How Online Retailers Use ChatGPT for Product Copy

Where ChatGPT Sits in the Catalog Loop

The common setup in Petaling Jaya and Shah Alam is not one-off chatting. Sellers paste two or three of their best-ranked listings into a fresh chat as few-shot examples, then feed the next SKU’s raw data from Shopee Seller Centre. A phone-accessory seller with 180 SKUs runs this manually: copy, paste, edit, approve. The moment a catalog passes 500 SKUs, that loop breaks.

Mid-size sellers move to the OpenAI API through Make.com scenarios or a plain Python script. The flow: export a CSV from Seller Centre, map the columns (product name, commodity specs, supplier description, material, weight), and push each row as a user message to GPT-4o mini. Output comes back structured — `title_ms`, `title_cn`, `title_en`, `desc_ms` — ready to re-import. The unit economics matter: GPT-4o mini charges US$0.15 per million input tokens and US$0.60 per million output tokens. A 320-token description costs about US$0.00019, roughly RM 0.0009. A 2,000-SKU refresh comes to under RM 2, and the batch endpoint discounts that another 50%.

Platform Limits: Shopee, Lazada, TikTok Shop

Each marketplace punishes copy that ignores its character ceilings. Shopee Malaysia trims titles at 60 characters; Lazada allows 120; TikTok Shop gives 255. Sellers bake these numbers into the system prompt, along with a rule that no title repeats the same keyword twice — Shopee ranks keyword stuffing down. The practical prompt line is blunt: “Title must be under 60 characters, include color, key variant, and brand suffix. If the result exceeds 60, retry.”

Lazada additionally pushes bullet-point-style descriptions for its mobile layout. The better prompts generate a 5-line bullet block in Bahasa Malaysia first, then reorder it for Chinese interfaces. Doing this by hand on a 1,000-item catalog is a two-week job. With a structured-output prompt, it is a 10-minute run.

Trilingual Prompts: Bahasa Malaysia, Chinese, English

Localised copy means more than translation. “65W fast charging” becomes “Pengecasan pantas 65W” in Malaysian Bahasa Malaysia — not the Indonesian “Pengecasan cepat 65W” that cheap translation tools emit. Chinese listings need “65W快充” plus the simplified-character preference used on Lazada’s Chinese backend. Sellers lock the approved terms into a Custom GPT or into the API system prompt: no Indonesian loanwords, no traditional characters unless specified.

One kitchenware seller in Shah Alam uses a glossary of 40 translated material and finish terms so “high-grade stainless steel” consistently renders as “keluli tahan karat gred tinggi” across every SKU. Without that glossary, ChatGPT picks synonyms at random, and search queries that matched yesterday stop matching today.

The API Batch Route for 1,000-SKU Catalogs

OpenAI’s batch endpoint is the difference between a hobby and a workflow. Sellers submit a JSONL file where each line holds a system prompt, the SKU data, and a `response_format` JSON schema defining exactly which fields come back. The job processes overnight, at half the synchronous price. A KL-based houseware seller runs this every Monday at midnight via a cron job; by 8 a.m. a copy lead reviews the output.

The same route enables A/B copy at scale. Instead of one description, the schema requests version A (spec-first) and version B (benefit-first) for the same SKU. Variances in tone are actually measurable: one batch put B variants through TikTok Shop ads and kept the higher CTR version on the catalogue page.

Review Gates and Hallucination Catching

ChatGPT will invent specs if the prompt allows it. Common failures: fabricating an IPX8 waterproof rating, adding “Made in Japan” to a product assembled in Puchong, or writing “Powerbank 20000” when the SKU is 10000 mAh. Sellers who skip review gates eventually get a product return spree or a Shopee policy warning.

The working pattern: the system prompt restricts the model to values present in the input spec sheet and adds “Do not introduce attributes absent from the supplied data.” A human still spot-checks the first 10% of every batch against the supplier’s datasheet columns. Retailers handling FAMA-related food claims or SIRIM-certified electronics keep humans in the loop permanently — the model has no idea whether a certification exists.