THE SHORT ANSWER
Key points
- Start the match in parallel with model generation.
- Never block the first token or completed answer on an ad request.
- Use strict timeouts and treat no response as no ad.
- Measure answer latency separately from placement latency.
01
The non-blocking pattern
- 01
The user sends a message
Your app starts its normal model request immediately.
- 02
Matching starts in parallel
The latest request and allowed context are evaluated without pausing the model.
- 03
The answer streams normally
The user sees the same response timing they would see without monetization.
- 04
The placement resolves
If a useful match is ready, render it after the answer. Otherwise render nothing.
02
Engineering rules that protect speed
- Set a short timeout for the placement request.
- Do not retry in the foreground during the same turn.
- Reserve no large blank space while waiting.
- Load optional images lazily.
- Record placement failures separately from answer failures.
03
What should you measure?
Track time to first token, time to complete the answer, placement response time, placement render time, and error rate. Compare these before and after installation.
Also watch product measures such as follow-up rate and return usage. A technically fast placement can still be harmful if it makes the experience feel less useful.
COMMON QUESTIONS
What else should you know?
Should an AI answer wait for an ad response?
No. The answer should be generated and shown independently of the ad request.
What happens when matching times out?
The app should continue normally and show no placement for that turn.
Can product images affect performance?
Yes. Use fixed dimensions, modern formats, and lazy loading so optional creative does not shift or delay the interface.
FOR AI DEVELOPERS