuration refresh, with a proposed fix still open at the time of writing.

Rivals argue the constraint should be removed rather than rationed. Magnite set out two routes for partner models to run inside its auction on August 6, 2026, and the ARTF container standard puts buyer logic next to the exchange. Once bidding code sits beside the auction, the cost DTE saves shrinks.

What it is not

Traffic shaping is the general practice of choosing which requests reach which buyer. DTE is one implementation of it, distinguished by the buyer writing the rules.

QPS caps and bid throttling limit volume. A cap says how many requests; DTE says which.

Curation packages inventory and data into deal IDs for particular campaigns. Katsur has likened DTE to curation applied across a DSP's entire client base, but curation produces a transactable deal and DTE produces none.

Pre-bid filtering usually means a buyer excluding unsafe or fraudulent inventory at bid time, after the request arrives. DTE acts before it leaves the exchange.

Recent developments

Amazon's July 21, 2026 update said Magnite was integrating DTE across all supported data centres, joining OpenX, PubMatic, TripleLift, Yieldmo and Equativ among disclosed sellers. The update dated OpenX's integration to September 2025, although OpenX announced it on December 22, 2025; the gap may separate testing from launch.

The agentic proposal remains the larger unknown. If autonomous systems can rewrite signal files every five minutes, the filter becomes a live expression of campaign strategy, and sellers would enforce decisions no human reviewed.

Timeline

  • June 2024: Magnite becomes one of the first partners in Amazon Ads' Certified Supply Exchange programme
  • October 2024: Amazon tests DTE with SSPs, reporting a reduction of more than 40% in low-demand requests from one partner
  • November 25, 2024: Amazon Ads announces Dynamic Traffic Engine in beta with Yieldmo and TripleLift
  • January 14, 2025: Specification version 1.1 released as a closed beta
  • January 29, 2025: Version 1.3 adds the learning flag for control groups
  • February 4, 2025: TripleLift discloses its participation in the beta
  • February 28, 2025: Version 1.4 adds failure handling
  • July 1, 2025: Version 2.0 adds model types, the deals model and a reference Dockerfile
  • July 2025: Equativ reports helping containerise the evaluator library
  • September 2025: OpenX integration begins, according to Amazon
  • September 24, 2025: Version 2.1 adds RPMA monetisation insight files
  • October 7, 2025: Microsoft Monetize joins the Certified Supply Exchange
  • December 22, 2025: OpenX announces its DTE integration
  • February 5, 2026: Version 2.3 generalises the manual's language
  • March 30, 2026: PubMatic reports its integration live globally
  • April 15, 2026: Amazon Ads donates DTE to the IAB Tech Lab Open Source Initiative
  • July 7, 2026: Amazon archives its repository in favour of the IAB Tech Lab copy
  • July 21, 2026: Amazon publishes OpenX results and Magnite's integration across data centres
  • September 14, 2026: A data race in the Go library is reported in the IAB Tech Lab repository

Summary

Who. Amazon Ads built the Dynamic Traffic Engine for Amazon DSP and donated it to the IAB Tech Lab. Disclosed sell-side integrations include Yieldmo, TripleLift, Equativ, OpenX, PubMatic and Magnite. Anthony Katsur of the IAB Tech Lab and Pieter de Zwart of Amazon Ads are the named spokesmen for the donation.

What. A file-based signalling framework in which a buyer publishes rules describing low-value or high-value bid requests, and an evaluator library inside the SSP applies them to each OpenRTB request, forwarding or filtering it while reserving a control group for learning.

When. Announced in beta on November 25, 2024, specified from version 1.1 on January 14, 2025 to version 2.3 on February 5, 2026, and donated to the IAB Tech Lab on April 15, 2026.

Where. Inside SSP ad-serving infrastructure, between the construction of a bid request and its dispatch to the buyer, with signal files hosted in cloud storage across four regions.

Why. Exchanges send buyers far more requests than they bid on, and each costs money to transmit and process. DTE lets buyers state what they want instead of cutting volume blindly, though the rules still reflect one buyer's interests and the reported gains come from the companies involved.