Correct the solicitation finding: recomputed on the published 79-product subset, where the within-product effect shrinks by a third rather than vanishing

README.md

@@ -58,13 +58,22 @@ The largest public G2 review corpus by a wide margin. Before this, the biggest a
 
 ## A result worth knowing before you model this
 
-Uncontrolled, solicited reviews look clearly worse than organic ones: Organic averages 4.35 stars, Seller invite 4.20, G2 invite 3.92.
+G2 invite reviews average **3.61 stars** against **4.35** for organic ones, a gap of nearly three quarters of a star. Seller invite sits at 4.10.
 
-Compare **within the same product** and the effect disappears. Seller invite comes out at +0.085 stars, higher in only 63 of 111 products. G2 invite at +0.046, higher in 45 of 87. Verification method is flatter still.
+Products do not solicit reviews at random, though, so part of that gap could simply be which products run invite campaigns. Comparing each channel against organic reviews **of the same product** tests that:
 
-So the apparent "invited reviews are worse" pattern is **a product-selection artifact**, not a rating effect. Products that run invite campaigns are simply different products. Any analysis of solicitation bias on review data has to group by product or it will find an effect that is not there.
+| Channel | Naive mean | Within-product lift | Products |
+|---|---|---|---|
+| Organic | 4.353 | baseline | |
+| G2 invite | 3.612 | **-0.499** | higher in 6 of 18 |
+| Seller invite | 4.099 | -0.151 | higher in 10 of 27 |
+| G2 invite on behalf of seller | 4.169 | -0.043 | higher in 5 of 12 |
 
-That is a caution rather than a headline, and it is why `product_slug` matters as much as the text columns.
+Controlling for product **shrinks the gap by about a third but does not remove it**. G2-invited reviews still run roughly half a star below organic reviews of the same product, and are higher in only 6 of 18 products.
+
+Two honest caveats. Only 18 to 27 products carry 30 or more reviews on both sides, so the magnitude is unsettled. And the same comparison run across the full 6,120-product cache this was drawn from gives a much weaker effect, which suggests the pattern is specific to heavily-reviewed products rather than universal.
+
+Either way, `product_slug` matters as much as the text columns: the uncontrolled comparison overstates the effect by a third.
 
 ## Why this is useful for ML