BidTrim Book a demo

Traffic shaping methods.

The four components a supply platform combines to decide which bid requests each demand partner receives, and what each one can and cannot see.

Last reviewed: 15 November 2026

Traffic shaping means sending each demand partner fewer bid requests, chosen so that the requests it does receive are the ones it is likely to bid on. In practice it combines four components: capacity limits, eligibility rules, buyer-provided signals and supply-side estimates. They differ in who makes the decision, what the decision is based on and how it follows a change in demand.

The four components

ComponentWho decidesWhat the decision is based onWhen it follows a change in demand
QPS limitThe demand partner sets the limit. The platform's rate limiter enforces it.How many requests are sent. Which ones depends on the platform's selection policy.When the limit is changed
RulesThe platform, often at a demand partner's requestAttributes of a request: country, format, size, publisher, deviceWhen the rule's configuration is updated, by hand or automatically
Buyer signalsThe buyer supplies signals or model outputs. The seller evaluates them and applies the agreed forwarding policy.Combinations of request attributes that the buyer has marked as wanted or unwantedWhen the buyer publishes new signals
Bid predictionA model on the supply side, with one estimate per request and demand partnerThe estimated probability of a bid, learned from past bids, no-bids and timeoutsAs new outcomes are incorporated into a refreshed model

QPS limits

Every demand partner states how many queries per second each of its endpoints accepts. The limit protects the demand partner's infrastructure and bounds its cost of listening. It says how many requests may be sent. It does not say which.

When a platform has more eligible requests than a limit allows, it has to select which ones to send, and the selection policy determines which opportunities reach that demand partner. A policy that ignores the content of a request, such as taking requests in arrival order, cuts a request that would have received a bid as readily as one that would not.

Google's documentation for Authorized Buyers describes one platform's approach. Its callout system combines configured quotas with spend-based limits, error throttling and prioritization based on history. The quota is set and enforced per bidder endpoint and trading location, and requests over it are throttled. Requests the bidder is likely to respond to may be prioritized on the basis of its recent bidding history, using the publisher ID, the environment type and the ad format. Programmatic Guaranteed requests are always prioritized.

Rules

A rule filters on attributes of a request: no display requests from these countries for this demand partner, no banner sizes it never buys, no publishers it has blocked. Rules are transparent and take effect at once, and some are compulsory, such as a demand partner's blocklist or a compliance requirement.

A rule is as good as the attributes, the granularity and the update process available to the platform. Rules maintained by hand lag behind demand, which changes by the hour, by campaign and by budget. Rules generated automatically move closer to the other components on this page. In both cases a rule cuts everything that matches it, including requests that would have received a bid.

Buyer signals

With buyer signals, the buyer tells the seller what it wants, in a form that software can apply to every request. The best-known open framework is the IAB Tech Lab's Dynamic Traffic Engine (DTE), which began at Amazon Ads and is published by the IAB Tech Lab under the Apache 2.0 license.

The DTE manual divides the work as follows.

  • The buyer hosts the signal and configuration files in cloud storage. The signals come from the buyer's own traffic-shaping rules and algorithms.
  • The seller runs the DTE evaluator library inside its ad stack, or as a microservice beside it. The library pulls the buyer's files and syncs every five minutes.
  • For each evaluated ad slot, the evaluator returns a filter decision: 0.0 to filter, 1.0 to forward. The seller enforces the decision on the buyer's behalf.
  • Part of the traffic can be allocated to the buyer's learning. Those requests are always forwarded, whatever the filter decision, and they serve as the buyer's control traffic.

Three limits follow from the design. A decision exists only for a buyer that publishes signals, so every other demand partner still needs another component. The manual narrows its own scope: sellers are told not to evaluate requests that carry Programmatic Guaranteed deals, connected TV requests or native requests. And the control traffic needs explicit treatment when another shaping layer is applied: the integration has to preserve the agreed control path, or the buyer's measurement breaks.

This section describes version 2.4 of the manual, dated 1 June 2026.

Bid prediction

Bid prediction moves the estimate to the supply side. A model looks at the context of a request, such as the app or site, the format, the device type and the country, and estimates for each demand partner the probability that it will bid. It learns from what the platform already receives: the bids, no-bids and timeouts that come back.

Supply-side prediction can cover demand partners that publish no buyer signals. Its coverage and its speed of adaptation depend on the outcomes available, on the traffic left unshaped for learning and on how often the model is refreshed.

Its weak point is the feedback loop. A filtered request produces no response, so the model cannot learn that the demand partner has started to want it. The remedy is a share of traffic that is never shaped, which keeps the model learning.

Two cautions apply. Bid probability measures the likelihood of a response, not its value, so an evaluation also has to account for auction revenue, delivery, timeouts and processing cost. And measurement needs a control allocation: equivalent traffic, left unshaped over the same period. Learning traffic can serve as that control only if it is allocated at random.

How the components combine

The components are layers, not alternatives. One common integration pattern applies them in the order below. Whatever the order, it has to preserve deal handling and experiment allocations.

  1. Rules run first. Blocklists and compliance requirements are not predictions.
  2. Buyer signals apply for the buyers that publish them.
  3. Bid prediction filters pairs of request and demand partner among those that remain.
  4. The QPS limit is the outer bound, enforced by the platform's rate limiter. When more requests are eligible than a limit allows, the platform's selection policy decides, and it can use the estimated bid probability to do so.
Requests for one demand partner, layer by layer Illustration, not to scale

Illustration: of all the requests that match one demand partner, rules remove some, buyer signals remove more where the buyer publishes them, bid prediction filters those unlikely to get a bid, and the QPS limit bounds what is sent.

Each layer removes requests that the next one no longer has to consider. Learning and control traffic, which is forwarded without shaping, is not shown.

BidTrim is a bid prediction engine. It can be integrated alongside rules and buyer signals such as DTE, and it leaves the enforcement of each QPS limit to the platform's own rate limiter. How it works describes the decision.

Questions to ask of any method

  • Failure. What happens when the decision is late or missing? A shaping layer should fail open, so that requests are forwarded under the platform's own routing.
  • Deals. How are requests that carry deals handled, including private auctions and guaranteed delivery?
  • Control. Is there a control allocation, how large is it, who chooses it, and is it separate from learning traffic?
  • Reporting. Does the report show only the volume removed, or also what happened to bid rate, timeouts and revenue per million inbound requests?
  • Data. What leaves the platform, and is any of it personal data?
  • Switching off. Who can switch the method off for one demand partner, and how fast?

Sources

Checked on 15 November 2026.