Prompt
We’re not getting good results from Google Cloud Recommendations AI
Latest observation
If Google Cloud Recommendations AI isn’t producing good results, the issue is usually one of these: weak/insufficient event data, poor catalog quality, the wrong recommendation use case, or not enough traffic for the model to learn.
Here’s a practical checklist to improve it:
1) Check data quality first
Recommendations AI depends heavily on user-event data.
- Are you sending enough events?
- Views, clicks, add-to-cart, purchases, search events, etc.
- If you only send purchases, training data may be too sparse.
- Are events correct and complete?
- Proper user/session IDs
- Correct item IDs matching the catalog exactly
- Accurate timestamps
- No duplicate or bot traffic
- Do you have enough volume?
- Small catalogs or low traffic often produce weak recommendations.
2) Improve the product catalog
Bad catalog data leads to bad recommendations.
- Make sure item titles, categories, brands, prices, and availability are accurate
- Use rich metadata:
- category hierarchy
- labels/tags
- color, size, style, audience, etc. if relevant
- Remove out-of-stock or discontinued items
- Ensure item IDs are stable and consistent
3) Confirm you’re using the right model/recommendation type
Different use cases need different setups:
- Frequently Bought Together
- Similar Items
- Recommended for You
- Trending
If the use case doesn’t match the user experience, results may seem poor even if the model is working as designed.
4) Make sure the model has enough time and training data
Recommendations improve only after sufficient learning.
- New catalogs or newly launched stores often need time
- Don’t judge too early
- Check model training status and whether it has enough recent data
5) Examine filtering and business rules
Too many constraints can make recommendations look bad.
- Overly restrictive filters can eliminate good candidates
- Verify category, inventory, price, and locale filters
- Check whether personalization is being unintentionally blocked
6) Test by placement
The same model can perform differently depending on where it’s shown.
- Product detail page
- Cart page
- Home page
- Search results page
For example, “similar items” works better on PDPs, while “frequently bought together” works better in cart/checkout contexts.
7) Use offline and online evaluation
Don’t rely only on subjective judgment.
- Review:
- CTR
- conversion rate
- add-to-cart rate
- revenue per session
- Compare against a baseline such as popularity-based or rule-based recommendations
- A/B test changes instead of guessing
8) Add fallback logic
If Recommendations AI has low confidence or sparse data:
- show trending/popular items
- show category-based recommendations
- show hand-curated items for key placements
9) Common integration mistakes
These are frequent causes of bad results:
- item IDs in events don’t match catalog item IDs
- user IDs reset too often
- event timestamps are incorrect or out of order
- missing purchase events
- training data not updated frequently enough
- using stale catalog data
10) Consider hybrid approaches
Recommendations AI may work better when combined with:
- search/query signals
- rules-based merchandising
- popularity/trending signals
- contextual signals like seasonality, location, or device type
If you want, I can help you troubleshoot this systematically. If you share:
- your use case,
- what event data you send,
- catalog size,
- and an example of “bad” recommendations,
I can suggest the most likely fixes.