ChatGPT cannot see a trustworthy live picture of Pakistani ecommerce demand unless you provide current evidence or use a configured browsing/data connection. Its useful job is to normalize observations, reveal weak assumptions and propose tests—not declare a “winning product.”
Build a Dated Evidence Packet
Collect 20–30 observations across a defined window. For each, save the platform, category, listing URL, capture date, normalized unit price, visible offer, review/count context if shown, stock or delivery signal if shown, and the exact customer problem implied. Add Google Trends only as relative search-interest evidence; it is not sales volume.
Before prompting, remove seller/customer identities that are not necessary. Then label every row, for example D12 or S07.
Analyze only the supplied, dated observations for category-screening.
For every conclusion cite row IDs. Separate: observation, plausible explanation,
missing evidence, and cheapest validation test. Never convert reviews, rankings,
search interest or social engagement into sales. Do not call anything a winner.
Ask for three outputs: repeated problem clusters, comparison gaps that make the data misleading, and a shortlist ranked by testability rather than predicted profit.
Worked Example
A learner captures listings for insulated bottles, desk organizers and rechargeable lamps. The first AI answer ranks lamps highest because several posts have large engagement. That violates the evidence boundary: social engagement is not paid demand, and rechargeable products add defect, warranty and shipping risks.
After correction, the model identifies a narrower hypothesis: desk organizers have comparable bundles, locally sourceable samples and low breakage risk. It cites eight rows but notes missing return and conversion evidence. The next action is not 100 units; it is three samples, a landed-cost calculation and a capped listing test.
Failure Cases to Diagnose
- Asking “What is trending in Pakistan?” without a date or evidence packet.
- Mixing single-item and multipack prices.
- Treating sponsored placement as organic demand.
- Ignoring regulated, fragile, perishable or warranty-heavy risk.
- Letting the model cite sources you never supplied or verified.
🇵🇰 Pakistan Angle
Test regional and seasonal explanations separately. Search behavior in Karachi during heat, Lahore during wedding season, or nationwide near Eid can shift, but an attractive story is not evidence. Also check whether an imported item’s replacement parts, labeling, warranty and currency exposure make the apparent trend unworkable.
Hands-On Exercise
- Define one category question and seven-day capture window.
- Collect 20 labeled observations from at least two legitimate sources.
- Normalize units, bundles and dates.
- Run the bounded prompt and audit each citation.
- Choose one category based on testability and write a capped validation plan.
Completion Rubric
- The evidence packet is dated and row-labeled.
- Sales are not inferred from proxy metrics.
- AI conclusions cite only supplied rows.
- Risk and missing evidence affect the shortlist.
- The output ends in a small test, not a purchase order.
Sources
- Google Trends Help — FAQ about Google Trends data
- Daraz Pakistan — Start Selling on Daraz
- Shopify Help Center — Analytics
Key takeaway: AI can compare a current evidence packet; it cannot turn marketplace proxies into guaranteed product demand.