The practical brief
The finding
Allegro’s additional recommendation retriever increased carousel-attributed merchandise value on organic product pages and in-cart checkout, but pre-cart tests generally showed no significant gains. [c3] [c5] [c6] [c7]
Why it matters
A recommendation strategy that succeeds at checkout may not help earlier in the shopping journey; better offline compatibility scores are not enough to establish commercial value.
| Organic placement | Web outcome | App outcome |
|---|---|---|
| Product page: complements plus substitutes | +8.05% carousel-attributed GMV versus production baseline; p<0.005 | +9.35% carousel-attributed GMV versus production baseline; p<0.005 [c5] [c6] |
| Pre-cart: expert-filtered training | Qualitative summary: no significant attributed-GMV gain; visit conversion declined significantly | Qualitative summary: no significant attributed-GMV or visit-conversion change [c5] [c7] |
| In-cart: same-category mapping | +21.25% carousel-attributed GMV versus production baseline; p<0.005 | +15.73% carousel-attributed GMV versus production baseline; p<0.005 [c5] [c7] |
What to try
BLURSOR’s practical interpretation
Our interpretation: evaluate recommendation changes separately by placement and business outcome, rather than applying one accessory-only policy across the entire journey.
Study boundary
These are short-term results from one marketplace with same-seller delivery incentives, not evidence about visibility in general-purpose AI answers.
Choose the recommendation policy by shopping context
Before making product suggestions more narrowly focused on compatible accessories, ask where shoppers will encounter them. Allegro’s study shows that commercial effects varied by placement: broader related-product discovery helped in some contexts, while stricter filtering did not reliably improve purchases.
This is research about marketplace recommendation discovery—not whether ChatGPT or another general-purpose assistant will mention your business. Its relevance for owners and marketers is narrower: how catalogue information, product relationships and placement rules influence which products a recommendation system surfaces.
The study separated buying together from working together
Allegro built AlleCompanion, a retrieval system that selects candidate products for recommendation carousels. It represents products through content such as title, price and category, with support for attributes and seller information. A separate category mapping combines purchase patterns, expert rules and human annotations, including an LLM-assisted workflow.
The distinction matters: complements work together, while substitutes offer alternatives. Frequent co-purchases can contain either—or unrelated items. In an internal exercise involving 400 product pairs, with two-annotator consensus, minimum-count filtering reduced unrelated pairs from 44% to 22%. Excessively high thresholds increased substitutes from 28% to 43%, while complements rose only from 29% to 36%. Those are relationship shares, not recommendation accuracy.
Offline experiments used 90 days of training transactions and a subsequent seven-day holdout. Expert filtering improved attribute consistency—matching-feature overlap used as a compatibility proxy—but reduced retrieval relevance. Training exclusively on synthetic expert-rule pairs performed poorly.
The live A/B experiments each lasted two weeks across Web and App. AlleCompanion supplemented existing production retrievers rather than replacing them. Organic recommendations were restricted to the same seller, supporting shoppers trying to reach a free-delivery threshold.
Attributed sales rose, but not everywhere
On organic product pages, mixing complements with same-category alternatives increased carousel-attributed gross merchandise value, or GMV, by 8.05% on Web and 9.35% on App. At in-cart checkout, same-category mapping increased attributed GMV by 21.25% and 15.73%, respectively. These were relative changes against production baselines, all marked significant at p<0.005—not percentage-point gains or increases in total marketplace sales.
Pre-cart tests generally produced nonsignificant GMV and carousel-conversion fluctuations. Expert-filtered training also reduced Web visit conversion—the share of visits resulting in a purchase—by a significant relative 0.28%.
Sponsored product-page recommendations generated approximately 50% more carousel ad revenue across both platforms, but GMV changes were not significant. Separately, the mapping’s much larger reported click and conversion lifts came from historical-traffic offline evaluation, not live A/B tests.
A realistic interpretation—and what remains unknown
Our interpretation is that recommendation quality should be judged against the job of each placement, not a universal definition of compatibility. The results support testing broader discovery in some contexts; they do not prove that alternatives always outperform accessories.
The study cannot establish long-term gains, effects on another retailer, or exhaustive item-level compatibility. Category granularity can obscure product nuances, new categories may lack coverage, and candidate generation has limited personalization. Production feedback also complicates offline evaluation.
- For catalogue teams: review titles, categories and compatibility attributes because these are inputs to the studied system—not because this paper proves that editing them increases sales.
- For recommendation owners: test placements separately and distinguish attributed sales, visit conversion and advertising revenue.
- For marketers: do not present these marketplace results as a proven tactic for winning AI-answer mentions.
The source and evidence
Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro
Study limitations and disclosures
- This is an internally conducted study of one marketplace. Same-seller restrictions and free-delivery incentives shape its organic recommendation results; it does not test discoverability or credibility in general-purpose AI answers.
- The live tests lasted two weeks each. The paper does not report arm-level sample sizes, allocation ratios, confidence intervals or absolute baseline conversion and GMV values. Reported effects are relative changes, and GMV is attributed to carousel interactions rather than total marketplace sales.
- The category-mapping click and conversion lifts were measured offline using historical traffic. Attribute consistency is a matching-feature proxy, not a direct measurement of compatibility accuracy.
- The authors acknowledge taxonomy-limited precision, possible coverage gaps in extremely new categories, limited personalization and production feedback loops. High traffic coverage does not establish complete catalogue coverage or item-level compatibility.
- Four authors are affiliated with Allegro.com; Michał Bień is affiliated with NVIDIA, with the note that his work was done at Allegro.com. All authors disclose equal contributions. The research evaluates Allegro’s own commercial recommendation system, creating a material employer interest in the reported outcomes. No separate funding statement is provided in the supplied paper.
Evidence behind this briefing
[c1] AlleCompanion uses product content rather than item IDs, including title, price and category, with support for product attributes and seller IDs. This supports retrieval across a volatile catalogue; ComCat separately supplies category-level guidance that can be changed without retraining.
§3.1.1 Two Tower; category-policy decoupling described in §3.3 Online Deployment · Read the study
[c2] Co-purchase frequency was not a reliable substitute for functional complementarity. In an internal annotation exercise involving 400 item pairs and two-annotator consensus, minimum-count filtering reduced unrelated pairs from 44% to 22%, but excessively high thresholds increased substitutes from 28% to 43%, while complements rose only from 29% to 36%. These are annotated relationship shares, not recommendation accuracy.
§3.2.2 Behavioural filtering and heuristics · Read the study
[c3] In offline experiments using 90 days of training transactions and a subsequent seven-day holdout, transaction-trained AlleCompanion performed best on Recall@20 and MRR@20. Expert filtering traded retrieval relevance for higher attribute consistency, a compatibility proxy; synthetic expert-rule training alone performed poorly. This does not establish that expert rules improve purchase outcomes.
§4.2.2 Discussion; Table 3; training and holdout protocol in §4.1.2; proxy definition in §4.1.1 · Read the study
[c4] The full ComCat ensemble, including same-category alternatives, produced reported relative increases of 450% in CTR and 398% in CVR against the heuristics-only baseline in an offline historical-traffic evaluation. These are online-like offline metrics, not live A/B-test conversion gains or direct measurements of user preference. Approximately 42% source-category coverage represented 99.8% of traffic in that evaluation.
§4.3 Complementary Categories Mapping; Table 4 · Read the study
[c5] Each online A/B experiment ran for two weeks across Web and App and routed 100% of platform traffic. AlleCompanion was an additional retriever alongside production models, not a standalone replacement. Organic carousels were restricted to same-seller items; GMV measured value directly attributed to carousel interactions, not total marketplace sales.
§4.4 Online A/B testing · Read the study
[c6] Against the existing product-page production baseline, sponsored complementary retrieval increased visit conversion by 0.53% on Web (p<0.005) and 0.13% on App (not marked significant), with approximately 50% higher carousel ad revenue but nonsignificant GMV changes. Organic retrieval mixing complements and substitutes increased carousel-attributed GMV by 8.05% on Web and 9.35% on App, both p<0.005, while visit conversion showed neutral fluctuations. These are relative changes, not percentage-point gains.
§4.4.1 Product Page; Table 5 caption defines statistical significance as p<0.005 · Read the study
[c7] Business effects depended on placement. Pre-cart strategies generally produced nonsignificant GMV and carousel-conversion fluctuations, and expert-filtered training reduced Web visit conversion by 0.28% (p<0.005). At in-cart checkout, same-category mapping increased carousel-attributed GMV by 15.73% on App and 21.25% on Web, both p<0.005, against the production baseline. These experiments support placement-specific effects, not a universal advantage for alternatives or stricter compatibility.
§4.4.2 Cart Placements; Table 5 · Read the study
[c8] The authors acknowledge taxonomy-bounded precision, possible extreme-cold-start coverage gaps, limited personalization in current candidate generation, and production feedback loops that complicate offline evaluation. Reported traffic coverage does not establish exhaustive catalogue coverage or item-level compatibility.
§5 Conclusions, final paragraph · Read the study