How AWS Elemental Inference contextual metadata reaches a live ad break
AWS Elemental Inference classifies live video at the shot level and the scene level. AWS says plainly that those classifications are there to enable contextual ad targeting (AWS Elemental Inference docs). That's a big signal for anyone selling live and FAST inventory: the biggest name in cloud video now treats what's on screen as something the ad decision should know about.
So how does the signal actually get from the encoder to the ad server? AWS documents it well. This is the short version.
What Inference produces
Inference runs alongside encoding, not after it. From live video it pulls out content categories, GARM brand suitability signals, the objects and actions it detects, and descriptions at shot and scene level (AWS What's New, 10 Sep 2026).
The path to the break
AWS's MediaTailor guide lays out four steps (AWS MediaTailor guide):
- Inference analyses the content and classifies each shot.
- MediaLive writes Inference details into the SCTE-35 ad markers, so MediaTailor knows where to look.
- At each break, MediaTailor asks Inference for the classifications covering the content just before it.
- The function picks out the content categories and GARM signals and hands them to the ad decision server.
You set it up as a MediaTailor function that runs just before the ad request goes out (AWS calls this the PRE_ADS_REQUEST hook).
Before the first break
Five settings matter. All of them come from the same AWS guide.
- Access. The Inference feed needs a resource policy that lets MediaTailor read it.
- Enrichment. Switch on contextual metadata enrichment on the MediaLive channel. AWS supports the SCTE-35 route for production and says manual timing is for testing only.
- The window. Each lookup covers content since the last break, going back 30 seconds at most. That's the maximum lookback Inference supports.
- Processing time. Summary generation and extended analysis are both on by default, and both add processing time. If you only need ad targeting, AWS recommends turning them off. Otherwise the content right before the break may not be ready yet.
- Timeout. The console default is 2,000 milliseconds. If the lookup fails or times out, MediaTailor carries on with the break without it.
Where the signal lands
The categories can ride on the ad decision server URL as player parameters, go out as request headers, or replace the ad server URL outright (same AWS MediaTailor guide).
Channels outside the AWS chain
All of this assumes MediaLive and MediaTailor end to end. Plenty of channels don't run that way. That's the gap Edge Video's Autonomous Commerce for ads fills.
Edge reads the channel's live stream and can tell what's on screen, down to naming the sport being played. It works upstream of ad insertion. By the time the ad decision is made, it already carries standard content categories describing the moment.
Edge Video builds Autonomous Commerce for ads.