I’m a solo founder working on an early-stage experiment called Pearch.
At a high level, it’s a Chrome extension that surfaces product recommendations while someone is browsing online, but the part I’m most interested in right now is signals.
The problem I’m exploring
Most recommendation systems I’ve worked with or studied lean heavily on one of two things:
- Browsing behavior (clicks, views, dwell time)
- Similarity signals (category, visual similarity, embeddings)
What I’ve been questioning lately is whether historic purchase behavior might be a stronger anchor for relevance than either of those alone, especially when combined with real-time browsing context.
In other words:
What if we treated what someone has actually bought as the primary signal, and everything else as supporting evidence?
Why this feels interesting (and risky)
Purchase data is:
- Sparse
- Delayed
- Messy across retailers
But it’s also the clearest expression of intent we have.
I’m trying to understand:
- Does anchoring recommendations on purchase history meaningfully improve relevance?
- Where does this break down at small scale?
- At what point does recency matter more than history?
- How do you avoid overfitting someone to who they were versus who they’re becoming?
What I’m not doing
- I’m not selling anything.
- I’m not claiming this is the right approach.
- I’m not optimizing for growth yet.
This is still very much an exploration of signal quality and system design, not a polished product.
What I’d love feedback on
If you’ve worked on recommendation
systems, personalization, or ecommerce tooling:
- What signals ended up being more valuable than you expected?
- What signals looked promising but failed in practice?
- How do you think about balancing long-term behavior vs in-session intent?
- Are there obvious pitfalls I should be pressure-testing earlier?
Happy to learn from anyone who’s been down this path before. Even strong skepticism is useful here.
Thanks for reading.
Top comments (0)