Checkout Dark Pattern Detector for South Asian Online Stores and Quick Commerce Sites

The problem: the price you see is not the price you pay

Online shopping, food delivery and quick commerce have grown fast across Pakistan, India and Bangladesh. So have tricks that push people into paying more than they meant to. A small "service fee" and "platform fee" appear only on the last screen. A donation or insurance add-on is ticked by default. A timer says the deal ends in 10 minutes, and resets when you reload the page. The button to skip an offer says "No thanks, I don't like saving money".

Each trick is small. Together they cost shoppers money, and they hit hardest the people who are new to online shopping and trust the screen. They also hurt honest stores, which look more expensive on the first screen because they show the full price up front.

What already exists and the gap

• Research crawls such as the Princeton study of 11,000 shopping websites found dark patterns at scale, but mostly on product pages, and on English-language sites.
• Regulation is catching up. India's Central Consumer Protection Authority issued the Guidelines for Prevention and Regulation of Dark Patterns in November 2023, listing 13 specified patterns including false urgency, basket sneaking, confirm shaming and drip pricing. Checking compliance still depends on manual complaints.
• Consumer-side tools are mostly ad blockers and price trackers. They do not detect manipulation inside the checkout.

The gap is an automated tool that reaches the checkout, tracks the price across steps, and understands pressure messages in local languages, including Roman Urdu.

Inspiration

"Dark Patterns at Scale: Findings from a Crawl of 11K Shopping Websites" by Mathur and colleagues (Proceedings of the ACM on Human-Computer Interaction, CSCW 2019) used a crawler and clustering of page text to find 1,818 instances across 15 types of dark pattern. Their code and data are public. Your project builds on that method but goes further into the flow, where drip pricing and basket sneaking happen, and adds the languages South Asian shoppers actually read.

What you build

• Consumer groups and journalists run the crawler across a sector and get a report of which stores use which patterns, with evidence.
• Store owners and agencies test their own checkout before launch and fix problems before customers complain.
• Shoppers (stretch goal) get an extension that flags a late fee or a pre-ticked add-on as it appears.

The technically interesting part

• Driving real flows safely. Every store's cart and checkout is different. Combine heuristics (find "add to cart", find the cart icon, find "checkout") with a small set of per-site hints, and always stop before any payment or order is placed. Use guest checkout where possible and never enter real card details.
• Price tracking. Extract the total at each step, normalise currency formats and digits (including Urdu numerals), and explain any increase line by line.
• Testing timers and scarcity. Load the same page several times over a period, with different sessions, and check whether "only 2 left" and countdowns behave consistently.
• Multilingual classification. Build a labelled set of pressure and shaming messages in English, Urdu script and Roman Urdu, and train a small classifier. Roman Urdu spelling varies a lot, so test how well your model handles it.
• Evidence quality. A finding must be reproducible. Save screenshots, DOM snapshots and timestamps for every claim.

How to evaluate it

• Flow success: share of stores where the driver reached the last step before payment.
• Detection accuracy: precision and recall per pattern, against a hand-labelled sample of stores.
• Classifier quality: F1 on held-out Urdu and Roman Urdu messages, compared with a keyword baseline.
• Stability: rerun the study after a month and report which findings persisted.

Suggested plan

• Weeks 1 to 4: read the paper and the guidelines, build the taxonomy, collect 20 example flows by hand.
• Weeks 5 to 10: flow driver, safety stops and price tracking.
• Weeks 11 to 15: basket, timer and scarcity checks, evidence packs.
• Weeks 16 to 20: labelled message set and classifier.
• Weeks 21 to 26: the 60-store study, dashboard, write-up and demo.

Path to a product

Consumer protection bodies, compliance teams and agencies in India, Pakistan and beyond need evidence, not anecdotes. A monitoring service with monthly reports, and a pre-launch checkout audit for agencies, are realistic offers. The study itself is publishable and makes a distinctive portfolio piece.

Scope warning

Never complete a purchase, never use real payment details and respect each site's terms and rate limits. Describe what you observed rather than calling a store illegal. Start with five patterns you can detect reliably; a smaller, accurate study is better than a broad one full of false alarms.

Learn what you need on Sefism

• Machine learning roadmap: building and evaluating the text classifier.
• Speech and Language Processing: text classification for the Roman Urdu messages.
• Playwright Documentation: driving checkout flows and saving evidence.
• scikit-learn User Guide: baselines and evaluation metrics.
• UI/UX design roadmap: what honest checkout design looks like, so you can describe the fix.

Get help with this idea

The risk is a crawler that breaks on every second store, or claims that do not survive scrutiny. Tauseef can help you narrow the patterns, choose the stores and design the labelled sets so your findings hold up. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session