Can a new product inherit a similar product's recommendations? We tested it

It did not work very well. The idea recovered roughly an eighth of the distance between the fallback most shops use today and the recommendations a product earns from its own history — and the interesting part is why, because the obvious explanation turns out to be wrong.

Designing the test around a shortage of cold products

The honest problem with evaluating a cold-start fix is that cold products are, by definition, the ones nobody interacts with. In our test window only 34 sessions began with a product that had no history at all — far too few to separate a good idea from a bad one.

So we ran a simulated cold start. Take products that do have history, hide their own co-view data, serve them their nearest neighbour's recommendations instead, and score on the same held-out sessions. That gives 2,453 sessions of evidence, a clean upper bound (what the product's own history achieves) and a clean lower bound (the category fallback). It is the standard way to get statistical power out of a cold-start question, and it is slightly generous to the method, since these products are better connected than a genuinely new line would be.

The results

What the shopper was shown	Opened one of the ten	Share of the gap closed
The product's own history	51.1%	— the ceiling
Inherited from its nearest neighbour	35.6%	12%
The category's top sellers	33.4%	— the floor

Inheritance beat the category fallback by 2.2 percentage points. On 2,453 sessions that is outside chance, but it is small: it closes about an eighth of the 17.7-point gap between the floor and the ceiling.

The matching was not the problem

The obvious diagnosis is that our similarity measure was too crude — shared words in the product name plus category, rather than a language model's embeddings. We expected to write that conclusion. Then we looked at the pairs it actually chose:

Product	Nearest neighbour it borrowed from
Google Crew Socks	YouTube Crew Socks
Google Camp Mug Ivory	Google Camp Mug Gray
Google Zip Hoodie F/C	Google Youth FC Zip Hoodie
YouTube Icon Tee Grey	YouTube Icon Tee Charcoal
Google Clear Pen 4-Pack	Google Cloud Pen
Google Land & Sea Cotton Cap	Google Land & Sea Unisex Tee

Those are the pairs a merchandiser would choose by hand. If the matching is already picking the near-identical product, a better similarity measure has very little room left to improve on it — which means embeddings would probably not rescue this either, and the assumption underneath the idea is what failed.

Why similar products don't share recommendations

A co-view list is not a list of similar products. It is a record of the paths people take through your shop — which pages follow which, driven by navigation, merchandising, search results and whatever was on the home page that week.

Two near-identical products often sit on different paths. The grey tee is reached from the tees listing; the charcoal one from a promotion; a mug from the Drinkware page and its twin from a campus collection. They look alike to a human and to an embedding model, and they are browsed in different company. That is the gap the method could not close.

What we would actually do about cold start now

• Still never render an empty slot. 36% beats 33% beats nothing. Inheritance is a marginally better fallback than category popularity and costs about the same to run, so use it — just do not expect it to solve the problem.
• Put the weight on earning real data instead. The gap that matters is the one between any fallback and the product's own history. The way to close it is to give new products a guaranteed share of impressions for a fixed window so they accumulate their own co-view data — the unglamorous step we listed third in part 3, which this result promotes to first.
• Measure the time-to-history, not just the quality of the fallback. How long does a new product take to reach the point where its own data beats the fallback? That number tells you what the impression budget needs to be.

Under the hood — for the technically curious

Similarity is Jaccard overlap on product-name tokens, after dropping words carried by more than 30% of the catalogue (“google” is not a distinguishing feature here), with ties broken by category match and then price proximity. A product may only borrow from a lender with at least 20 sessions of history:

neighbour AS (
  SELECT item_a AS item_key, item_b AS lender
  FROM similarity s
  JOIN item_n ON item_n.item_key = s.item_b
  WHERE item_n.n >= 20                      -- the lender must have something to lend
  QUALIFY ROW_NUMBER() OVER (
    PARTITION BY item_a
    ORDER BY jaccard DESC, same_cat DESC, price_gap ASC) = 1
)

Two limits worth stating. We used text similarity rather than embeddings because no Vertex AI connection exists on this warehouse yet; given that the matching already produces near-identical pairs, we do not expect embeddings to change the conclusion, but we have not proven that. And 3.4% of test sessions had no eligible neighbour at all, which a genuinely new product in a thin category would hit more often.

Data and licence

bigquery-public-data.ga4_obfuscated_sample_ecommerce — the obfuscated GA4 export of the Google Merchandise Store, published by Google through the Cloud Public Datasets Program. Trained on data to 10 January 2021, tested on sessions from 11 to 31 January; crawler traffic excluded throughout, for the reasons in part 1.