Signal watches every click, skip, and dwell — then serves the next item they didn't know they wanted. In under 40 milliseconds. No hand-curation. No stale batch jobs. Just behavioral inference at production scale.
Choose your inference mode.
Signal ships three production-grade algorithms. Toggle between them and watch how your metrics shift in real time.
Learned ensemble + contextual bandit
Dynamically weights collaborative and content signals per user based on interaction density.
Signal vs. the alternatives.
No marketing copy. Just numbers. Every cell you read down this grid is a reason the benchmark runs itself.
| metric | signal | legacy provider a | open-source fork b |
|---|---|---|---|
| p99 latency | 38ms5.5× faster than legacy | 210ms | 890ms |
| cold-start handling | Hybrid fallbackAutomatic, zero config | Manual rules | None |
| integration steps | 3SDK → webhook → done | 47 | 120+ |
| real-time ingestion | < 1s lagSub-second behavioral sync | Batch (6h) | Batch (24h) |
| catalog size limit | Unlimited | 500K items | 100K items |
| A/B testing built-in | Yes — native | Third-party | Manual |
| nDCG@10 (benchmark) | 0.891+20% quality lift | 0.742 | 0.681 |
| monthly cost (1M MAU) | $2,4007.5× cheaper than legacy | $18,000 | $9,200 infra |
| SLA uptime | 99.97% | 99.5% | No SLA |
| explainability | Per-item reasons"Because you viewed X" | Black box | Black box |
Paste a product ID. Get recommendations.
This is Signal running live. No mock data, no pre-baked results. Your catalog ID in, five ranked items out.
Run a benchmark
against your stack.
We'll spin up a Signal instance against your current recommendation layer, run 72 hours of A/B traffic, and send you the full precision/recall report. No credit card. No sales call until you ask.
need data before the demo?
Read the Latency Report →trusted by product teams at