GMV Dropped: A 30-Minute Variance Decomposition Method

The identity that makes a drop answerable

Sales on a marketplace reduce to three multipliers: sessions, units per session, and average selling price. Availability and the featured offer act on the middle term. Once you accept that, a sales drop stops being a debate and becomes arithmetic.

The discipline is to decompose before you diagnose. Teams that skip the step spend a week rewriting creative when the real cause was an offer they lost 18 points of. The decomposition takes 30 minutes and it removes the loudest opinion in the room from the decision.

The reports you need, and their lag

On Amazon India the spine is the business report for detail page sales and traffic by child ASIN. It gives sessions, page views, featured offer percentage, units ordered, unit session percentage and ordered product sales, at daily granularity, generally available the next day and retained for around two years. Four sources sit around it.

• Brand analytics for category demand and click share, weekly, so you can separate a you problem from a category problem.
• The inventory ledger for out of stock and stranded hours, which explains conversion gaps that look like listing problems.
• Advertising reports for paid clicks, so sessions can be split into paid and the rest.
• The fee preview or settlement file for realised price after fees, because the ASP on a sales report is not what you banked.

Write the lag beside each source inside the dashboard itself. Half of all reporting arguments are two people comparing a next day number against one that settles three days later.

A worked decomposition

Take a realistically shaped example. Week on week, ordered product sales fall from 42.6 lakh to 34.9 lakh, down 18.1 percent. The three multipliers move like this.

• Sessions: 96,400 to 92,500, down 4.0 percent.
• Unit session percentage: 12.1 to 10.4, down 14.0 percent in relative terms.
• Average selling price: Rs 1,148 to Rs 1,152, up 0.3 percent.

Multiply the ratios. 0.960 times 0.860 times 1.003 gives 0.828, which is down 17.2 percent. That accounts for nearly all of the observed 18.1 percent, leaving about a point to mix and multi unit orders. Now go one level down on the weakest term. Featured offer percentage moved from 96 percent to 78 percent across the same week.

The answer is no longer arguable. Traffic held. Price held. Conversion collapsed because you lost the featured offer for roughly a fifth of the week. The action is a pricing and offer fix, live the same day, and not a creative refresh scheduled for next month.

Two rules keep the method honest. Never read a percentage without its denominator, because a conversion rate on 300 sessions moves on noise alone. And always check that your three ratios multiply back to the observed change. When the residual runs above three or four points, something outside the identity moved, usually assortment mix or a variation that went out of stock, and you keep digging rather than declaring an answer.

Thresholds that trigger a look

Rules stop a weekly review turning into a scroll through charts. Set five and let them do the alerting for you.

• Featured offer percentage below 90 percent for two consecutive days on any top 50 SKU.
• Unit session percentage down more than 12 percent relative, week on week, at child ASIN level.
• Page views per session above 2.5, which usually signals variation sprawl or a gallery that forces hunting.
• Sessions down more than 10 percent while category demand is flat, which points at organic rank or paid share.
• Out of stock hours above 2 percent of the week across the top 50 SKUs.

When several fire at once, rank them by contribution to the variance and act on the largest only. Fixing the third largest driver first is how a quarter disappears while everyone stays busy.

Flipkart, with less instrumentation

Flipkart gives you product views and units with roughly a day of lag and no direct equivalent of featured offer percentage. Build the proxy yourself: a daily check on the top 50 product identifiers recording who holds the default offer and at what price, stored in your own sheet. After 30 days that series answers most of your pricing questions.

Reconcile units to the settlement report monthly, never daily. Ordered, shipped and settled will not agree on the same day and they are not meant to. Use ordered for demand diagnosis and settled for money questions, and keep them in separate charts with separate owners.

The weekly review that stays 30 minutes

Instrumentation is rarely the constraint. Discipline is. Keep a one page metric dictionary naming the exact report and field behind every number, and freeze those definitions for a full quarter so trends remain comparable when someone improves the pipeline mid quarter.

Bring four numbers to the review: sessions, unit session percentage, featured offer percentage, and realised ASP, each against the prior week with the decomposition already done. End with written decisions, each carrying an owner and a date. A dashboard with 40 tiles that produces no decisions costs more than no dashboard at all, because somebody is paid to maintain it.