HARBORLINEmarketplace
/
Sellers

Live catalog · updated 4 min ago

Everything worth hunting for.

Independent sellers, honest condition grades, and a search bar that admits when it found nothing.

Listings48
Sellers12
Searches today0
Zero-result0.0%
48 listings 

  








```

**`claude_search_analytics.html`**

```html





Harborline — Search & Discovery Analytics









  
H
HARBORLINE ANALYTICSsearch · discovery · conversion

  
 postgres://harborline_prod · schema discovery · 2,614 search_log rows

  

    

    
  






  


  
  

    
Q1

      
Most searched terms — and the searches that returned nothing

      
Every query is normalized (lowercase, trimmed, stem-mapped) before it lands in search_log. Zero-result rows are the single cheapest source of catalog and synonym backlog.

      

    

    

  


  
  

    
Q2

      
Conversion rate by category

      
Session-level funnel per top-level category: search → click → add-to-cart → purchase. Category is attributed from the clicked/purchased product, falling back to the query's predicted category.

      

    

    

  


  
  

    
Q3

      
Top performing sellers — revenue vs. reputation

      
Revenue is SUM(order_items.unit_price * qty); reputation is rating volume + refund rate. The gap between the two leaders is where merchandising effort goes.

      

    

    

  


  
  

    
Q4

      
Search-to-purchase funnel

      
Distinct sessions, not rows — a session with three searches and one purchase is one session at every stage. Stage 2 drop is pure result-quality loss.

      

    

    

  


  
  

    
Q5

      
Average time from first search to purchase

      
Computed with a window function: MIN(occurred_at) OVER (PARTITION BY session_id) anchored to each purchase event. Median beats mean here — the distribution is right-skewed.

      

    

    

  


  
  

    
DDL

      
Schema, normalization & indexes behind the search bar

      
3NF with junction tables for tags; the search log is deliberately denormalized (star-ish) so analytics never joins against the hot catalog path. Term is stored both raw and normalized so synonyms can be backfilled without re-running the collector.

    

  


  
  

    
QED
What the numbers say to do next

      
Generated from the computed result sets above — not hardcoded.

    

  


  
Harborline Marketplace · claude_search_analytics.html · dataset = seeded deterministic session model (mulberry32, seed 0x5EED) mirroring the event schema written live by claude_marketplace.html · all metrics recomputed client-side from the same SQL semantics shown in each panel.