6 min read
Galeria had to show that AI search would earn more than its new usage bill cost.
Anyone who has typed a product into a shop’s search box and scrolled past the wrong results knows how quickly patience runs out. Galeria’s old site search had reached its limit. The replacement promised better product discovery and changed the bill. Moving to AI commerce search meant paying according to usage, so the German department-store group needed evidence that the commercial return would grow faster than the new variable cost.
That turned a technology migration into a growth experiment. Would more relevant search results create enough incremental revenue to justify rollout, or would shoppers click different products without buying more?
In a case study published on 13 April 2026, implementation partner e-dialog says Galeria tested the new search across 1.5 million sessions. Galeria’s ecommerce director, Bert Middendorp, is quoted saying the test produced a 15% increase in sales. The click rate remained stable.
A variable bill required a commercial baseline #
Galeria is headquartered in Düsseldorf and operates a large department-store business online and through physical locations. The e-dialog account says the retailer was evaluating Google’s Vertex AI Search, which was later folded into Google’s AI Commerce Search product.
The growth constraint was unusually clear. The two options carried different cost structures. Galeria’s existing system represented the baseline. The proposed system would process search and customer signals through a machine-learning layer while creating usage-based AI costs that rise with activity.
The source supports a simple hypothesis. Galeria wanted statistical proof that better search relevance would create a commercial upside before rollout. It discloses no break-even target, forecast cost or minimum acceptable lift.
The test followed revenue past the click #
Before the ecommerce search A/B test, e-dialog says the product feed was cleaned and prepared for model training. The team then ran the comparison across 1.5 million sessions and built tracking that attributed revenue to search results.
That sequence matters. A ranking model can score better on an offline relevance test and still produce no commercial gain. It can also increase clicks by making results more interesting without improving completed purchases. Galeria kept revenue in the measurement path and treated engagement as a step along the way. Allegro’s recommendation test showed a similar gap, where a plausible AI pick still produced an awkward commercial result.
e-dialog reports that clicks stayed stable while purchase intent rose. Middendorp says sales increased 15%. The case page also refers to purchase-rate and add-to-cart lifts, but its accessible text publishes no values for them. It also gives no absolute revenue baseline, search-session share or cost for the new system.
The result therefore supports a narrower claim than the case study’s promotional language suggests. The reported revenue lift occurred during a large test of the new search setup. The net profit effect after cloud fees, implementation costs, merchandising changes and margin differences between the products shown remains unknown.
The control is described less clearly than the treatment #
The case calls the work a scientific A/B test, but it gives no traffic split, test duration, unit of randomisation (whether shoppers were split by session or by user), confidence interval or exact control experience. The natural inference is that the previous search was the comparator, because the business question concerned migration. The source leaves that configuration unconfirmed, so it cannot be presented as fact.
The report also leaves the ranking model tangled up with the preparation around it. Product-feed optimisation happened before the test. Better titles, attributes or availability data can improve a search system regardless of the model. The 15% figure belongs to the combined treatment and says little about an isolated algorithm.
The sourcing has another limit. e-dialog helped design the validation and published the case. A second implementation partner, KPS, independently describes Galeria’s use of Google’s retail search and recommendation systems, yet its page omits the 1.5 million-session design and the 15% figure. It reports a 17% rise in search conversion and a 43% rise in average order value instead, so the two partners do not cite the same results. The measured result remains vendor-published and company-endorsed, and no outside party has audited it.
Copy the migration test before buying the model #
A growth team can copy the operating move without building a search model. Define the incumbent experience as the commercial baseline before signing off a usage-priced replacement. Randomly split eligible search sessions, preserve a stable control and measure the entire path from query to purchase.
The primary metric should match the bill. If the supplier charges by query or session, track incremental revenue and gross margin per eligible search session. Add the model fee, implementation cost and merchandising workload to the same decision sheet. A revenue lift is useful, but the migration earns its place only when the contribution after those costs clears the agreed threshold.
Keep click-through rate as a diagnostic and let sales decide the verdict. Galeria’s reported pattern, stable clicks alongside higher sales, shows why. The system may have changed which products shoppers saw, how quickly they found a viable item and what they eventually bought, with interaction levels unchanged.
The next test should examine durability. Hold back a fixed share of traffic after rollout and compare net revenue, margin, returns and cloud cost over a longer window. That would show whether the 15% revenue lift persists once novelty, feed clean-up and launch conditions fade.
Galeria’s experiment offers a practical rule for AI procurement. When software introduces a metered cost, demand a controlled business metric that uses the same unit of activity. Model quality can explain why a system might work, and a commercial holdout shows whether it deserves more traffic.
Sources #
- [e-dialog case study on Galeria](https://e-dialog.group/en/success-story/galeria-from-vision-to-measurable-performance-with-google-cloud-ai/)
- [KPS reference on Galeria’s Google Cloud search implementation](https://kps.com/references/galeria/)
- [Google Cloud documentation for AI Commerce Search](https://docs.cloud.google.com/retail/docs/what-is-it)
- [Bert Middendorp on LinkedIn](https://nl.linkedin.com/in/bertmiddendorp)
Compare the evidence and operating lessons from more AI experiments in our Growth Signal Index, or browse benchmark tooling in the Industry Contents Lab.
Spotted an error? See our corrections policy.