Skip to content
Feedback · Acquisition & Conversion

When an almost-right recommendation is enough to lose the conversion

A search engine can return products broadly close to a query while violating its most important criterion. On a catalogue with strong constraints, the value of a search depends not only on what it goes back, but on what it can exclude.

  1. Preferencescan influence classification
  2. Obligationmust filter eligibility
  3. Uncertaintymust be climbed, never guessed

This feedback comes from a project prior to CAIAC. It illustrates the operational mechanisms that contributed to the platform's design. The results are specific to the project context and do not constitute a performance guarantee.

The challenge

Why the business result was out of reach

For a vegan person, recommending a non-vegan product is not a small inaccuracies. For a parent looking for a natural product suitable for a child, an ambiguous feature is enough to stop the purchase.

In these situations, a "almost correct" recommendation may be more damaging than any recommendation: the error does not cause a product to lose, it loses confidence in the entire catalogue.

The traditional internal motor was based on lexical and frequent correspondence: it could lack synonyms, overvalue a term present in a description or cause a product to be traced up incompatible with an implicit constraint.

The starting point

Data, processes and constraints

  • An e-commerce of natural, organic and vegan products.
  • Very demanding audiences, whose criteria are not preferences but conditions of purchase.
  • Multiple intentions: vegan products, untested products on animals, natural products adapted to certain family uses.
  • A project carried out in 2022, with less mature models and a still large control load.
The approach

What was implemented

  1. 01
    Understanding intentBringing together a demand for relevant products even when the same words are not used in the catalogue.
  2. 02
    Distinguishing preference and obligationA preference is used for ranking; an obligation becomes an exclusion rule. Consolidating the two exposes the user to an unacceptable recommendation.
  3. 03
    Check the quality of attributesA system cannot guarantee "vegan" or "untested on animals" if the information is missing, ambiguous or unverified: uncertainty must be visible.
  4. 04
    Align the page with intentSelection, title and explanations responding to the same request — a good list under an overly broad promise remains a risk.
  5. 05
    Maintaining a Human-in-the-LoopAmbiguous cases, new categories and sensitive information are climbed, never solved by a supposition of the model.
Results

What has been measured, and what it means

  1. Observed result

    Personalized pages were compared with the experience based on the site's internal search. The historical test reported 3.2 times higher conversion for the approach combining semantic understanding, constraint-based selection and dedicated pages.

  2. Interpretation

    This result is notable, especially since it dates from 2022. However, it is not sufficiently documented to be included as a promise: exposed populations, definition of conversion, duration, distribution of traffic and significance remain to be established. What the case clearly demonstrates, however, is that semantic similarity alone is not enough: We need to govern eligibility.

Why does it matter

How this translates into business impact

  • Fewer exits due to poor recommendation

    An error in an identity or ethical criterion can lead to a loss of confidence in the entire catalogue.

  • Better use of paid traffic

    When the campaign promises a specific feature, the destination must respect it.

  • Fewer operational errors

    Explicit constraints limit requests to customer service, cancellations and returns.

  • Reusable product knowledge

    Formalized attributes, exclusions and confidence levels are then used for research, campaigns and CRM.

The principle

What this project inspired in CAIAC

This project has taught us that improving conversion on a high-constraint catalogue requires more than better similarity: we need to transform non-negotiable expectations into explicit rules and know how to increase uncertainty.

CAIAC today industrializes this principle: intent, mandatory constraints, eligible products, semantic ranking, control, page, test and learning. Sensitive attributes, certifications and exclusions remain traceable. No algorithm, catalogue, page or rule of the historical project is included.

About this feedbackShow less

This feedback comes from a project prior to CAIAC. It illustrates the operational mechanisms that contributed to the platform's design. The results are specific to the project context and do not constitute a performance guarantee.

Items to be documented:

  • Exposed populations on both sides of the test.
  • Definition of conversion.
  • Duration of test and distribution of traffic.
  • Average basket and campaign sources.
  • Statistical significance: the 3.2× ratio is not highlighted until the protocol has been documented.

Anonymized project · pre-CAIAC · unsecured results. Deployment is adapted to the data, systems and levels of autonomy available in the company environment.

Is this your issue?

This axis brings together the levers that work together and are measured on the same indicators.