What AliExpress's search experiment shows AU retailers

Shoppers and product catalogues do not use the same words. A customer types "cosy jumper for winter"; the catalogue says "knitted pullover, acrylic blend". Search engineers call this vocabulary mismatch, and it is one of the oldest unsolved problems in ecommerce search. A paper posted to arXiv on 4 September 2026, SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce, asks whether a model that looks at the product photograph as well as the title can write better substitute queries, and whether training that model against commercial outcomes rather than semantic similarity produces a measurable lift in a live store.

What they did

The technique being extended is called Doc2Query. Instead of waiting for a shopper's words to match the seller's words, the system generates the queries a product ought to match and stores them alongside the product, so the index contains the language customers actually use. SAM-D2Q adds two things to that idea. First, it generates from the image as well as the text, using a vision-language model fine-tuned on ecommerce data and then trained on augmented examples designed to capture attributes visible in the photograph but missing from the title. Second, it applies reinforcement learning to align the generated queries with search demand and conversion, rather than with semantic similarity alone, so the model is rewarded for producing queries people actually search and buy from.

The paper reports two kinds of evaluation. Offline, it compares retrieval performance against conventional Doc2Query approaches. Online, it was tested in the production search system of AliExpress, the cross-border marketplace, in a live A/B test.

What they found

The online result is the number worth carrying: gross merchandise value increased 3.38 per cent and pay count increased 2.27 per cent against the control. Offline, the authors report substantially improved retrieval performance compared with traditional Doc2Query methods; the abstract does not put a figure on the offline gain, and the specific benchmark metrics are not stated there.

Two things are worth noticing about the shape of that result. GMV moved more than pay count, which means average order value rose as well as order count, though the paper as summarised here does not decompose it. And the lift is a single-digit percentage on an existing, heavily optimised search system, which is what a credible production result in this area looks like. Search improvements that claim double-digit GMV lifts in a mature marketplace are usually measuring something other than what they say.

What it means operationally

The practical reading for an Australian retailer is about where this technique pays. Document expansion earns its keep when the gap between customer language and catalogue language is wide, which is a function of who writes the product copy. A marketplace or a long-tail catalogue fed by supplier data has enormous mismatch: titles are keyword-stuffed, written for a different market, sometimes machine-translated. A curated catalogue where a merchandising team writes every description has much less. If a retailer's product copy is already written in customer language, the headroom this paper is measuring may not exist in their store.

The multimodal part is the more portable idea, and it points at apparel, homewares and anything where the photograph carries attributes the title omits: colour family, pattern, neckline, sleeve length, leg shape, finish, the room a piece of furniture belongs in. Those are exactly the terms shoppers search and exactly the terms missing from a supplier-supplied title. A retailer does not need to reproduce this paper to act on it; the question to put to a search vendor is whether the index is built only from text, and if so what the plan is for the attributes that only exist in the image.

The reinforcement learning element deserves a guardrail. Training generated queries against conversion means the system will learn to surface what sells, which is not always what was asked for. Any retailer trying this should measure relevance separately from revenue, because a search engine optimised purely for GMV has an incentive to drift towards popular products regardless of query intent, and that damage shows up slowly in repeat visits rather than quickly in a test.

How much to trust it

The online test is the strongest part of the evidence. A live A/B test in production supports a causal claim about the effect of the change in that system, which is more than most published work in this area can offer. It does not support a claim about the size of the effect anywhere else.

The limitations are real and should be stated. The paper is a preprint, version 1, posted on 4 September 2026; its comments field records acceptance by CIKM 2026 as an oral full paper, which means it has passed conference peer review, not that the result has been replicated by anyone else. The evaluation was conducted by the operator of the platform whose GMV improved, which is normal for industrial search research and is still a conflict worth naming. The abstract does not state the duration of the A/B test, the traffic volume, confidence intervals, or the offline benchmark figures, so the precision of the 3.38 per cent cannot be assessed from the abstract alone.

The setting matters most. AliExpress is a cross-border marketplace with third-party sellers and highly variable product data, and that is the environment where vocabulary mismatch is worst and document expansion helps most. Applying the result to an Australian retailer with a curated catalogue would be extrapolation, and it is flagged here as extrapolation. The honest summary is that this paper shows a specific technique produced a small, real, causally attributable lift in one large marketplace, and that the mechanism it exploits is one most Australian retailers have not tested for in their own catalogues.

Citation

Zhou, H., Ji, J. H., Ma, L., Xiao, R. and Zeng, X. (2026). SAM-D2Q: Aligning Multimodal Doc2Query with Search Demand and Conversion for E-commerce. arXiv preprint arXiv:2609.04961 [cs.IR], 4 September 2026. Accepted by CIKM 2026 as an oral full paper. https://arxiv.org/abs/2609.04961