Product Review & Ratings Data Across Six Indian E-commerce & Quick-Commerce Platforms

The Client

A consumer brand operating across India's beauty, personal-care, and packaged-goods categories — selling through the full spread of platforms Indian shoppers actually use: Flipkart and Nykaa and Purplle for considered purchases, BigBasket and Blinkit and Zepto for the everyday and the instant. Their products lived on six platforms; their understanding of what customers thought of those products lived nowhere, in aggregate. Each platform showed reviews on its own pages, in its own format, and nobody on the brand's side had a unified, structured view of the voice of their customer across the places that voice was actually being expressed.

They came to Actowiz Solutions for one thing: all the review and ratings data for their products (and key competitors') across those six platforms, in one clean structured dataset — deduplicated, normalised, and ready for the sentiment and theme analysis their insights team wanted to run.

The Challenge

Review data across six diverse platforms is a harder problem than it appears:

• Six platforms, six review structures. Flipkart's reviews carry ratings, titles, verified-purchase flags, helpful votes, and structured attributes; Nykaa and Purplle (beauty-native) carry skin-type and concern tags alongside the review; BigBasket, Blinkit, and Zepto (grocery and instant) carry lighter, higher-velocity review structures. Extracting a consistent review record across these very different formats — while preserving the platform-specific richness (beauty tags matter enormously for a beauty brand) — is the core design challenge.
• Multilingual and code-mixed reviews. Indian reviews arrive in English, Hindi, Hinglish, and regional languages, frequently code-mixed within a single review ("product bahut accha hai, but delivery was late"). Any sentiment or theme analysis downstream depends on handling this correctly — and treating a Hinglish review as noise, or mis-processing it, throws away a huge share of the actual signal in the Indian market.
• Volume, pagination, and completeness. Popular products carry thousands of reviews across pages; collecting them completely (not just the first page the platform shows) while deduplicating is essential — a review dataset that silently captures only recent or "top" reviews gives a skewed picture of sentiment.
• Reviews contain personal data. Reviewer names, profile references, and sometimes personal details in review text are PII — and under DPDP, handling this correctly (masking reviewer identity at the edge, retaining only the commercial signal) is both a compliance requirement and, frankly, all the brand actually needs.
• Fake and incentivised review patterns. Review streams contain incentivised reviews, template spam, and increasingly AI-generated text — noise that pollutes sentiment analysis if not filtered, a problem we handle as a standard curation stage.
• Quick-commerce reviews are a newer, different signal. Blinkit, Zepto, and Instamart reviews skew toward delivery experience and freshness rather than deep product assessment — a different (and, for grocery, highly valuable) signal that has to be understood distinctly from a considered Nykaa beauty review.

The Actowiz Solution

1. Unified review schema across six platforms.

Per-platform extraction feeding one normalised schema: product reference, platform, rating, review title and body, review date, verified-purchase flag, helpful votes, and platform-specific enrichment fields (beauty tags from Nykaa/Purplle, delivery-experience signals from q-commerce) preserved in typed extensions — so the common structure enables cross-platform analysis while the platform-specific richness isn't flattened away.

2. Multilingual and code-mixed handling.

Language identification per review (including Hinglish and code-mixed tagging), with processing that preserves meaning across English, Hindi, Hinglish, and regional languages — the multilingual depth from our regional-language work, applied to the review-sentiment use case. Original text retained; language tagged for downstream analysis.

3. Complete, deduplicated collection.

Full review collection across pagination (not just first-page or "top" reviews), with deduplication within and across platforms, so the dataset represents the genuine distribution of customer sentiment rather than a skewed sample.

4. PII masking at the edge.

Reviewer identity masked during collection — names, profile references, and incidental personal details in text handled per DPDP — retaining the commercial signal (rating, text themes, sentiment, verified flag) and nothing that creates personal-data liability. This is both compliant and sufficient: the brand needs the voice, not the identity.

5. Noise filtering.

Incentivised-review patterns, template spam, and AI-generated text flagged and filtered as a curation stage, so the delivered dataset is sentiment-analysis-ready rather than polluted.

6. Sentiment-ready structuring.

Reviews delivered structured for the client's analysis — clean text, language tags, ratings, and platform context — with an optional theme-and-sentiment enrichment layer (aspect-level: product quality, delivery, value, packaging) that decomposes the voice into the dimensions the insights team wanted rather than a single star average.

7. Delivery and cadence.

Delivered in the client's structured format on a recurring cadence with historical retention (so sentiment trends over time became visible), per-record lineage, and DPDP-mapped compliance documentation.

The Outcome

The brand's insights team got, for the first time, a single unified view of their customer's voice across every platform their products sold on — and the cross-platform view immediately surfaced things single-platform browsing had hidden. The same serum praised for results on Nykaa was criticised for packaging on Flipkart and delivery on q-commerce — three different truths about one product, actionable by three different teams (formulation held steady, packaging flagged for redesign, q-commerce fulfilment raised with partners). The aspect-level sentiment decomposition meant "4.2 stars" became "loved for texture, dinged for a fragile dropper" — a product-development instruction rather than a vanity metric.

The multilingual handling proved its worth in the Indian context specifically: a large share of the most detailed, most useful reviews were Hinglish and code-mixed, and a pipeline that mishandled them would have discarded exactly the richest signal. The q-commerce reviews, distinct in character, gave the brand its first structured read on delivery-and-freshness perception — the axis that increasingly decides repeat purchase in instant commerce.

The engagement continues as a recurring feed, with sentiment trends now tracked over time (did the packaging redesign move the packaging-complaint rate?) and the panel expanding to more products and competitors — the voice-of-customer becoming a monitored metric rather than an occasional manual audit.

Why This Pattern Repeats

Every multi-platform consumer brand faces the same blind spot: their customer's voice is fragmented across the platforms they sell on, in multiple languages, in incompatible formats, laced with PII and noise. The transferable design: a unified-but-extensible review schema across platforms, genuine multilingual and code-mixed handling, complete deduplicated collection, PII masking at the edge, noise filtering, and aspect-level sentiment structuring. The unified, clean, compliant voice-of-customer is the asset — and in India specifically, the multilingual layer is what makes it real.

Frequently Asked Questions

Can reviews be collected across very different platforms into one dataset?

Yes — a unified schema captures common review fields across all six platforms while preserving platform-specific richness (beauty tags from Nykaa/Purplle, delivery signals from q-commerce) in typed extensions, enabling cross-platform analysis without flattening.

How are Hinglish and code-mixed reviews handled?

With language identification (including Hinglish/code-mixed tagging) and meaning-preserving processing — essential in India, where a large share of the most detailed reviews are code-mixed and would otherwise be lost as noise.

Is reviewer personal data collected?

No — reviewer identity is masked at the edge per DPDP, retaining only the commercial signal (rating, text, themes, verified flag). The brand needs the voice, not the identity.

Can sentiment and themes be delivered, not just raw reviews?

Yes — an optional enrichment layer decomposes reviews into aspect-level sentiment (quality, delivery, value, packaging), turning star averages into actionable product insight. Contact Actowiz Solutions to scope a review-data programme.