What changes when a retailer uses Haggl?

Methodology

The same shopping task, tested both ways.

What the numbers mean

Recommendation rate

The share of valid paired tasks in which the agent chose the target retailer. It is not checkout conversion or organic agent discovery. Across complete pairs, recommendations rose from 89/357 (24.9%) to 232/357 (65.0%): a 160.7% relative lift.

AOV when recommended

The average net merchandise value of the target basket, only when that retailer was recommended. Agents choosing competitors are excluded from this average. The headline +6.5% is the median of the 15 category AOV changes; it does not pool coffee, furniture and hotel dollars. The shoppers contributing to conditional AOV differ between arms.

Primary outcome: value per task

Net target-store merchandise value per eligible task, with zero when another retailer or no basket was chosen. Hospitality includes the complete stay’s room and services, excluding pass-through taxes and fees. This metric captures both winning recommendations and basket value.

The Haggl policy

Shared shopper context, frozen model predictions and a merchant agent selected executable discounts, upgrades, bundles, gifts or delivery offers. Incentives were capped at 15% of the shopper’s budget, with an additional cash-discount cap and a 20% modeled contribution floor. The comparison measures the complete policy, including its incentive spending.

All results, including uncertainty

Nine categories met the prespecified positive threshold for value per task, three were inconclusive and three had incomplete valid pairs. The green lifts in the summary show the direction of observed changes; they are not significance labels.

Intervals use 20,000 paired shopper bootstrap draws. The adjusted intervals use Bonferroni alpha 0.05/15 across the 15 primary comparisons. Small cohorts and percentile-bootstrap tails make this adjustment approximate. A positive finding requires all 24 valid pairs and an adjusted lower bound above zero.

What this benchmark supports

Evidence for the combined Haggl offer policy under the disclosed synthetic retail and hospitality conditions. It does not establish real-world sales, LTV, universal shopping-agent compatibility, or the incremental benefit of ML alone. It also does not test SaaS, utility or financial-service purchases.

Every category is reported, including AOV declines. Training and test personas were disjoint, but predefined shopping situations recur. The small samples do not represent all shoppers or merchants. A separate statistical calculation reproduced the category estimates and intervals; an audit checked saved agent responses, quotes, frozen sources and model hashes.

Three early training attempts failed because concurrent processes shared a temporary instruction file. Those failures were retained and an operational amendment isolated subsequent shopper processes before evaluation. No catalogs, prompts, offer policy or statistical rules were changed using test outcomes.