Experiment summary #
eBag used shopper-level AI to show past purchases missing from the current grocery basket. Releva reports that 30% of 693,666 orders contained at least one clicked recommendation, but the Bulgarian retailer kept no holdout.
- Bulgarian online grocer eBag connected its live catalogue and order history to Releva’s recommendation system.
- The busiest placement showed each shopper products from previous orders that were missing from the current cart.
- From 1 June to 28 September 2026, 210,119 of 693,666 orders contained at least one product clicked in a recommendation.
- The cart-page block produced 20.7 orders per 100 clicks and 68% of revenue attributed to recommendations.
- These are click-attributed results, not measured lift. eBag had no holdout, and Releva excluded two months after finding a counting error.
11 min read
Just before an eBag shopper checks out, one recommendation block asks a narrow question. Which products has this person bought before but forgotten today?
The answer might be milk, bread or coffee. The products must still be available and absent from the open basket. The shopper decides whether any of them belong in this week’s order. There is no chatbot and no campaign, just a prompt inside a task the customer is already doing.
That makes the eBag AI recommendations case more useful than its largest number first appears. Releva says three in ten orders included a product found through a recommendation during a four-month window. The system was present across several page types, yet the cart-page reminder accounted for most of the attributed recommendation revenue.
The question for a growth team is whether a retailer can turn a repeated basket into a small, testable product decision, then measure it without confusing a shopper’s existing intent with lift. Personalisation can always produce an attribution total; the holdout is what tells you its value.
A thirteen-item basket leaves useful gaps #
eBag describes itself as a Bulgarian online supermarket carrying more than 15,000 grocery and household products. It delivers in Sofia and several other Bulgarian cities. Its terms name Convenience AD as the operator.
Releva’s case says eBag processed about 180,000 orders in September 2026, with roughly 13 items in each order. That is a particular kind of conversion problem. A shopper has already decided to buy, but a long repeated basket can still be incomplete. Finding one relevant omission can add value without creating a new visit or discounting the cart.
The incumbent process was ordinary site merchandising. Product lists appeared on the home, category, product and cart pages. The change was to let an algorithm choose the items for the person in front of each placement, using the retailer’s own catalogue and order history.
eBag has used Releva since 2019, according to the supplier. The 2026 reporting does not describe a single launch day or a clean switch from manual merchandising. It documents a mature recommendation programme and the placements that were operating during a defined measurement window. That distinction matters because this is an operational case, not a launch experiment with a before-and-after control.
The cart asked what was missing #
The Releva AI setup began with two data feeds. eBag connected products, availability and language, then supplied each shopper’s orders. The system used those inputs to decide which products appeared in four placements.
The most specific placement sat on the cart page. It showed products from that shopper’s past orders that were not in today’s cart. The home page carried themed rows. A merchandiser named a theme, such as seasonal fruit or pharmacy, and set one filter, while the system selected the products within that boundary for each shopper. Category pages showed popular items, and product pages showed products bought together.
Human control remained deliberately small. The team chose themes and rotated the rows. Releva says 206 themed rows ran during the June to September window, with 372 launched since January 2025. The algorithm handled product selection inside the row. The rules were limited to stock availability, shopper language and the theme filter, with no manual boost for price or brand.
This is online grocery personalisation as a constrained decision system. The retailer supplied the catalogue, inventory and transaction history. The merchandiser supplied the commercial frame. The model ranked products, but a person still chose where the module appeared and what sort of selection the page should make.
The cart page carried most of the value #
From 1 June to 28 September 2026, Releva counted 693,666 eBag orders. Of those, 210,119 contained at least one product that the shopper had clicked in a recommendation. That is the reported 30% share. The placement breakdown is more instructive than the headline. The cart page recommendations generated 20.7 orders for every 100 recommendation clicks and accounted for 68% of recommendation-attributed revenue. Home-page themed rows generated 14.8 orders per 100 clicks and 13% of attributed revenue. Category pages produced 15.8 orders per 100 clicks and 12% of attributed revenue. Product pages produced 5.2 orders per 100 clicks and 8% of attributed revenue.
Those figures point to a commercial owner and a decision. This is mainly a product and merchandising question, with basket size and order conversion as the business outcomes. The cart-page module reached shoppers at the moment when remembered staples could still be added, and its rule was easier to explain than a broad promise of AI personalisation.
Releva also gives a trend across three months. The share of orders containing a recommended product was 28.0% in June 2024, 20.3% in June 2025 and 30.4% in August 2026. Recommendation clicks per 1,000 page views went from 31.5 to 34.5 to 47.9, and the share of online revenue attributed to recommendations from 3.6% to 2.7% to 4.3%. The latest reading is higher than the first, but it is not a steady climb.
The supplier says other things changed too, so this is a trend, not a controlled result. Themed rows launched during 2025, and traffic and customer mix moved as well. Monthly orders grew from 74,582 in June 2024 to 177,681 in August 2026.
Thirty percent is attribution, not lift #
Releva credits a product when the shopper clicks it in a recommendation and later buys it. An order counts once in the 30%, even if it contains several recommended products. Revenue by placement is the value of the clicked products that were bought.
That method establishes a traceable path from module to order. It does not establish what would have happened without the module. A shopper who clicks familiar milk may already have intended to buy it. Someone who engages with a recommendation is also likely to have more purchase intent than a visitor who does not.
The conversion gap in the case shows the selection problem clearly. Releva reports that 40% to 53% of influenced visitors bought, against 2% to 3% overall. Influenced visitors had opened at least one page through Releva and chosen to click. The supplier warns that part of the gap is intent. It should not be read as a conversion lift caused by the software.
eBag had no holdout group. The strongest defensible result is therefore that recommendations participated in a large share of completed orders and that the cart module produced the largest attributed contribution. The case cannot say how many orders, items or euros were incremental.
A counting error changed the window #
The source discloses an unusually useful failure. April and May 2026 were removed because a counting error credited 5.7 recommended items per order instead of the usual 2.2. Releva does not explain the code path or how the mistake was found, but publishing the exclusion prevents the inflated months from silently supporting the result.
For an ecommerce recommendation attribution pipeline, that is an operating lesson in its own right. The event model needs an order identifier, product identifier, placement, click timestamp and attribution rule that can be reconciled against the final basket. A ratio such as recommended items per attributed order can act as a simple guardrail. When it jumps far outside its usual range, the team should stop the report before it reaches a commercial decision. The case also defines conversion as the share of a month’s visitors who bought during that month, not a visit-level rate. That choice can be valid for a repeated grocery business, but it makes direct comparison with session-based ecommerce benchmarks unsafe.
The next test belongs on one placement #
Releva’s own proposed next step is a holdout on the cart block. Some visitors would see the usual-items module and others would not. Assignment should persist at shopper level so repeat visits do not leak between treatment and control.
The primary metric should be revenue or gross margin per eligible visitor, not clicked-product revenue. Orders per visitor, items per order and average order value would explain the movement. Checkout completion, product removals, substitutions and fulfilment margin would be useful guardrails. A module that adds low-margin staples while slowing checkout can look better in attributed revenue than it is for the business.
The test can stay bounded. Run only the cart-page recommendations, keep the other placements unchanged and report the eligible population, allocation, duration and confidence interval. If the module wins, a second test can compare the shopper’s own past purchases with generic bestsellers. That would tell eBag whether personal memory is doing work beyond a well-timed reminder. It is the analyse, prioritise and test loop that Hacking Growth by Sean Ellis and Morgan Brown describes: each result decides the next experiment.
| Experiment detail | What the reporting shows |
|---|---|
| Commercial problem | Help repeat grocery shoppers complete a long basket without relying on a new visit or discount. |
| Intervention | Shopper-level recommendations across the cart, home, category and product pages, using live catalogue data and order history. |
| Most specific lever | Show products from the shopper’s past orders that are available and missing from the current cart. |
| Window | 1 June to 28 September 2026, covering 693,666 orders. |
| Reported result | 210,119 orders, or 30%, contained at least one product clicked in a recommendation. |
| Cart-page result | 20.7 orders per 100 clicks and 68% of recommendation-attributed revenue. |
| Counterfactual | None. eBag had no holdout, so the figures measure attributed participation rather than incremental lift. |
| Data limit | April and May 2026 were excluded after a counting error overstated recommended items per order. |
| Next test | A persistent visitor-level holdout on the cart block, judged on incremental revenue or margin per eligible visitor. |
Sources and disclosure #
- [Releva’s eBag case study](https://releva.ai/case-study-ebag)
- [Releva’s method and setup notes](https://releva.ai/case-study-ebag-method)
- [eBag’s official service information](https://ebag.bg/en/llm-info)
- [eBag’s terms and company details](https://ebag.bg/en/terms)
Releva supplied the recommendation system and published both the case and its methodology, so it has a commercial interest in the result. The case names no individual at eBag. eBag has not published the underlying order data or an independent methods report.
FAQ #
What did eBag use recommendations for? #
eBag connected its live catalogue and order history to Releva so shoppers see recommended products on the home, category, product and cart pages. The cart-page block shows items from past orders that are missing from the current basket.
What result was reported? #
Releva reports that 210,119 of 693,666 orders, about 30%, contained at least one product clicked in a recommendation between 1 June and 28 September 2026.
Does 30% mean recommendations lifted orders by 30%? #
No. It is click attribution, not measured lift. eBag had no holdout group, so the case cannot say how many orders, items or euros were incremental.
Which placement mattered most? #
The cart page, with 20.7 orders per 100 clicks and 68% of recommendation-attributed revenue.
Who published the result? #
Releva supplied the recommendation system and published the case and its methodology. It excluded April and May 2026 after a counting error overstated recommended items per order.
What can another growth team copy? #
Use past purchases in one bounded cart-page module, reserve a persistent holdout and judge incremental revenue or margin per eligible visitor. Compare more measured experiments in the Signal Index, browse the Growth Signal archive, or use the benchmark tools in the IC Lab.