Why Taobao's Search Ranking Stopped Optimizing for Prediction: A Causal Rebuild That Lifted GMV 0.36%

The 0.36% that moved in a 41-day A/B test

Alibaba ranking model gains 0.36% GMV in 41-day A/B test

Seven researchers at Taobao & Tmall Group of Alibaba posted a paper to arXiv on August 26, 2026 describing a ranking layer that learns which conversion signals to weight for each shopper, and reported a 0.36% gain in gross merchandise value against the conventional formula across a 41-day online experiment.

ppc.land

What decides the order of products is a fusion formula that bundles several predicted values into a single score. Alibaba's search team has taken a direct run at the question of how the contents of that formula should be decided.

On August 26, 2026, seven researchers at Taobao & Tmall Group posted DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search to arXiv. Across a 41-day online A/B test, GMV rose 0.36%, click count rose 0.36% and purchase count rose 0.12%. The control arm was the GMV proxy long used in practice: predicted click-through rate multiplied by predicted conversion rate multiplied by expected transaction value conditional on purchase.

The number looks modest on its own. But this is a relative gain against a production ranking tuned over years, not against no ranking at all. At marketplace volumes, 0.36% is not a small absolute figure. That said, Alibaba has not disclosed platform-level GMV for Taobao and Tmall since 2020, so what 0.36% amounts to in currency terms remains undisclosed.

Predicting well is not the same as moving the needle

The paper starts from a structural mismatch inside e-commerce search.

Ranking assigns a score to each product within a single request. The objective the business watches, by contrast, accumulates across one person over several days: cumulative purchases or cumulative GMV. The paper calls this the granularity gap. Using the user-level figure directly as a training label makes every item shown to the same person inherit the same label, so there is no way to tell which item did the work.

Industry practice has bridged that gap with multi-objective fusion. Predictions from upstream models for click, cart addition, purchase and transaction value are combined into one ranking score, and the parameters are tuned by hand through repeated A/B tests. The paper names two limits to this. A small set of globally shared parameters offers limited personalization, and repeated online experiments are costly and slow.

Here is the heart of it. The authors argue that a predictive association is one thing, and moving the ultimate objective by pushing that metric up is another. Purchase intent leaks into every stage of the funnel. A signal that correlates well with high-value users may simply be identifying them. Identifying and changing behavior are different capabilities.

The reference cited for that distinction is Tyler VanderWeele's 2013 work on surrogate measures, published in Biometrics. A surrogate measure is an easily observed metric used in place of the outcome you actually care about. Medical statistics has long known the surrogate paradox, in which a treatment improves the surrogate while the clinical outcome worsens. The proposition that a good predictor need not be a good intervention target holds just as well in e-commerce search.

The alternative the authors adopt is the relative causal effect. Item-level proxy scores are aggregated into a user-level proxy metric, and the question becomes how much the ultimate objective improves under a given relative increase in that metric. The learning problem is reframed as finding the item-level score that maximizes it.

The aggregation design is more careful than it looks. Both daily action count and daily action rate depend on impression count, and ranking itself moves impression count. The paper therefore fixes impressions at a reference value of 100 and defines the proxy metric as the action rate at that fixed count.

Only the actor reaches production

DCEO uses an actor-critic structure, the reinforcement learning design that splits the side choosing actions (the actor) from the side judging them (the critic).

The actor takes user features and request features and emits non-negative weights that sum to one. Those weights mix the upstream predicted scores into an item-level proxy score. Because the weights are generated per request, two shoppers issuing the same query receive different blends of click probability and high-value purchase probability.

The critic's job is counterfactual estimation. It produces the difference between its four-day GMV prediction at the current proxy metric and its prediction under a hypothetical 5% increase. The actor is trained to maximize that difference, which is the causal effect loss. Because the causal effect loss alone makes optimization unstable, a conditional normalized ranking loss is added as a stabiliser, with a coefficient of 0.3 in the final configuration.

At serving time only the actor runs. Its output enters the existing fusion formula as one additional logarithmic term, leaving the original score components and weights untouched. Independent analyst Eric Seufert noted this in his LinkedIn commentary, contrasting a layer that sits on top of an existing prediction stack with deep learning recommenders that demand an end-to-end rebuild.

How the weights landed across 17 scores

The final configuration draws on 17 upstream scores: nine starting at an impression (click probability, cart-addition probability, purchase probability, generated GMV, and purchase probability above transaction values of 10, 30, 100, 300 and 1,000), seven starting at a click, and one starting at a cart addition.

The largest learned weights are as follows.

The allocation reads as counterintuitive. If GMV is the target, weight might be expected to concentrate on purchase probability or transaction value predictions, yet the largest weight went to impression-to-click probability. Second place goes to purchase probability above a transaction value of 1,000, while plain purchase probability after an impression, by contrast, takes just 0.032. These weights also carry standard deviations from 0.046 to 0.115 across impressions, meaning the blend shifts with context rather than staying fixed.

The loss ablation is clean. Optimization based on predictive correlation reached 0.022, replacing it with the causal effect loss reached 0.031, and adding the ranking loss produced the final 0.053, which the paper describes as a 2.41 times improvement.

The score-set comparison is the more telling one.

The configuration using only a score that predicts GMV directly was the weakest of the five compared. Predicting exactly what you want is not the optimum. That is the empirical substance of the paper.

Change the objective and the shelf changes

The experiment that swapped only the ultimate objective, holding everything else constant, left the result that matters most to merchants.

Optimising four-day click count sends 0.999 of the weight to impression-to-click probability alone, with effectively nothing elsewhere. Optimising four-day purchase count drops the click weight to 0.531 and moves weight onto click-to-purchase and purchase above a transaction value of 10. Optimising four-day GMV spreads weight across the transaction-value threshold scores.

DCEO does not learn a fixed fusion rule. It changes the very composition of the proxy metric according to the objective it is handed. Put the other way round, the same catalog is ordered differently depending on an objective function chosen inside the platform and never published.

The paper's own ethics section addresses this plainly, stating that objective selection requires human oversight and consideration of affected stakeholders, and that the system may change the exposure of different items and merchants.

The caveats the authors placed themselves

Read as a vendor's claim, this paper is unusually self-critical. At least three items in its limitations section sharply narrow how far the numbers travel.

The relative causal effect is a model-based local estimate from a critic trained on observational logs. A causal reading requires user features to capture the major confounders between proxy metric and ultimate objective, sufficient data support at the increased level, and accurate critic predictions inside that local region. The authors state that unobserved confounding may bias causal identification, and that the small 5% intervention magnitude reduces extrapolation without eliminating that bias. As future work they propose randomly varying the boost strength to collect genuinely interventional data.

The A/B test does not rescue this. The paper states that the online experiment validates the end-to-end effectiveness of the learned proxy score but does not directly validate the numerical estimate. The 0.053 and the 0.36% measure different things.

The evaluation also covers one search system and horizons of up to four days. The paper itself calls for validation on additional platforms, objectives and longer horizons. It is worth adding that the result holds on Taobao's particular traffic scale and particular user behavior. The paper does not name the platform; the identification with Taobao and Tmall rests on the authors' affiliations. Whether an internal A/B result reproduces on a mid-sized store is something this paper cannot speak to.

What merchants should take away

This is a story about Alibaba's search team, but the implications sit wider.

The most common move among merchants trying to optimize AI-driven exposure and recommendation is to find a metric that correlates well with sales and chase it. Mention counts in AI search exposure, citation counts, referring domains. Any of these may correlate with revenue. But by DCEO's framing, that correlation may amount to the metric identifying customers who were already going to buy. Nothing guarantees that revenue moves proportionally when budget goes into lifting the metric.

The second takeaway concerns the risk of investing in metrics you cannot validate through an A/B test. The authors were in a position to run online experiments and still placed caveats on the causal reading of their offline estimate. Most merchants cannot even construct a control group for AI search exposure. The way generative AI obscures where a purchase journey began makes the measurement problem heavier still. For proxy metrics that cannot be validated, the minimum discipline is to ask whether the metric is genuinely an intervenable variable.

Parallel arguments are surfacing elsewhere. A Zalando researcher's August 2026 paper on marketing mix models showed a standard specification reporting 10.61 times return on ad spend against a true value of 4.20 times, though the caveat that it is a single-author preprint on synthetic data, not yet peer reviewed, applies. Correlational fitting on observational data breaks on contact with intervention. DCEO takes that argument down one layer, from budget allocation into the ranking function itself.

For sellers, the practical starting point is the fact that a change in objective changes the weight allocation. What needs work on a product page differs between a phase weighted toward clickability and a phase weighted toward high-value purchase probability. Platforms will not disclose which objective they optimize for, but a merchant can still separate information work that holds under either regime from tactics that only pay off under one.

Conclusion

The figure that will circulate is 0.36%. The one that changes how ranking is understood is 0.999. Hand the model click count as its objective, and it puts nearly all the weight on a single click prediction. Order is downstream of a decision that is never published.

What to watch next is whether the randomized boost strength the authors listed as future work is actually implemented. Interventional data, rather than estimates from observational logs, would raise the verifiability of claims like this considerably. And when the same idea moves from search ranking into the recommendation order of AI assistants, merchants will have one more invisible objective function to contend with.