Interpreting analysis results

#First ask what a number represents

Different figures do not automatically mean one report is wrong. Check scope, date basis, time zone, currency, row grain, status, and freshness before expecting equality. AI should distinguish retrieved facts, hypotheses based on them, and causes that still need verification.

#The same date can describe different business events

For cross-Shop comparison, specify local calendar days or one common time interval. Clarify whether “last seven days” includes the unfinished current day. Do not directly compare part of today with a complete day.

#Avoid double-counting

Orders, SKUs, sample relationships, and advertising creatives have different row meanings. An order containing two products, attributed to a Creator and to advertising, does not represent four separate sales.

• Do not sum product order counts into a Shop order count.
• Do not add Affiliate platform GMV to sample-conversion GMV.
• Deduplicate order lines linked to multiple Sample Applications for unique sales totals.
• Do not repeat request amounts on each aftersales product line.
• Count a shared WMS warehouse once, not once per connected Shop.

Ask for raw row counts, deduplication rules, and summarized counts when reconciling results.

#Distinguish amounts, statuses, and rates

Effective GMV excludes canceled orders but is not net income after refunds and fees. Buyer payments, Affiliate commissions, statements, and seller payouts are different amounts. A refund request does not mean the refund is completed.

Do not add different currencies. Before converting, agree on exchange-rate source, date, and method, retaining original amounts.

ROI is generally a multiple; conversion percentages need denominators. “Ten orders paid after outreach” also requires the number actually reached. Satisfaction needs the number of rated sessions. Do not ignore sample size or replace missing values with 0%.

#Attribution is not incrementality

“Paid after outreach” describes timing and attribution, not proof that the buyer would otherwise never have paid. Evaluate additional sales through mature randomized control samples, separating experiment versions and observation windows.

Mature holdout comparison includes natural payments and orders that received no message. Do not restrict the comparison to sent orders. Use Unpaid-order outreach for full A/B/C comparisons and formal experiment conclusions; ordinary attribution tables do not establish significance or causality.

Ads and Affiliate attribution are not automatically incremental Shop revenue either. AI can suggest explanations, but “possibly influenced by advertising changes” and “proven to be caused by advertising” are different claims.

#Blank, incomplete, and zero are different

• Zero: an observed zero within covered data and the selected filters.
• Blank: the field was not supplied, such as anonymous identity or an unsupported WMS quantity.
• Unready or incomplete history: data is unavailable, not evidence that no business occurred.
• No access: the current identity lacks permission; another chat entry point does not grant it.

Authorization does not guarantee all platform history has synchronized. Explicitly request 1–5 stars for an overall review analysis: the default 1–3-star focus does not represent the whole distribution.

#Current observations versus historical reports

Current WMS stock and SPS can be analyzed alongside historical sales, but timestamp each separately. Do not substitute today's stock for last month's inventory or interpret a single SPS metric's diagnostic trend as daily historical overall scores.

Historical daily summaries can lag realtime queues. Platforms can also revise commissions, refunds, and fulfillment outcomes. Preserve report generation time, scope, and raw detail for important reconciliations, then check these conditions before reproducing a result. See Reports and exports.