Rank Decay and the Half-Life of a Sale

Rank responds to a sale sharply and then drifts back. Understanding the drift explains most of the shapes you will see in a rank chart.

The behaviour

A sale produces an immediate improvement, visible at the next refresh.

With no further sales, rank worsens steadily. The improvement decays.

The decay is not linear. It is faster immediately after the sale and slower afterwards, which is what a weighted recency calculation produces.

Steady sales produce a stable rank. Equilibrium is reached where the decay is offset by new sales.

This is why a flat rank line means constant sales, not zero sales. A product with no sales does not hold a flat line; it drifts downward continuously.

Estimating the decay rate

You can measure this from your own data without any inside knowledge.

Find a product with a clean isolated spike — a promotion, a mention, one busy day followed by quiet ones.

Plot the recovery. How long until the rank returns to its pre-event level.

Repeat across several events and several products in the same category.

The pattern is category-specific, and knowing yours is worth the afternoon it takes.

Typical observations put the meaningful influence of a sale at somewhere between one and several days, with faster decay in high-velocity categories. Treat any specific figure as an estimate for your category, not a constant.

What the decay explains

The sawtooth pattern in low-volume products. Sale, sharp improvement, gradual decay, sale, repeat. The teeth are individual sales.

Why a slow product's rank looks volatile while a fast product's looks smooth. The fast product's sales overlap continuously; the slow one's do not.

Why promotional spikes always look bigger than the sustained effect. The peak reflects a burst; the new equilibrium reflects the sustained rate, and the two are very different numbers.

Why post-campaign analysis run too early overstates the result. Measuring three days after a campaign captures decay in progress rather than the new baseline.

The measurement this enables

Wait for equilibrium before judging a campaign. The relevant question is not the peak but the level the product settles at afterwards, compared to before.

Allow at least a week of quiet after the event, longer in slow categories.

Compare pre-period and post-period baselines, both measured over several days, rather than comparing a peak to anything.

A campaign that produced a large spike and no baseline change generated a burst of sales from people who would probably have bought anyway, or from an audience that did not persist. That is a real and common finding and it is invisible if you only look at the peak.

Reading a sawtooth

For products in the long tail, the chart is a series of teeth and each tooth is approximately one sale.

Counting teeth over a period gives you an approximate sales count, which is a more direct estimate than any conversion table for slow products.

Tooth height indicates how deep the catalogue is around you. Large jumps mean few products with similar velocity nearby.

Tooth frequency is the sales rate.

This is the one place where rank data gives you something close to a unit count, and it only works for slow-moving products where sales are separable.

The equilibrium framing

The useful mental model is a level that settles where new sales balance decay.

Rank measures a rate, not a total. A product that sold a million copies over ten years and nothing this month ranks poorly. A product that sold forty units this week ranks well.

This is why bestseller lists reflect current velocity rather than cumulative success, and why a rank chart is a picture of momentum.

And why "our book hit rank 300" is a statement about one week rather than about the book.

Using decay to estimate slow-product sales

For long-tail products, counting sawtooth teeth is more direct than any conversion table.

Confirm the teeth are separable. If improvements overlap, the product is too fast for this method.

Count distinct improvement events over a period. Each is approximately one sale, or one order of possibly several units.

Divide by the period for a sales rate.

The error is real: multi-unit orders count as one tooth, and simultaneous sales merge. This biases the estimate downward, so treat it as a lower bound.

It requires sub-daily sampling. Daily observation misses teeth entirely on a product selling every few days.

Where it works, it beats a conversion table by a wide margin, because it observes the events rather than inferring them from a fitted curve built on someone else's category.