From request to decision.
BidTrim is delivered as a container that runs in your own infrastructure, next to the auction. Your platform calls it for each request and gets back which demand partners to forward to. Your platform sends the requests. BidTrim never changes a bid request.
Illustration: inside your network, your auction calls the BidTrim container for each request and gets back a decision per demand partner. A request with a high bid probability is forwarded and one with a low bid probability is filtered. A request that carries a deal is not shaped. If no decision arrives by the deadline, the request is forwarded. Learning traffic is never shaped. Every request that is sent passes your own rate limiter before it leaves your network.
Interface
Your platform calls the container once per request, before fan-out, with the request and the demand partners it would send to. BidTrim returns a decision for each of them, with a reason code.
Signals
BidTrim reads request context: environment, device type and OS, app bundle or domain, publisher ID, country, format and size, video duration, pod and slot fields, floor, whether an identifier is present, and each demand partner's recent outcomes. It does not use identifier values and does not change the request.
Model
Machine-learning models estimate, for each demand partner, the probability that it bids on a request, learned from that demand partner's own outcomes. Bids, no-bids and timeouts are recorded separately. Models are trained by BidTrim and delivered to the container as files.
Decision time
A decision involves no call outside your network. Your platform sets the decision deadline on each call, and decision time is measured on your hardware before go-live.
Failure mode
If BidTrim does not answer by the deadline, returns an error or has no current model, nothing is filtered: your platform forwards the request under its own routing and rate limits. A model that can no longer be refreshed is used for a limited period, then switched off.
QPS limits
Your platform's rate limiter enforces each demand partner's QPS limit, per endpoint or data center. Each BidTrim decision carries the estimated bid probability, which the limiter can use when more requests are eligible than a limit allows.
Deals
Requests that carry PMP or programmatic guaranteed deals are not shaped. They are forwarded as your platform routes them.
Controls
You control it per demand partner: on or off, the share of traffic that is shaped, and the learning share that is never shaped. One switch turns BidTrim off everywhere.
Learning and rollout
A learning share of every demand partner's traffic is always sent unshaped. It keeps the models current. A rollout starts on a small share for one demand partner, is compared with a control allocation of equivalent traffic over the same period, and ramps in steps.
Reporting
The container reports its own decisions in aggregate, inside your network: forwarded and filtered volume by demand partner, segment and reason, and decision time. Revenue, spend and fill stay in your own reporting, where every rollout metric is split by shaped traffic and control group.
Container
The container image is signed, runs as a non-root user and needs no outbound network access to make a decision. It has health and readiness checks and a metrics endpoint, and your team can inspect it before deployment.
Protocols
BidTrim reads OpenRTB 2.5 and 2.6 bid requests in JSON. Pod fields are read from 2.6 requests.
What a rollout is measured on.
Each metric is compared between shaped traffic and the control group, over the same period.
| Metric | What it shows | How it is calculated |
|---|---|---|
| Bid rate | Requests that come back with at least one bid, out of the requests sent to that DSP. | requests with a bid ÷ requests sent |
| Outbound QPS | Requests sent per second to each DSP endpoint, against its limit. | requests sent ÷ elapsed seconds |
| Filter rate | Pairs of request and demand partner filtered by BidTrim, out of the pairs submitted to it and eligible for shaping. | filtered pairs ÷ eligible pairs submitted |
| Win rate | Bids that win the auction, out of the valid bids received. | winning bids ÷ valid bids |
| Decision time | Time BidTrim takes to return a decision, read as a distribution against the deadline your platform sets. | measured per call |
| Decisions not returned | Calls that returned no decision by the deadline, or an error. Those requests are forwarded unshaped. | calls without a decision ÷ calls |
| DSP timeouts and errors | Requests a demand partner did not answer in time, or answered with an error. | timeouts and errors ÷ requests sent |
| Fill rate | Impression opportunities that end in a billable impression. | billable impressions ÷ opportunities |
| Cost per million outbound requests | Processing and egress cost of the requests you send. | cost ÷ requests sent × 1,000,000 |
| Revenue per million inbound requests | Inbound traffic is the denominator, so sending to fewer DSPs does not inflate the number by itself. | revenue ÷ inbound requests × 1,000,000 |
| DSP spend | Billable spend per demand partner, over equivalent traffic and periods. | spend per demand partner |
Filtering is reported by demand partner and reason.
Example rows of the aggregate filter report, produced by the container inside your network, by demand partner, segment and reason.
- DSP BCTV · US · first in pod · floor above USD 9 CPMLow bid probability
- DSP DWeb · UK · outstream videoLow bid probability
- DSP AIn-app · DE · banner 320×50Low bid probability
- DSP EWeb · UK · outstream videoLow bid probability
Demos for supply platforms.
A demo covers integration options and the measures used to evaluate a rollout. Each request is reviewed before the time is confirmed.
Book a demo