Generative retrieval for search: What a study shows

Search is where most e-commerce demand is won or lost, and a preprint posted to arXiv on 13 September 2026 takes aim at an unglamorous part of it: how the engine decides which products to even consider before it ranks them. The paper, VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search, asks whether that first retrieval step can be taught to weigh business value, not only relevance, without breaking the ranking that follows.

The question the paper asks

An e-commerce search engine usually works in two stages. First it retrieves a broad pool of candidate products that look relevant to the query; then a separate ranker orders that pool into the list a shopper sees. Business value — margin, stock position, promotional priority — has traditionally been applied only at the ranking stage. The authors ask whether value can be built into the first stage, so the engine generates a better pool to begin with, and whether doing so helps or hurts overall results.

What they did

VARG replaces the relevance-only retrieval step with a generative model that produces candidate items directly and hands them to the existing final ranker. In plain terms, instead of looking products up in an index, the model is trained to generate the identifiers of products that fit the query. To make that possible it builds compact semantic identifiers for every product (using a technique called an RQ-VAE, which compresses each item into a short sequence of codes), and it learns to match queries and items in both directions. The novel part is what the authors call value-ordered tokens: within a group of similar products, the model is trained to generate higher-value items first. Training runs in three supervised stages, followed by a reinforcement-learning step the authors name Prefix-GRPO that aligns what the model generates with what the ranker actually rewards.

What they found

Offline, on a catalogue of tens of millions of products, the method improved both retrieval quality and what the authors call head-level value recall — its ability to surface high-value items rather than merely relevant ones. The more meaningful test is the online one: a 14-day A/B test on 20% of a single platform's live search traffic. There, gross merchandise value rose 1.45%, page click-through 0.31%, and items viewed per user 0.22%. The direction is consistent across the three measures, and the GMV figure is the one that matters: a 1.45% lift on a large platform's search channel is a substantial sum in absolute terms, even though the percentage looks modest.

What it means operationally for an Australian retailer

The model is out of reach for most retailers; the idea behind it is not. The useful takeaway is that business value belongs in the retrieval step, not only in a re-rank at the end. Very few Australian retailers will build generative retrieval in-house, but almost all of them buy search from a vendor, and this is a question worth putting to that vendor: where in the pipeline is value applied? If your platform optimises purely for relevance and only layers margin, availability and promotions on at the final re-rank, VARG is evidence that moving some of that logic earlier can lift sales. The risk to watch for is the obvious one — an engine tuned toward value can quietly drift away from relevance and erode trust — so the same test the authors used, a controlled A/B on live traffic with GMV and engagement measured together, is the right way to check it locally.

How much to trust it

Carefully. This is a preprint, not a peer-reviewed paper, so it has not been through independent review. The online result comes from a single platform's own A/B test, run by the team that built the system; that is the standard industrial setup, but it is also one setting and a clear conflict of interest, and the result has not been replicated elsewhere. The lifts, while valuable at scale, are small, and the paper offers little public discussion of failure cases or limits. Nothing here was tested on Australian catalogues or shopper behaviour, so any read-across to the local market is extrapolation, not evidence. Treat VARG as a promising direction and a good prompt for a vendor conversation, not as a proven playbook.

Full citation: Chu, X., Zhu, J., Jin, M., Wang, J., Fang, X. and Zhang, W. (2026). VARG: Value-Aware and Ranking-Aligned Generative Retrieval for Dynamic E-commerce Search. arXiv:2609.14493 (preprint).