Contextualizing Ad Recommendations Through Meta's AI Business Assistant
Ads Manager shows advertisers recommendations for actions that could improve their account and ad performance. Advertisers didn't always act on them — the recommendations were necessarily short, given limited real estate in the UI, and often didn't explain why a given recommendation mattered to that advertiser's specific account. The team added two entry points into the AI business assistant's chat experience to address this: one on the recommendation cards themselves, and another as a dropdown at the top of the campaign table, aiming to increase engagement with the assistant and adoption of the underlying recommendations.
Approach
I wrote the system prompts for this use case, instructing the assistant to contextualize each recommendation within the advertiser's own account — current performance, previous performance, and the advertiser's stated goals — and to explain in plain language why the recommendation was being made and how it could help. I also shaped the structure of the response to be easy to scan and to tell a clear story for why the recommendation mattered, since the goal wasn't just to inform advertisers but to motivate them to act.
For the campaign-table entry point, I wrote the dropdown content itself, grounding the topic options in user engagement data and vetting the most relevant topics with marketing science. I then wrote the system prompts that triggered the assistant's personalized response for each of those topics, so selecting a topic from the dropdown led to the same kind of account-grounded, plain-language explanation as the recommendation-card experience.
A short recommendation on a card can tell an advertiser what to do. It takes account context to tell them why it matters to their business — and why is what actually moves someone from reading a recommendation to acting on it.
Key decisions
- Instructed the assistant to ground every recommendation in the advertiser's own account data — current performance, previous performance, and stated goals — instead of generic reasoning.
- Wrote prompts that required the assistant to explain the "why" in plain language, translating account signals into a rationale advertisers could actually follow.
- Structured responses to be easy to scan and built around a clear narrative, since the goal was to motivate action, not just deliver information.
- Grounded the campaign-table dropdown's topic options in user engagement data, then vetted the most relevant topics with marketing science before writing the content.
- Wrote system prompts that triggered a personalized, account-grounded response for each dropdown topic, keeping the experience consistent with the recommendation-card entry point.
Outcome
- Increased the adoption rate of guidance surfaced through the assistant.
- This use case accounted for a quarter of all guidance adoption across these types of conversations.