Key Concepts
- 1Recommendation in an agent flow is not one algorithm but a policy: when to explore (show variety) vs exploit (double down on the best-fit item)
- 2Constraint-satisfaction first: filter to what's actually valid (budget, stock, size), then rank the survivors by fit — never rank first and apologize about constraints later
- 3Diversity matters: three near-identical black jackets is a worse result set than three genuinely different options spanning price and style, even if all match intent
- 4Explanation is part of the recommendation: 'this one because it's within budget and ships by Friday' converts better than a bare list, because it reduces the user's evaluation effort
- 5Cold-start (no history) leans on catalog signals and stated intent; warm users blend in preference memory — the agent must degrade gracefully when data is thin
- 6Avoid the popularity trap: purely best-seller-driven recommendations erase the personalization advantage that justifies an agent in the first place
- 7Set size is a UX decision: too few options feels limiting, too many reintroduces the paradox of choice the agent was supposed to solve — three to five is a common sweet spot
Why It Matters for PMs
Directly drives AOV and conversion: The ranking and framing of options is the moment of truth. Good strategy lifts basket size; poor strategy sends users back to manual browsing.
Personalization is the moat: For SMBs, tailored recommendations are the whole value proposition. Defaulting to best-sellers throws away the differentiation an agent is supposed to provide.
Explainability compounds trust: Recommendations paired with clear reasons reduce returns and second-guessing, improving both margin and CLV over time.
Test Your Understanding
Question 1 (Multiple Choice)
What should happen BEFORE ranking products by fit in an agent recommendation flow?
Question 2 (Multiple Choice)
Why is diversity in a result set often better than three near-identical matches?
Question 3 (Multiple Choice)
Relying purely on best-sellers for recommendations primarily undermines...
Question 4 (Short Answer)
How would you balance exploration (variety) vs exploitation (best single match) in an agent's recommendations?
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
Why might pairing each recommendation with a one-line reason increase conversion and reduce returns?
Think about your answer before revealing the explanation.