Blinkit Flash: AI Shopping Assistant

The tension

Blinkit has won the replenishment habit, and that same habit may limit category expansion. The better it gets at helping people retrieve what they already need, the less reason they have to explore.

The target segment: high-frequency customers with narrow baskets. They already trust the platform and order often, but their purchases stay in familiar aisles.

Evidence, not assumptions

Instead of guessing where category expansion breaks, I combined public customer conversations, AI-assisted retrieval, structured behavioural tagging and human synthesis.

1. 01

   Collect

   ~3,776 public conversations: Play Store, App Store, YouTube, competitor reviews

2. 02

   Retrieve

   Custom engine finds discovery and consideration talk: 711 candidates

3. 03

   Classify

   AI tags stage, barrier, JTBD, category and confidence: 38 signals

4. 04

   Synthesise

   Human review of high-confidence signals: 3 opportunity areas

The classifier agreed with 50 human labels 92% of the time (Cohen's κ = 0.89). The schema allows "unknown" when evidence is thin. Code is on GitHub ↗.

What the reviews showed

Only ~1.2% of conversations touch category discovery, but that thin signal is strikingly consistent. It lines up with three barriers: know ("I didn't know they sold this"), consider ("I know, but it's not where I'd buy it") and confidence ("I can see it, but I can't choose confidently").

Click to enlarge

Click to enlarge

Public data can see two of the three barriers. Consider barely shows up, because nobody reviews a platform they never thought of. That made it the priority question for interviews.

What interviews changed

Six in-depth interviews with high-frequency, narrow-basket customers in Bengaluru, Mumbai, Hyderabad and Delhi (~35 minutes each).

• 5 of 6

  Went elsewhere

  For their last non-grocery need, and 3 of them already knew Blinkit carried it.

• 4 of 6

  Underestimated the range

  Named fewer than half the non-grocery categories Blinkit sells.

• 4 of 6

  Needed confidence

  Wanted reviews, comparisons or return assurance before trying beauty or electronics.

Click to enlarge

The problem

High-frequency Blinkit customers order 4+ times a week, but their baskets stay in one or two categories. The moment a purchase needs reviews, comparison or return assurance, Blinkit isn't in the consideration set. It owns the urgent, low-decision trip, but it isn't yet the platform customers consider when a purchase takes thought.

Three options

• Option 1

  Awareness campaigns

  Push category ads. Solves know, but not consider: 3 of 6 already knew and still opened Amazon.

• Option 2

  Discovery rails

  Top picks and an AI next-product rail. Only reaches people already in the app, before consideration.

• Option 3 · Recommended

  Flash

  A situation-first AI layer. Describe the moment in one line and get a ready basket. Solves know, consider and confidence, with the highest build effort.

Nobody shops for a category; they shop for a situation. For products, Amazon is the default. For situations, there's no default yet, and that's the consideration set Blinkit can own.

The experience

Click to enlarge

Click to enlarge

• No searching. Describe a situation instead of hunting for products one by one.
• No decision fatigue. At most one clarifying question, then a complete basket.
• No forgotten items. Flash remembers what the situation needs.

Flash understands Hinglish too: "bacche ko bukhar hai", "my kid is burning up" and "fever 102 what do I do" all lead to the same fever kit. A child's fever leads to pharmacy, a new puppy to pet care, an exam tomorrow to stationery. The customer never picks a new category. The category comes with the solution.

How Flash works

Click to enlarge

• Grounded, not guessed

  Only real, in-stock SKUs. Responses follow a strict schema, so free text can't break the UI.

• Confidence-aware

  60%+ match builds the basket directly, 30 to 60% asks one question, under 30% shows suggestions.

• Safe by design

  Prompt-injection safe, medical caution, no upsell, no sensitive inference, and the customer always starts the conversation.

Cold start is easy: the grocery basket is already the signal. Shaving gel suggests razors, a thermometer suggests a fever kit. Built with Next.js, Supabase and the Claude API, deployed on Vercel.

Measuring success

1. 01

   Conversation successPilot target ≥70%

   Users get a relevant basket with 0 to 1 clarifications.

2. 02

   New-category buyersTarget +20%

   vs a control group of non-Flash users.

3. 03

   AOVTarget +₹125

   Incremental value per order.

4. 04

   Repeat Flash useTarget >30%

   Users come back to Flash.

5. 05

   AI qualityTargets

   Intent accuracy ≥90%, grounding ≥95%, 0% hallucinated SKUs, under 3 seconds.

All targets are projections for a pilot, not shipped results.

Roadmap

• 0 to 3 months

  Pilot

  10 high-intent situations, 5% of users in one city. Validate impact and unit economics.

• 4 to 9 months

  Expansion

  50+ situations, voice input, 25% of users across more cities.

• 10 to 12+ months

  Platform

  Flash becomes the default discovery layer, with personalisation and cross-category intelligence.