Score the lift, not the likelihood: a B2B uplift model tripled revenue per account by skipping the sure things.

Your best-scored accounts may be your worst targets. A propensity model answers one question: how likely is this account to convert? It then ranks the surest things to the top. But many of those accounts would have bought without the call. Spending a rep on them is motion, not lift. The revenue you can actually create lives one rung down the list, in the accounts that move only because someone reached out. That is the persuadable segment, and a propensity model does not see it. This is the honest reframe of lead scoring, and the direct sequel to our Issue 01 piece on funnel-stage ranking: score the incremental effect of the touch, not the raw odds of the outcome.

A new preprint puts numbers on the idea. We walk its tables, its one field test, and its arithmetic below, and we are careful about which numbers travel.

The paper, read properly

The paper is “VALOR: Value-Aware Revenue Uplift Modeling with Treatment-Gated Representation for B2B Sales,” posted to arXiv in April 2026 by Guduguntla, Soni, and Das [1]. The research question is the one every revenue operator should be asking: under finite sales capacity, which accounts should a model put at the top, the ones most likely to convert, or the ones whose conversion the touch actually changes? VALOR is built to rank the second group, and to weight that ranking toward high-value accounts rather than cheap ones.

The evidence comes in two settings [1]. First, a synthetic benchmark called B2B-Mimic: 236,421 accounts, 100 features, with over 80 percent non-conversion, built to mimic a zero-inflated B2B revenue distribution. Second, a proprietary production dataset of roughly 80,000 billing accounts from the SMB segment of a global cloud provider, where the outcome is continuous, zero-inflated incremental annual recurring revenue. The synthetic generator is potentially reproducible. The production data is not released, so its offline numbers cannot be checked from outside the company.

The design is a methods and benchmark paper plus a single-company field experiment. Offline, VALOR is benchmarked against causal-uplift baselines on both datasets. Online, the authors ran a four-month randomized A/B test on the cloud provider’s Long Tail sales program, 80 percent of accounts treated and 20 percent held as an untouched control [1]. Per-arm sample sizes are not disclosed in the available text. This is an arXiv preprint, version 1, not peer-reviewed.

The outcome definitions matter, so write them down. Offline, the headline metric is the Qini coefficient, a measure of how well the model ranks accounts by true incremental effect, with the random-lift curve subtracted out [1]. Qini is a ranking score, not a calibration score: it tells you whether the order is right, not whether the predicted dollar amounts are correct. Other offline metrics include AUUC (area under the uplift curve), Lift@30 (lift inside the top 30 percent of ranked accounts, a proxy for finite capacity), and KRCC (rank correlation with ground-truth uplift). Online, the two reported metrics are opportunity rate (the share of assigned accounts that became a qualified opportunity) and incremental revenue per account (the change in annualized revenue between the 90-day post-assignment and 90-day pre-assignment windows) [1].

Now the results. On the synthetic benchmark, VALOR’s best variant (CFR-WASS) scored Qini 0.3049 against 0.2596 for the best state-of-the-art baseline, RERUM-DragonNet [1]. The authors describe this as a 20 percent improvement in rankability. The exact ratio is 0.3049 divided by 0.2596, which is 1.174, about 17.4 percent (derived), so the authors round up. Treat it as a 17 to 20 percent range, not a clean number. Notably, that best baseline is reported as failing to converge on the harder production data, so the field comparison runs against the incumbent T-Learner instead.

The architecture earns one paragraph because its ablation is load-bearing. VALOR’s two named ingredients are a gated treatment interaction (a bilinear gate that keeps the “was this account touched” signal from being swamped in high-dimensional feature space) and a value-weighted pairwise ranking loss (which pushes the model to rank high-value accounts well, not just any account). Removing the gated treatment interaction caused the largest Qini drop on both backbones tested. On the DragonNet backbone, the full system scored Qini 0.2821, falling to 0.2539 without the gating and to 0.2799 without the value-weighted ranking [1]. The gate matters most. We return to why in the mechanisms section.

The field number, worked honestly

The result that will travel fastest is the field one, so handle it carefully. In the four-month A/B test, the VALOR-driven targeting policy delivered $1,185 in incremental revenue per account against $445 for the incumbent T-Learner, a net gain of $740 per account [1]. Opportunity rate moved from 9.3 percent to 17.6 percent, a lift of 8.3 percentage points. The authors attribute about $30M in annualized aggregate lift to the rollout.

What does that gap mean on a real targeting budget? Walk it line by line.

The lever here is selection, not effort. VALOR optimizes who you touch, not when or how hard; the authors state that intervention timing is left to future work [1]. The same 3,000-account capacity, re-ranked, also moves the opportunity count.

Now the honesty. The field gap is one experiment at one firm. The paper reports a 95 percent confidence interval on the per-account lift (the +$740 difference over the incumbent) of roughly $429 to $1,051 [1]. Even at the pessimistic bound the advantage stays positive, about $429 per account, which is consistent with the stated significance at p below 0.05 (derived). At the optimistic bound the per-account gain is about $1,051. So the lift looks real across the interval, but the size of it swings by more than a factor of two depending on which end you bank.

Four traps, in plain words, before any of this goes on a slide.

Ranking is not calibration. Qini, AUUC, and Lift@30 measure whether the model orders accounts well, not whether its dollar predictions are right. A model can rank perfectly and misprice every account. Do not read the Qini gain as forecast accuracy.

Offline Qini does not equal field dollars. The jump from a 17 percent offline ranking gain to a 2.7x field revenue gain is asserted through the A/B test, not derived from the Qini number. These are two different measurements on two different datasets. Better ranking plausibly drives better selection drives more revenue, but the magnitude is confounded by being one program at one firm.

The 2.7x is a ratio of two point estimates. 1185 divided by 445 is 2.66 (derived). The paper reports a confidence interval on the per-account lift, not on the ratio itself. A ratio of two noisy numbers is noisier than either. Quote 2.7x as a headline, not a guarantee.

Per-arm sample sizes are undisclosed. The split was 80 percent treated, 20 percent control, over four months, but per-arm counts are not in the available text, so the stated p below 0.05 cannot be independently checked. The opportunity-rate confidence interval was flagged in our source review as possibly garbled, so we do not print it.

Why uplift beats propensity

VALOR is an engineering paper. It does not test a psychological mechanism: there is no buyer-behavior data, no manipulation of attention or persuasion. The “persuadable account” is a decision-theoretic segment, not a measured cognitive state. So the honest read is two-layered. The reason uplift beats propensity is a durable, formalized result in the causal-inference literature. Why this particular model beats prior uplift models rests on the paper’s own ablation, not independent replication. Four mechanisms, labeled by strength.

You can only move the persuadable. Strength: established. A propensity model ranks how likely an account is to convert; an uplift model ranks how much the touch changes that. They disagree exactly where it matters. High-propensity “sure things” buy anyway, so spending finite capacity on them creates zero incremental revenue. The persuadable-versus-sure-thing decomposition, and the formal case for modeling the incremental effect rather than the raw response, are laid out in Gutierrez and Gerardy’s review of uplift modeling [2]. This is not a claim about VALOR. It is the reason any uplift approach beats any propensity approach under a capacity constraint.

Uplift has to be ranked, because it is never observed. Strength: established. You see what happened when an account was touched, or when it was not, never both. There is no per-account uplift label to regress on, which is why uplift is a ranking problem and why VALOR’s headline metric is Qini rather than error. Radcliffe and Surry’s foundational work established that modeling treatment-caused change needs different methods from response modeling at every stage, and developed the Qini-style evaluation now standard in the field [3]. VALOR’s choice to optimize a value-weighted pairwise ranking loss descends directly from this, and its ablation (removing that loss lowers Qini) is consistent with it.

The incumbent was a T-Learner, a known-weak baseline. Strength: established (the limitation is documented; VALOR’s specific margin over it is single-study). The replaced system fit one model on treated accounts, one on control, and subtracted. Kunzel, Sekhon, Bickel, and Yu formalized this family of metalearners and showed why the T-Learner is fragile: each model minimizes its own prediction error, so each can spend capacity fitting structure that cancels out of the difference, and the approach degrades when treated and control groups differ in size or distribution [4]. That is precisely the VALOR field regime: 80 percent treated, 20 percent control, a zero-inflated outcome. The direction of VALOR’s advantage is well motivated; the size of it is setting-specific.

Balancing treated and control in a learned representation. Strength: plausible (established design principle; VALOR’s specific gating gain is internal ablation only). VALOR’s backbones descend from Shalit, Johansson, and Sontag, who showed you can bound the error of a treatment-effect estimate by the sum of a prediction error and the distance between treated and control distributions in a learned representation [5]. The intuition: in high-dimensional account space, touched and untouched accounts differ on many dimensions, and a naive model can mistake those background differences for the effect of the touch. Balancing strips that confound out. VALOR’s finding that removing its gated treatment interaction causes the largest Qini drop is consistent with the idea that the treatment signal needs structural protection. Two honest limits: the bound assumes no hidden confounders, which holds under the randomized A/B test but not necessarily in the offline production data, and the gating-specific gain is the paper’s own ablation, not replication.

Why a revenue team should care

If those mechanisms are why this works, the asset is already on your books. Every account you touched and every account you did not is a row in a natural experiment you are probably not measuring. The reframe costs nothing to test: stop asking your score who is likely to buy, and start asking who buys because you called. The two questions produce different lists, and the gap between them is the budget you are burning on sure things. That is Signal and Measurement work, the discipline of ranking persuadable high-value accounts under finite capacity.

What to do Monday morning, by maturity level

One move per rung of the ladder.

1. Manual. Run the “would they have bought anyway?” audit by hand. Pull your last two quarters of closed-won deals a rep actively worked. For each, ask the closer one question: would this account have bought without our outreach? Tag every deal sure-thing, persuadable, or unknown. If 40 percent of your worked wins land in sure-thing, you just found the budget you are burning. No model required.
2. Assisted. Split your score into two columns. Column one is propensity, which you likely already have. Column two is a persuadability flag. Build it with a holdout: stop touching a random 10 to 20 percent of your mid-tier accounts for one cycle and compare conversion against the touched group. The gap is your lift. Rank by the gap, not the score. This is the cheapest causal experiment in B2B, and most teams have never run it.
3. Orchestrated. Gate targeting on predicted lift, with a permanent control arm. Move from a score threshold to an uplift threshold in routing. Hold back a never-touched control (the VALOR test used 20 percent) so you can measure incremental revenue per account the way the paper does: the 90-day post-assignment delta against the 90-day pre-assignment baseline. The control arm is not lost pipeline. It is the only way to know your targeting is doing anything. Watch the 90-day window honestly; the paper does not establish lift persists beyond it.
4. Autonomous. Weight the model toward the whales, and watch for treatment-signal collapse. Rank by incremental revenue, not incremental conversion, so the system chases persuadable high-value accounts rather than persuadable cheap ones. And internalize the ablation: removing the gated treatment interaction caused the largest ranking drop in the paper, the authors’ evidence that in wide feature spaces the “did we touch it” signal gets drowned out unless the model is built to preserve it. In a wide feature space, the treatment variable needs structural protection or your uplift model quietly degrades into a propensity model wearing a costume.

Where this lands by industry

Four archetypes, parameterized rather than named, with the arithmetic carried from the worked examples. Only the first is a close analog to the paper’s setting; the rest are translation.

Limits and caveats

The honest perimeter. This is an unreviewed arXiv preprint [1]. No psychological mechanism is tested; the persuadable account is a decision-theoretic segment, not a measured cognitive state. The field magnitude is one global cloud provider, one Long Tail program, four months, with no external replication and no explicit external-validity analysis beyond that setting. The 2.7x is a rounded ratio of two point estimates ($1,185 divided by $445 is 2.66, derived), and the reported confidence interval covers the per-account lift ($429 to $1,051), not the ratio. Per-arm sample sizes are undisclosed, so the p below 0.05 claim cannot be checked from outside. The production dataset is proprietary and unreleased, so the production Qini near 1.40 cannot be independently verified; only the synthetic B2B-Mimic generator is potentially reproducible. Durability is unknown: lift is measured over a 90-day window, and persistence beyond it is not established. And a ranking model only orders the list it is given; it does nothing for a list built on a dead ICP, which is a Market problem no scoring system fixes.

What would change our read: an independent replication on a second firm or industry, a public release of the production data or code, or a longer measurement window showing the lift persists. Until then, the portable claim is the question, not the number. If you want the next test of a number like this, we publish one every week. Get the Weekly Research.

What you learned

Propensity ranks the sure things; uplift ranks the persuadable, and under finite capacity that difference is the whole game. VALOR scored about 17 to 20 percent better on Qini than its best baseline and, in one field test, moved per-account revenue from $445 to $1,185. On a 3,000-account budget that is roughly $2.22M of measured 90-day incremental ARR at the point estimate, with the paper’s CI on the per-account lift putting the floor near $1.29M and the ceiling near $3.15M on the same capacity. The mechanism is established in the causal-inference literature. The magnitude is one company’s, on an unreviewed preprint. Split your score, run the holdout, and measure the lift before you bank the multiple.