CASE STUDY · TRAVEL AND LUGGAGE
Provogue
4,045 shopper messages, and a query log the search box never saw.
4,045
Shopper messages
1,332
Sessions that engaged
6
Settled orders
₹171.61
Model spend, 90 days
The problem
Provogue relaunched into luggage as a fashion category and scaled fast across D2C and marketplaces. On its own storefront, shoppers were being handed a category tree when the questions they actually arrived with were constraint-shaped: a size, a budget, a feature. A filter cannot answer "cabin trolley bag under 2000", and a search box never hears it.
The intervention
- 01Deployed the on-site assistant against the live Shopify catalog, grounded at variant level so availability and price come from the store rather than from a stale index
- 02Routed the predictable turns - greetings, browse, sale, top-rated, show-me-more, checkout hand-off - through deterministic fast paths instead of the model
- 03Enabled in-chat cart and checkout so a shopper who decides in the conversation does not re-enter the funnel
- 04Added voice alongside chat, which shoppers used in English, Hinglish and Hindi
- 05Published the whole event stream to an analytics dashboard: seven-step funnel, query log, product engagement, model cost and error classes
The result
Over 90 days the assistant handled 4,045 shopper messages across 1,332 engaged sessions, produced 97 add-to-carts and 92 checkout starts, and recorded 6 settled orders. Total model spend across the same window was ₹171.61 across 1,115 model turns, comfortably smaller than the payment processing on the same orders. Roughly three thousand further turns never reached a model at all, because deterministic fast paths served them at zero cost and zero hallucination risk. These are recorded outcomes, not a measured lift: there was no holdout arm, so nothing here establishes what the assistant added. Order values are the merchant's to disclose, so this page publishes counts and not revenue.
Growth table
Metric
Before
After
Change
Sessions that sent a message
-
1,332
90 days
Ran a search
-
732
55.0% of engaged
Viewed a product
-
278
20.9% of engaged
Added to cart
-
97
7.3% of engaged
Started checkout
-
92
6.9% of engaged
Settled orders
-
6
0.5% of engaged
Category research
How luggage actually gets bought, according to 4,045 shopper messages
This is the part no keyword tool can give you. Site search only records queries a shopper already knew how to phrase in your vocabulary; the assistant records the rest. These are the most-asked queries across the 90-day window, taken from the deployment's own log.
Query
Times asked
Sessions
show me cabin luggage
237
224
more
230
119
show me medium luggage
151
136
show me your highest rated products
130
108
show me items on sale
104
91
show me large luggage
85
81
show me backpacks
71
66
show me cheaper options
67
66
show me luggage sets
56
55
show me the Lattitude collection
41
37
- 01
Size is the primary buying axis, not style or brand. Cabin, medium, large and sets together account for 529 asks - more than every other axis combined. Luggage is bought by cabin-fit first.
- 02
"more" is the second most-asked query in the whole log, at 230 asks across 119 sessions. Shoppers exhaust the first result set and ask for the next one. A first page of five products is not enough for this category.
- 03
Price pressure is explicit and constant: 104 asks for items on sale, 67 for cheaper options, and a steady tail of stated ceilings - "show me travel bags under 3000 rupees", "Cabin trolley bag under 2000". Shoppers say a budget out loud to an assistant that they never type into a search box.
- 04
Social proof behaves like a filter rather than a badge. 158 asks combined for highest-rated and top-rated products, ahead of every individual product line.
- 05
Some filters shoppers want do not exist as catalog fields at all: TSA locks, and capacity in litres ("45 Litar capacity bags", "Back bag 35lites"). Until those are structured attributes, neither an on-site assistant nor ChatGPT can answer them.
- 06
Post-sale questions arrive in a pre-sale tool - "Any order tracking link", "Bag is locked". Shoppers do not know the scope of the thing they are talking to, which is a scoping decision to make deliberately rather than by accident.
- 07
Voice was used in three languages across the window: 1,171 sessions in English, 54 in Hinglish, 3 in Hindi. Hinglish is not a rounding error in this market.
Live, right now
Try the assistant on provogue.com
The deployment in this case study is public. Open the storefront, use the assistant in the corner, and ask it something hard - a budget ceiling, a TSA lock, a size that is sold out. Everything on this page came out of that same system.
www.provogue.comWhat was deployed
The products behind this engagement
AI Shopping Assistant
The on-site AI shopping assistant itself: catalog-grounded at variant level, chat and voice, with in-chat cart and checkout so a shopper who decides in the conversation never re-enters the funnel.
MCP deployment
The same enriched catalog is exposed to external agents over MCP. Provogue's endpoint is live, so the product data the on-site assistant reads is readable by off-site agents on the same terms - one enrichment, two surfaces.
Agentic checkout
Checkout completes inside the conversation rather than bouncing the shopper to a cart page, and the merchant stays the merchant of record.
The full document · 7 pages · PDF
Provogue - the full engagement brief
Every funnel step, the model bill, the error classes, and the query log that shows how luggage actually gets bought.
- All six funnel steps from first message to settled order, with the denominators named
- What 90 days of model spend actually came to, and why most turns never reach a model
- The category research: the ten most-asked queries and what they say about buying luggage
- The three instrumentation faults we found in our own dashboard, published rather than hidden
In their words
4,045
Shopper messages in 90 days · Provogue
“It handles the questions our filters cannot - a budget ceiling, a TSA lock, a capacity in litres - and it comes back fast enough that shoppers stay in the conversation instead of dropping to the category page. Engagement is clearly up. On conversion we are early and still measuring, so we are reading the whole funnel rather than one number.”
Atul Gosavi
VP - Strategy, Growth & Business Transformation · Provogue
Key lessons
Lesson 01
Most shopper turns are deterministic problems wearing conversational clothes - roughly 3,000 of them were served by routing rules at zero model cost, and a fast path cannot hallucinate
Lesson 02
Model spend is not the constraint on this class of product: ₹171.61 over 90 days was a third of the payment processing on the same orders
Lesson 03
The query log is the asset the category does not talk about - budgets, feature filters and sizing language that never reach a search box
Lesson 04
Publish the instrumentation faults with the numbers. Chat-open fires on a minority of entry paths and voice call-length capture is wrong; both are named openly rather than hidden behind a headline rate
ROI summary
Investment
On-site AI shopping assistant on the live Shopify catalog, chat and voice, with in-chat checkout
Outcome
4,045 shopper messages and a first-party query log, on ₹171.61 of model spend
Window
90 days
The lesson
The most valuable output was not the orders. It was the query log. Shoppers stated budgets out loud - "show me travel bags under 3000 rupees", "Cabin trolley bag under 2000" - asked for features that have to exist as catalog fields before anything can filter on them, TSA locks and litre capacity among them, and used sizing language the taxonomy does not contain, like "any size smaller than cabin size". None of that appears in a site-search log. It is a merchandising input that no other system on the storefront produces.