Explore

Hero image revamp reads

Summary -

Team identified failures in the hero product classification model during CPDs and initiated optimisation and restoration to assess platform conversion potential.

A category-level A/B test was conducted to gain directional insights quickly, avoiding the complexity of a full user-facing A/B test. The experiment was launched for ~4L products contributing to ~3% platform OC across Home Decor, Men’s Fashion, and Women’s Kurtis/Kurta sets.

Early results show positive impacts, with an expected 0.6% platform conversion (O/V) improvement upon scale up.

Key takeaways -

Catalogs with updated hero registered 8-10% improvement in O/V across fashion super portfolios, while non-fashion (Home decor) remained flat

Previously underperforming by ~5% in terms of O/V, the updated catalogs have now closed the conversion gap and are performing better, particularly in fashion portfolios

View share for corrected catalogs increased by ~1pp, improving conversion for an already significant view share with incorrect hero products (~10%)

After removing SP growth rate and weighing views, fashion portfolios showed a ~1% O/V improvement, while non-fashion stayed flat

Extrapolating a 1% O/V boost in fashion, which holds 70% of views, the platform is expected to see a 0.6% O/V uplift

Problem context -

Hero product classification was previously managed by a Spark model, selecting the product image on PLP cards based on the highest transactors.

However, the model had not been running for over a year due to query failures from -

Unoptimised queries running every 2 hours, scanning lifetime orders for all listed catalogs

Data stored as logs in append mode, significantly inflating table size

Hero product update events to taxonomy required self-referencing prior day’s logs which was incorrectly handled

The issue impacts ~11% of OC (~22L catalogs) on PLPs, displaying the first product uploaded by the seller instead of a high-order product. This was identified as a quick win to experiment and improve platform conversion.

Experiment background -

The team opted not to run an A/B test due to the cross-org effort required across taxonomy, feed aggregator, ranking, and search tech.

Instead, we targeted high VC portfolios in fashion and non-fashion to estimate directional platform impact.

The legacy hero product model was restored, updating ~4L products (~3% platform OC) across Home Decor, Men’s Fashion, and Women’s Kurtis/Kurta sets.

Hero products were selected using the existing logic of choosing the most transacted item in the catalog.

The experiment went live on 30th Aug.

Note - Base business logic has not yet been rectified and will be assessed iteratively

Detailed experiment insights -

We have directionally positive reads from hero product model restoration and expect 0.6% jump in platform conversion (O/V) upon scale up.

Conversion (O/V) for updated catalogs improved significantly by ~8-10% for fashion super portfolio, while it remained flat for non-fashion super portfolios (Home decor)

Previously underperforming by ~5% in terms of O/V, the updated catalogs have now closed the conversion gap and are performing better, particularly in fashion portfolios

View share for the corrected catalogs was previously ~10% across super portfolios, which has now improved by ~1pp. The optimisation has enhanced conversion on a significant VC base with incorrect hero products

To isolate the impact of the updated catalogs onto super portfolio’s performance (SP), SP’s growth rate was removed and view weighted resulting in net super portfolio’s conversion improvement by ~1% O/V jump for fashion portfolios, trend remained flat for non fashion

To extrapolate O/V impact on platform, we boosted fashion O/V for fashion portfolios by ~1%, which account for 70% of views leading to expected platform O/V uptake by ~0.6%

⁠

to link G sheet data cuts / and update rating and NQD profile change for the affected super portfolios

⁠

Next steps -

Pre sale freeze - Assess reverting hero products in catalogs where sale discounts apply to other products

Post sale freeze - Gradually scale to more super portfolios and OC to monitor platform metrics

Update the base business logic to dynamically surface hero products using additional ranking parameters

Conversion (O/V) improved significantly by ~8-10% for fashion super portfolio where the hero product was updated, while it remained flat for non-fashion SPs

⁠

⁠

Previously underperforming by ~5%, the updated catalogs have now closed the conversion gap and are performing better, particularly in fashion portfolios

⁠

⁠

View share for the impacted catalogs was previously ~10% across super portfolios, which has now improved by ~1pp. The optimisation has enhanced conversion on a significant VC base with incorrect hero

⁠

⁠

Net conversion impact on super portfolio - To isolate the impact of the hero product update per super portfolio (SP), the SP growth rate was extrapolated and removed from the conversion gains. This showed a ~1% jump in O/V for fashion portfolios, while non-fashion portfolios remained flat

⁠

⁠

Net conversion impact on platform - With fashion holding a 70% VC share and a 1% jump in O/V, while other non fashion SP remain flat, the expected platform-wide impact is a 0.6% increase in O/V

⁠

Want to print your doc?
This is not the way.

Try clicking the ··· in the right corner or using a keyboard shortcut (

CtrlP

) instead.