1688 P4P Ads Optimization Methodology — AI Operations
P4P ads optimization on 1688 is not about any single trick — it is about running a repeatable weekly routine across three levels: the whole store, each campaign, and each product. This guide lines them up: what to look at, in what order, and what decision each situation calls for.
Quick Start
Fix a weekly review slot and walk through in this order:
- Store first: is overall efficiency sliding along its own historical worst band? (Top-down, never the reverse.)
- Then campaigns: underperformers get five checks before any stop decision.
- Then products: money-burning products get six issue tags to decide keep, fix, or remove.
- Finally, add: screen organically selling products as promotion candidates.
One data rule governs everything: inquiry data is ready early; transaction data must settle. A re-collection experiment measured transaction recording running as late as day 29 after week end — see Is 16 Days Enough for Marketplace Ad Data?.
Walkthrough
Task 1: Store-level efficiency check
Rank the past ~26 weeks of store-wide ads ROI. The worst band (roughly the bottom fifth) is your "normal deviation" range. Several consecutive weeks dipping into it, with spend-per-revenue also worsening, means a store-level alert: look for the common cause before touching individual campaigns. New accounts with less than half a year of data should accumulate history first.
Task 2: Campaign keep-or-stop
Before stopping an underperforming campaign, run five checks: enough runtime? enough spend? continuous delivery? learning period granted? and is it truly zero-inquiry? The first four failures all mean "wait"; only accumulated spend with persistent zero inquiries means "stop now". See the full checklist on our blog.
Stop-and-go campaigns do not enter efficiency judgment at all — restore continuous delivery and collect comparable data first.
Task 3: Product keep-or-stop
Tag unproductive products with six issue labels: consecutive GMV decline, cost up with GMV down, zero add-to-cart, high bounce, bottom-tier impressions within the campaign, and dual-low conversion. One label means observe and fix; two or more stacked means a removal candidate.
Two special cases: store staples (top contributors to shop revenue) get downgraded to observation instead of immediate removal — the cost of a false positive is far higher. Already-paused products need a restart evaluation against four scenarios; unprofitable before pausing means stay paused.
Task 4: Refill the pipeline
Screen never-promoted products that already sell organically through three gates: enough distinct buyers, meaningful GMV, and a conversion rate no worse than your store's promoted-product median. All three pass, promote; any miss, wait. For products running in multiple campaigns, compare acquisition cost across campaigns every cycle and shift budget toward the winner.
Notes
Three data disciplines
① Judge only on settled, complete periods; ② do not refresh baselines or make removal decisions during abnormal periods (promotions, holidays); ③ every "normal level" comes from your own store's history, never industry averages.
FAQ
How long does the weekly review take?
Under an hour once familiar: a few minutes on the store level, campaigns only where something changed, tags only on new anomalies. Most weeks nothing needs doing — the routine exists so nothing escapes notice.
How do inquiry data and transaction data divide work?
Traffic capability is judged on inquiries (clicks, inquiries — stable within days). Money capability is judged on settled transactions and ROI. Mixing the two speeds is the main source of misjudgment.
Can this be automated?
The judgment rules can run automatically — our AI analytics produces weekly reports following exactly this logic — but stop and budget decisions deserve human confirmation. Automated diagnosis, human decisions: the stable division of labor.