Explore

Intent channelisation

S&F Problems

DS rationalisation

PRD - Interstitial filters

HVF v2

Hero image revamp reads

PRD - Sort bugs

Objective -

Enable need based users to discover their desired product seamlessly

For users with high intent, identify intent signals and enable the user to discover the right products immediately

For users with low to medium intent, shape intent towards the right products

User journey map -

⁠

⁠

Definition of users based on intent

Users with purely exploratory intents are triggered to open Meesho with the

Tryst to Discover ‘NEW’ Products

Discover variety in ‘TRUSTED’ Categories

Discover products according to their ‘INTEREST’

Users with a Need Based or Specific Intent are triggered to Open Meesho depending on

Perceived VARIETY in the desired product or product category

Perceived LOWEST PRICE FOR THE QUALITY in the desired product / product category

Types of Users:

High Intent: User knows exactly what kind of product they want to buy eg. a blue XS Kurti

Medium Intent: User knows which category of products they want to purchase eg. a Kurti

Low Intent: User is just exploring the app, might buy something if they like it

A user may be a high intent user during a particular session and a low one during another. This will depend on the categories the user is interested in at that time, the time of the day etc.

User problems -

Basis historical LODs, TBR -

User Problem

Low Intent

Medium Intent

High Intent

Users when they know what they are looking for (either they already knew or realised they have some hard filters after scrolling) are not able to efficiently browse for just those products

Not applying filters even when they know they want to purchase a product of say 4 and above rating

Users going through widgets even when they have a further specification in mind eg. clicking on kurti when they are looking for cotton kurtis specifically

Users not being able to find the right search terms

Users not applying/not being able to find the right filters

Users not able to find similar products easily

Users not going to PDP Recommendations or search if they have to find similar products, will try to scroll till they find it

Users not knowing how to look for different color/fabric etc in the products they are already viewing

Users are not able to explore more products and categories easily

Users not knowing about the wide variety of products available on Meesho

There are no rows in this table

⁠

Understanding current capabilities -

Net filter application - Filters are applied on ~8% of all high intent / search feed opens

Of these, 56% usage comes from HVFs, 15% from IFs and remaining from basic / dynamic filters

⁠

Basic filter definition on tech has remained stagnant and has configuration issues (Eg - New arrival / Meesho Mall etc)

Eg - Mall filter historically applies to the original set of ~8 Mall seller’s listings and doesn’t reference the later setup taxonomy attribute

HVF details - HVF model has remained inactive for close to an year and has been serving stale filters for existing feeds

HVF usage breakdown - Other than Fabric, Category invariant filters account for ~70% of historical HVF usage

⁠

IF usage - Reads from experiment detail IF usage to be heavily indexed on Category (~50%), Price (~25%) and Gender (~13%)

IFs show up at a gap index of 6 catalogs, Of these the most interactions comes forth till 3rd/4th IF and starts dissipating post due to low engagement at increased view depth

Search for demand signalling - ~X% of search sessions are high intent where users specifically pinpoint key attributes as part of the query

Filter definitions -

Filters across the platform typically fall down under -

CIF / Category invariant filters - Includes Top quality / New arrival / Program stores

Price filters

Gender filters

Category filters

Rating filters

Size / Colour filters

CSF / Taxonomy filters

Outside IN -

Feature

Amazon

Flipkart

Temu

Myntra

Ajio

Consolidated learnings

HVF

Amazon has HVFs laid out in 2 rows, The first universal HVF pushed toggle for Prime followed by rating, brand and price filters. The next row is dynamic HVFs which are more gender or sub category specific

Flipkart - HVFs only pass in category invariant filters like top brands, new arrivals etc. Similar for Myntra

Temu optimises pricing filters on HVFs followed by rating, category and some category specific attributes

Amazon has each IF pertaining to one filter class label similar as ours, Hierarchy given to Gender followed by category which then extends to size, colour and other taxonomies

Flipkart and Myntra pass on taxonomy filters in IF and also integrate the RE to surface similar search queries with visual references

Temu only surfaces similar search queries interstitially

Ajio pushes brand store cards in interstitial RE

Basic filters

There are no rows in this table

⁠

Summarised view -

Competition generally uses HVFs to surface category invariant filters like price, stores or ratings

BAU filters are much more prominent on comp with filter labels showcased in a horizontal scroller and values showing up in the same screen

IFs are generally used to map in high level demand signals first including category / gender before extending on to category specific feeds

IFs are also used to surface similar search terms and store front cards

Solutioning considerations basis user persona -

Reducing interactions via simplified flows have historically helped forth with conversion (Eg - CBD→ PBD)

Textual comprehension is pretty poor for our users and majority of the information consumption happens through visual cues

Utilise HVFs to isolate prelim evaluation demand channel, i.e. Stores / Gender / Category / Price. Isolate CSF to only surface if used extensively or surface in IF

Tentative solves -

KR name

Sub KRs

Description

Reference

Status

POC

Timelines

H1 - Filter expansion for programs

Adding program wise filters (Gold / Mall etc)

Creating filters for different strategic programs to be used in static as well as HVF filters

Picked

⁠

@Ena Robinson

⁠

@Ayushi Tulsyan

⁠

@Nihal Singh

⁠

22nd Mar

H1 - Base filter definition refresh

Cleaning up outdated legacy filter conditions

Picked for ops clean up

⁠

@Revanth Gundapaneni

⁠

Mar end

H1 - HVF - Extending for strategic programs

Extending program filters across all HVFs

Surfacing program HVFs across all feeds

Picked for analytical model updates

⁠

@Ena Robinson

⁠

@Nihal Singh

⁠

25th Mar

H1 - HVF - Base model reinstatement

Increasing coverage of HVFs across all relevant feed

Picked

⁠

@Charchit Bharani

⁠

Apr 2nd

H1 - HVF / IF - Med term changes

Finalising HVF framework in unison with IF filters to maximise intent channelisation for a given feed

Extending IFs to also surface similar search terms

In product solutioning

⁠

@Nihal Singh

⁠

Apr 3rd week

H1 - Dynamic filter definition

Filter definition basis DS embeddings

Utilising DS embeddings to fetch catalogs mapping to a given filter condition / Do a one time clean up using LLM (Not recommended)

Currently definitions are static and have to be maintained manually leading to ops gaps

⁠

@Nihal Singh

⁠

Next R2R

H2 - Visual discovery for exploratory intent

Creating a fresh discovery flow to visually assist users with exploratory intent to shop by mood

TBP

⁠

@Nihal Singh

⁠

Next R2R

H2 - Basic filter UI refresh

Ui updates to reduce interactions to apply BAU filters

TBP

⁠

@Nihal Singh

⁠

Next R2R

H2 - HVF / IF - User personalisation

Personalising HVFs and IFs to surface filters basis user’s intent profile

⁠

@Nihal Singh

⁠

Next R2R

H2 - Filter rationalisation

Showcasing category specific filters post recording category intent

⁠

@Nihal Singh

⁠

Next R2R

There are no rows in this table

⁠

Sort and filter - Service architecture

⁠

Screenshot 2024-04-30 at 11.54.57 AM.png

⁠

⁠
`CPD`⁠
⁠

Proposed solve(s) - Intent channelisation (IG bits to be covered later)

Filter label service rationalisation -

Pack of 1 vs Single

Black vs Jet Black - Such values to be normalised into one

We will be building a ML (Embedding based Clustering) + LLM (Maintaining Quality) pipeline that will regularly pick up all unique products’ attribute values and normalise them and update this config file, thus changing from a static setup to a dynamic setup.

ML pipeline to select dynamic feed specific filters -

Build out variant support to test dynamic category specific filters that are selected by a ML model after taking in signals from feed’s content and how users are interacting with these feeds across filter classes.

Signals for feed specific filters

Filter usage - Which Filters are being used by users while searching products / Upon historical filter application

Feed’s conversion profile - Which are the products in a category that are being ordered/wish-listed the most. Look at common attributes and prioritise them

UGC - Attributes that are being talked about in the reviews of the products.

Engagement post filter application - Post filter application, how much time the user is spending on the feed

Filter attribute coverage - Percentage of pareto products under a feed having an attribute (predicate probability)

Attribute distribution - How evenly are the attribute values distributed among the products under a category (coefficient of variance)

Attribute cardinality - How many unique attribute-values are there among all products under a category (attribute cardinality)

IG / Demand shaping - Measure of uniqueness or specificity in terms of describing products under a category

How unique is an attribute to a category 𝑢𝑛𝑖𝑞𝑢𝑒𝑛𝑒𝑠𝑠 = (Pareto viewed product count with the attribute in a category) / (Pareto viewed product count with the same attribute within similar categories / portfolio)

Test model output through LLMs

Present filters+category+search tuple to LLM and ask whether the filters make sense or not for that category/search

Use the response to compute model’s precision with a mix of real world usage

ML user persona setup to select feed agnostic filters (Programs / Promos / Deals etc)

User’s persona profile - To be setup basis -

Interactions with primary REs per feature / program -

HP widget / filter / FIF interactions per program (Loyalty / Mall / Gold)

Order / Browsing history - Views and engagement across products part of programs

Feature signals -

FY / Sorting / Browsing activity / Order usage towards high quality / low price / high price / freshly arrived listings / return profile

Demographic signals - Device / Apps / BNPL to map user profile to High price / High quality / Low ASP

Later, Map user level signals to stack rank sort / filter and FIF position per user

Filter relevance model testing

Build variants that -

Prioritises feed agnostic filters (A) + feed specific filters (B) and curates a default set of primary filters to be showcased across HVF / IF and static / dynamic filters with

Priority given to A vs B

Mix of A and B basis net relevance

Experiment and align on the winning variant per cohort X feed

S&F UI / Hygiene updates -

Solve tech bugs related to -

Attribute update event misses on Elastic across platform

Comp / DS / Explore / Ads swaps post sort application

Any other attribute persistence bugs on Elastic

Build tech fallback systems that -

Stop allowing filter labels with single label value

Scale tech systems to read from DS feed’s schema with A/B support to experiment with filter/ sort ranking across filter types (Static + Visual)

Filter label value → Audience ID → Position on HVF → Position on IF → Position of static filter

P1 - Dynamic sorting of sort options

UI / Design updates -

Static filter screen revamp -

Incorporate visual filters on static filter screens

Label name rationalisation to curb overflowing label names

Simplified usability with section separators / infinite page scrolling with auto section change

EXP / P1 - Search within filters

HVF usability update - Making HVFs sticky : Display on feed up scroll, Hide during down scrolls

IF usability updates -

Given lower effective real estate for horizontal IC/IG components within a vertical scroll feed, Optimise IFs design to showcase all applicable filter labels as follow -

⁠

⁠

Clicking on More continues session visually over static filter screens

Label stepper on top to be pre selected with label per IF position

Unified horizontal IC / IG service architecture to experiment and scale FIF and IF

Scalable tech system to experiment with variants and scale horizontal IC/IG components per feed basis inputs from -

DS / Heuristics model

Manual admin input as fallback

Prioritisation / Phasing of solves -

⁠

⁠

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

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

CtrlP

) instead.

Intent channelisation

Objective -

User journey map -

Definition of users based on intent

User problems -

Understanding current capabilities -

Outside IN -

Solutioning considerations basis user persona -

Tentative solves -

Sort and filter - Service architecture

⁠CPD⁠⁠

Proposed solve(s) - Intent channelisation (IG bits to be covered later)

Prioritisation / Phasing of solves -

⁠
`CPD`⁠
⁠