Self-build
When it fits
- Maximum control
- In-house ad-tech team
- Own ops and billing
Why teams still switch
- Keep UX and data rules
- Avoid auctioning and SSP ops
- Outsource proof and billing
Two paths, one moment.
When your user sends a prompt, two things happen at the same time. The model thinks. An ad is matched. Neither waits for the other.
User sends prompt
Model path
Your API call runs exactly as usual
~800ms+
Ad path
Filtered signal goes to the ad market
< 20ms
Response + ad delivered
Filter
Only topic and language leave your system. No prompts. No PII. No history.
Match
OpenRTB 2.6 auction during model inference. No extra user-side roundtrip.
Prove
Every impression gets signed proof. Auditable. Billable. No self-reporting.
What data leaves your system
Your app
"Plan me a weekend trip to Barcelona with museum visits"
Raw prompt blocked
After firewall
topic: travel, lang: en
Filtered signal allowed
Ad market
Receives only abstract category and language. It never sees the original prompt, user data, or conversation history.
You control what leaves. Config is explicit. Missing config means blocked by default.
The timeline
0 ms5 ms15 ms20 ms800 ms+
Ad pathdone ~20 ms
Model path~800 ms+
Firewall 0.22 ms
SSP roundtrip 15.28 ms
Model inference ~800 ms+
The ad is ready before the model finishes. Zero added latency to the user.
Proof chain
1
Slot created
2
SSP fills
3
Beacon sent
4
Proof signed
Every impression is cryptographically linked to a compute unit. Settlement is signed, auditable, and billable. No self-reported counting.
In code
wrapper-flow.ts
typescript
SDK FLOW
const job = await client.createJob({
job_type: "chat",
model_id: "gpt-4o-mini",
locale: "en-US",
consent: { semantic_targeting: true },
prompt: { text: userMessage },
});
const decision = await client.getDecision(job.slot_ids[0]);
if (decision.fill) {
renderAd(decision.creative);
await client.sendBeacon({
beacon_type: "rendered",
asset_token: decision.asset_token,
});
}Compliance
Built on Compute Sponsoring v1.0
Profile: CS-S (S1/P0)
✓Ads and AI stay architecturally separated
✓Sponsors never receive user-level data
✓Semantic targeting requires explicit consent
Why wavebird
Ads are usually not the hard part. The real choice is ownership: market access, consent, proof, billing, and rollout.
Traditional ad stack
What works
- Mature demand
- Liquid market
- Standard OpenRTB rails
What is missing
- No GenAI-specific paths
- No semantic-consent layer
- No proof per compute unit
Subscriptions only
When it fits
- High ROI workflows
- Clear willingness to pay
- Strong paid conversion
Where it falls short
- Most users stay free
- Usage still creates compute cost
- No revenue on the free tier
Direct SSP path
What you get
- Direct demand access
- Full bid and policy control
- Custom consent and billing
Why wavebird sits on top
- Keeps market compatibility
- Adds filtered AI signals
- Handles proof and rollout