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.

In short

Alibaba researchers built a system that decides, for every search request, how much a click prediction counts against a purchase prediction when ordering products. Instead of tuning those weights by hand, the system learns them by asking which combination actually moves four-day sales rather than which combination best predicts them. In a live test the version increased gross merchandise value by 0.36%, clicks by 0.36% and purchases by 0.12%, and the authors state plainly that the offline estimate behind it rests on assumptions their data cannot verify.

The gap the paper is trying to close

Search ranking scores items. Businesses count users.

That mismatch is the whole problem the paper sets out. A ranking system assigns a number to each candidate product within a single request and sorts by it. The commercial objective, by contrast, accumulates across a person over days: cumulative purchases, or cumulative gross merchandise value over an n-day window. The paper calls this the granularity gap, and it explains why the user-level figure cannot serve directly as a training label. All items shown to the same shopper would inherit the same label despite contributing differently to the outcome.

Industrial systems bridge the gap with multi-objective fusion. Upstream models predict click, cart addition, purchase and transaction value at item level; a fusion module folds those predictions into one ranking score. According to the paper, practitioners then improve that score by adding signals or tuning fusion parameters through repeated online A/B tests, a process with two named limitations: a small set of globally shared parameters gives limited personalisation across users and requests, and repeated online experiments are costly and slow.

The conventional shortcut in this setting is a predicted GMV proxy, written in the paper as the product of predicted click-through rate, predicted conversion rate and expected transaction value conditional on purchase. That is the control arm of the experiment described later.

Why predicting is not the same as moving

The paper's central objection is stated early and is a familiar one to anyone who has argued about measurement. According to the authors, predictive association does not imply that increasing the proxy metric through ranking will produce a large improvement in the ultimate objective. Purchase intent leaks into every stage of the funnel, so a signal that correlates with high-value users may simply be identifying them rather than changing what they do.

This is endogeneity in the ranking layer rather than in the media plan. The paper cites Tyler VanderWeele's 2013 work on surrogate measures as the reference for why a good predictor need not be a good intervention target.

The alternative the authors adopt is to define alignment as a relative causal effect. They aggregate item-level proxy scores into a user-level proxy metric, then ask how much the ultimate objective improves under a given relative increase in that metric. A larger positive value indicates better alignment. The learning objective becomes finding an item-level score that maximises it.

The choice of what to aggregate is more careful than it first appears. Neither daily action count nor daily action rate works as a proxy metric, because both depend on impression count, which ranking itself affects. The paper gives a worked case: deboosting items with high predicted dislike scores may lower the dislike rate while raising user activity, pushing impressions up, so the daily dislike count can rise even as the rate falls. The authors therefore fix the impression count, using a reference value of 100, and define the proxy metric as the fixed-impression action rate.

Actor, critic, and two losses

DCEO is an actor-critic architecture, and only half of it ever reaches production.

The actor takes user features and request features and emits a set of non-negative weights that sum to one. Those weights form a convex combination of selected upstream predicted scores, producing an item-level proxy score. Because the weights are generated per request, two shoppers issuing the same query can receive different weightings of click probability against high-value purchase probability.

Those item-level scores are averaged across a user's impressions on a reference day, then passed through a calibration model that maps the average at the observed impression count to its counterpart at the reference count of 100. The calibration model is trained with mean squared error and its gradients are blocked during backpropagation.

The critic estimates the ultimate objective from user features and the calibrated proxy metric, also under mean squared error. Its role is counterfactual: the actor is trained to maximise the difference between the critic's prediction at the current proxy metric and its prediction under a hypothetical 5% increase. That difference is the causal effect loss. The critic's parameters are frozen while it is computed.

The authors optimise the unnormalised difference rather than the ratio itself, because backpropagating through the denominator would make gradients sensitive to fluctuations in a baseline response that varies little across training batches anyway.

The causal effect loss alone produces what the paper describes as a highly non-convex optimisation landscape. A second term, the conditional normalised ranking loss, acts as a stabiliser. It standardises the proxy metric and the ultimate objective into a common distribution using a model that predicts their conditional means and log standard deviations, then applies a Bradley-Terry pairwise loss over ordered user pairs. The two losses are combined with a coefficient of 0.3 in the final configuration. All four models are multilayer perceptrons over feature embeddings, trained jointly.

At serving time, only the actor runs. Its output enters the existing fusion formula as one additional logarithmic term, with the original weights and score components left untouched.

Seventeen scores, and which ones the model actually uses

The final configuration draws on 17 upstream predicted scores spanning two entry points. Nine begin at an impression: click probability, cart-addition probability, purchase probability, generated GMV, and purchase probability above transaction-value thresholds of 10, 30, 100, 300 and 1,000. Seven begin at a click, covering purchase probability, generated GMV and the same five thresholds. One begins at a cart addition. A transformation maps every score into the interval from zero to one while holding its mean near 0.1, so that average weights stay comparable.

Offline training uses 14 consecutive days of search logs with evaluation on the following day. Unless stated otherwise, the ultimate objective is cumulative GMV over four days from the reference day.

Under that configuration, the relative causal effect reaches 0.053, meaning the critic estimates a 5.3% relative increase in four-day GMV under a 5% increase in the proxy metric. The weight distribution is uneven. Impression-to-click probability takes the largest mean weight at 0.404. Click-to-purchase-above-1,000 takes 0.205, click-to-purchase-above-10 takes 0.154, click-to-purchase-above-100 takes 0.105 and click-to-purchase-above-30 takes 0.079. Direct purchase probability after an impression receives 0.032.

Every one of those six scores carries a non-zero standard deviation across impressions, which the authors read as evidence that the model learns context-dependent weights rather than a single fixed combination. The standard deviations range from 0.046 to 0.115.

The ablations matter more than the headline. Replacing predictive association optimisation with the causal effect loss raised the metric from 0.022 to 0.031. Adding the ranking loss took it to 0.053, which the paper describes as a 2.41 times improvement over predictive association optimisation. The trade-off coefficient behaves as a regulariser rather than a free parameter: 0.1 reached 0.052, 0.3 reached 0.053, and 1.0 fell back to 0.047. The ranking loss used alone reached 0.048, higher than the causal effect loss alone.

The score-set ablation is the one with the clearest read-across. Using impression-to-GMV alone produced 0.027. A basic impression-level set of four scores reached 0.039. A value-aware set reached 0.041 and a conversion-funnel set 0.040. The full 17-score set reached 0.053. Predicting GMV directly, in other words, was the weakest of the five configurations tested.

The proxy changes shape when the objective changes

Four different ultimate objectives were trained under otherwise identical settings, and the resulting weights diverge sharply.

Optimising four-day click count sends 0.999 of the weight to impression-to-click and effectively nothing anywhere else. Optimising four-day purchase count keeps a large click weight at 0.531 but moves 0.101 to click-to-purchase and 0.256 to click-to-purchase-above-10. Optimising four-day GMV spreads weight across the transaction-value thresholds, including 0.205 on the highest band. A composite objective of purchase count plus one tenth of GMV lands in a related but distinct allocation, with zero weight on plain click-to-purchase.

According to the paper, these changes show that DCEO does not learn a fixed fusion rule, and that its proxy composition responds to the specified user-level objective.

That is a finding with consequences well beyond one search engine. The composition of the ranking signal is downstream of an objective someone chose. The paper's own ethics section makes the point without hedging, stating that objective selection requires human oversight and consideration of affected stakeholders, and noting that the system may change the exposure of different items and merchants.

The horizon result, and a wrinkle in it

Holding the objective to GMV and varying the accumulation window from one to four days produced the result that travelled fastest.

Impression-to-click weight measures 0.331 at a one-day horizon and 0.404 at four days. The paper reads a larger click weight as a preference for items likely to be clicked, which raises click count, which in turn indicates that users explore more items. Its conclusion is that DCEO encourages more exploration when optimising GMV over a longer horizon.

The intermediate values complicate the line slightly. Two-day GMV produces 0.379 and three-day GMV produces 0.417, which is higher than the four-day figure. The direction across the endpoints holds; strict monotonicity does not. The paper does not address the three-day peak.

Eric Seufert, an independent analyst, summarised the paper on LinkedIn shortly after it appeared, opening with the question of whether "a personalized fusion of funnel predictions improve search ranking". He flagged the horizon result as one of two findings he found particularly interesting, describing the authors as interpreting the shift as "preferencing exploration (vs. immediate conversion) when optimizing over a longer horizon". His second observation was that a personalised mixture across the funnel aligned better with long-term GMV than predicted GMV alone, suggesting that "no single terminal metric captures all of the behaviors that generate long-term commercial value".

Seufert also identified the deployment as sitting inside Taobao and Tmall. The paper itself does not name the platform. It describes the deployment only as a large-scale industrial e-commerce search system, and the identification rests on the authors' affiliations rather than on any statement in the document. The distinction is worth preserving, since the paper's own limitations section restricts its claims to one system.

The online test

A 41-day online A/B test compared the DCEO proxy against the conventional GMV proxy. Both were integrated into the existing fusion formula through the same mechanism at the same boost strength, which isolates the proxy rather than the integration.

GMV rose 0.36%. Click count rose 0.36%. Purchase count rose 0.12%. The authors treat GMV as the primary metric because it corresponds directly to the ultimate objective, and read the click and purchase gains as evidence that the GMV improvement did not come at the expense of either.

That figure is a relative change against an already-tuned production baseline, not against no ranking at all. At marketplace volumes, a third of a percent of GMV is a large absolute number, which is why a paper reporting sub-one-percent movement is read as a result rather than as noise.

Seufert's architectural point is the one most likely to determine whether other teams attempt it. He described the approach as attractive because "it fuses the point models in an existing prediction stack rather than displacing it with an end-to-end rebuild", contrasting it with deep learning recommender models that require exactly such a rebuild.

What the authors say the result does not establish

The limitations section is unusually direct, and four items in it constrain how far the numbers travel.

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

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

Calibrating to a fixed impression count is a deliberate exclusion rather than an oversight. It removes changes in the ultimate objective that arrive through ranking-induced changes in user activity and impression count, which means the paper frames the offline metric as an alignment measure at the reference impression count rather than an estimate of total deployment effect.

The actor is also deliberately restricted. A more general version would map impression-level features straight to a proxy score, but the authors report that the resulting function space was difficult to optimise stably under a critic-based objective. Constraining it to a convex combination of selected upstream scores buys stability and lightweight serving at the cost of expressiveness: it cannot recover information absent from those scores.

The evaluation also covers one search system and horizons of up to four days. The paper states that validation on additional platforms, objectives and longer horizons is needed to establish broader generalisability.

Why this matters for the marketing community

The paper is not about advertising, and that is precisely why it is worth reading in an advertising context.

Every automated bidding product sold to marketers performs the operation DCEO performs: it collapses several predicted intermediate outcomes into one score, then uses that score to order things. The industry has spent two years arguing about whether the resulting numbers describe causation or correlation, mostly at the level of budget allocation. Google research on ad auctions, produced with Yale academics, framed the same tension using Markov Decision Processes in a shift from cost-per-click bidding toward user lifetime value. Journey Aware Bidding extended optimisation across full customer paths using biddable and non-biddable signals. A new Google metric counts conversions up to 180 days after a click, which is the same horizon question the DCEO ablation runs at four days.

DCEO relocates the argument one layer down, into the ranking function itself, and answers it with an explicit causal target rather than a better predictor. A Zalando researcher's August 2026 paper made a parallel case for marketing mix models, showing a standard specification reporting 10.61 times return on ad spend for a channel whose true return was 4.20 times, with better controls closing only part of the gap. Both papers arrive at the same conclusion from different directions: correlational fitting on observational logs does not survive contact with intervention.

The retail media read is more immediate. Sponsored listings on a retailer's site sit in the same result set as organic ones, ordered by a blended score whose weighting is not disclosed. IAB Europe published an incrementality framework for commerce media distinguishing causal impact from correlation between sales increases and campaign timing, alongside a measurement framework requiring credible counterfactuals. What the DCEO paper adds is a demonstration that the organic ranking function on the other side of that page is being retuned against a long-horizon objective by a model that changes its weighting per shopper. Onsite placement economics rest on that function.

The objective-selection finding is the part with the sharpest edge for merchants and brands. Weights trained against click count, purchase count and GMV produce materially different proxy compositions, which means the same catalogue is ordered differently depending on a decision taken inside the platform and never published. TikTok's GMV Max drew the same criticism when its ROI Protection guarantee arrived, and the later GMV Max Pro release moved optimisation past top-line gross merchandise value to include coupons, affiliate costs and platform commissions. Amazon's Rufus layered a language model over the A10 ranking algorithm and a shopper profile layer on top of that. Algorithmic bidding systems already set prices from more than twenty signals at auction time.

None of that is new. What is new is a published, reproducible description of a marketplace deciding, per request, that a shopper's four-day value is better served by weighting click probability at 0.404 and high-value purchase probability at 0.205 than by predicting sales directly. The number that will circulate is 0.36%. The number that changes how ranking is understood is 0.999, the weight the model assigns to a single click prediction when the objective it is handed is clicks.

Timeline

Summary

Who: Junzhao Zhang, Tao Zhang, Liren Yu, Feiyi Dong, Zhixuan Zhang, Dan Ou and Haihong Tang, all listed under Taobao & Tmall Group of Alibaba, wrote the paper. Independent analyst Eric Seufert circulated it on LinkedIn. The findings concern e-commerce search teams, marketplace sellers, retail media buyers and anyone reasoning about automated bidding systems that fuse multiple predicted outcomes.

What: DCEO, or Direct Causal Effect Optimization, is an actor-critic layer that learns request-specific weights over 17 upstream predicted scores to build an item-level proxy score, aggregates it to a user-level metric calibrated at a reference count of 100 impressions, and trains the actor to maximise a critic's estimate of the change in four-day gross merchandise value under a hypothetical 5% increase in that metric. Offline, the relative causal effect reached 0.053 against 0.022 for predictive association optimisation. Online, gross merchandise value rose 0.36%, clicks 0.36% and purchases 0.12%.

When: The paper was posted to arXiv on August 26, 2026, carrying identifier 2608.25635v1. The online experiment ran for 41 days; the paper does not state its calendar dates. Offline training used 14 consecutive days of logs with evaluation on the following day.

Where: Filed in the machine learning section of arXiv. The deployment is described only as a large-scale industrial e-commerce search system; the authors' affiliations place the work at Taobao & Tmall Group of Alibaba in Hangzhou, with one author in Beijing.

Why: Ranking assigns scores to items within a request while commercial objectives accumulate across a user over days, and hand-tuned fusion weights cannot personalise across that gap. Optimising a proxy for its correlation with sales rewards signals that identify valuable shoppers rather than signals whose promotion changes what those shoppers do, which is the distinction the paper's causal target is built to enforce.