Repricing Rules That Survive Peak Season

Why Repricing Rules Break in Peak Week

A repricing rule is a bet on its inputs: that the competitor price it just read is current, that the competitor can actually sell the item, that your cost figure is the real one, and that the last observation arrived. Peak week breaks several of those at once, and the rule keeps running anyway.

The failures are rarely dramatic. A rule matches an offer attached to nothing anyone can buy. A floor defined against a rival slides down as the rival slides. A feed goes quiet for forty minutes and the software, having no instruction for silence, reverts to a default. Each of those is a design decision nobody made rather than a bug.

What follows is the set of guardrails that keeps a rule inside sane bounds, the order those rules have to run in, and the behaviour almost nobody specifies: what happens when the data does not arrive. The Black Friday price monitoring playbook covers the monitoring cadence and watch list that feed all of it.

What Actually Changes in Peak Week

The numbers below make a case about spread rather than depth: the extra discount points are modest, and what they do to the distance between categories is the part rules handle badly.

DataWeave tracked roughly 80,000 US SKUs across Amazon, Walmart, Target, Macy's, Home Depot and Sephora, comparing the pre-Black-Friday period through 23 November against Black Friday week, 24 November to 1 December 2025 (DataWeave, December 2025).

Category	Pre-peak average discount	Peak-week uplift	Peak-week total
Consumer electronics	14.6%	+2.6 pts	17.2%
Apparel	13.8%	+3.8 pts	17.6%
Home and furniture	13.9%	+1.7 pts	15.6%
Health and beauty	7.1%	+4.8 pts	11.9%
Grocery	5.2%	+1.5 pts	6.7%

Read the absolute column and peak week looks tame: between 1.5 and 4.8 extra points. Read it against each category's own baseline and it does not. Health and beauty discounting rises about 68% above its own normal, while electronics rises under 18%. A single threshold expressed in percentage points, applied across a mixed catalogue, misses real moves in grocery, where the entire event is worth 1.5 points, and fires constantly in electronics, where that much movement is an ordinary Tuesday.

The gap between the two measurement styles is what rule design has to absorb. Adobe recorded peak discount rates of 30.9% in electronics, 29.6% in toys and 25.1% in apparel for the 2025 season (Adobe Analytics, January 2026). Those are peak-day rates; DataWeave's 17.2% is a period average over the same category. Adobe's peak-day rate sits 13.7 points above DataWeave's week-long average for a comparable category, a rough measure of how far the deepest day runs from the week's mean. The two figures come from different panels, Adobe covering US online transactions and DataWeave about 80,000 SKUs at six retailers, so read it as an order of magnitude rather than a measurement. Rules calibrated on weekly averages meet that gap on the Friday morning and behave as though something unprecedented has happened.

Two Failures Worth Knowing by Name

Two documented incidents cover both directions a rule can run away.

The loop with no ceiling, April 2011

Two sellers listed the same used copy of "The Making of a Fly". One priced at 0.9983 times the other's price. The other priced at 1.270589 times the first. Each cycle multiplied the pair by just under 27%, once a day, with nothing to stop it, until the listing reached USD 23,698,655.93 (The Register, April 2011).

Nobody wrote a rule saying "charge twenty-three million dollars". Two reasonable relative rules pointed at each other with no absolute bound anywhere in the system. That is the argument for an explicit ceiling, and inverted, for a floor that is not relative either.

The failure state that priced to the bottom, December 2014

On the evening of 12 December 2014, a fault in the RepricerExpress service dropped Amazon UK listings to GBP 0.01. Around 2,000 sellers were affected in roughly an hour, and one merchant reported losing GBP 100,000 of inventory before it was contained (Washington Post, December 2014).

What matters is what the software did when it lost confidence in its own state: it priced to the bottom. A system that held its last known price instead would have produced a boring hour of slightly stale prices. This is the most consequential line in a rule set and the one that almost never gets written down.

Eight Guardrails Your Rule Set Needs

Start with the failure state, because everything else assumes the data arrived.

1. On stale or failed data, hold the price

Define what happens when an observation is missing, older than your freshness limit, or fails a sanity check. The correct default is to keep the last published price and raise an alert, never to fall back to the floor, to a competitor default, or to a "safe" aggressive rule.

Set the freshness limit relative to how often your data refreshes: if monitoring runs hourly, an observation from three hours ago is a memory rather than evidence. Refusing to act on it costs one cycle; acting on it in peak week can cost the day.

2. Compute the floor from landed cost, never from a competitor

Landed unit cost, plus referral and fulfilment fees, plus a returns provision, plus your target margin. That is the floor. A floor expressed as "lowest competitor minus 1%" has no bottom, because the competitor's floor is relative to yours: two of those facing each other is the 2011 loop running downward. Do the pre-peak floor-price work before October, where the arithmetic shows how much the returns provision alone moves the floor.

3. Set an explicit ceiling

Give every SKU a maximum price. The 2011 loop had no absolute bound anywhere, which is why two sane relative rules could multiply for days. A ceiling costs nothing while the rule behaves and is the only thing that stops it when the rule does not.

4. Drop unbuyable competitors from the reference set

An out-of-stock listing at a low price is not competition. Filter out-of-stock, unshippable and add-to-cart-disabled listings before the rule computes anything, and treat a stale stock flag as a failed observation under guardrail one.

There is one clean measurement of how far survivors move when supply disappears. Studying Amazon listings for masks and hand sanitiser in early 2020, Cabral and Xu found that when Amazon itself was out of stock, entrant sellers raised prices by 178% while continuing sellers raised by 56.7% (Cabral and Xu, Economic Inquiry, July 2021). That was a pandemic scarcity shock rather than a Black Friday, so read it as an upper bound on reference-set movement, not as a peak-season forecast. The direction is what your rule has to survive, and what a competitor stockout is actually worth covers valuing the opportunity rather than just avoiding the trap.

5. Cap net movement per SKU per day

Cap the net change per SKU per 24 hours, typically 5 to 10%, and require a human to release anything larger. Cap the net, not the individual adjustment: the 2011 loop was built from steps that were each plausible and compounded daily. A cap is no defence against 2014, where the drop to a penny was instantaneous and the software's degraded behaviour was to price to the bottom rather than to do nothing. That one belongs to guardrail one. The cap will bind more often in peak week, which is the point: that is when a runaway costs the most.

6. Cool down after every cut

After any downward move, hold that SKU still for 30 to 60 minutes. This kills tit-for-tat oscillation with a rival repricer reading you while you read it. Marketplace price changes also take time to surface: Walmart's own repricer documentation notes that a strategy-driven change can take up to four hours to appear on the item page (Walmart Developer, April 2026). A loop that re-evaluates faster than propagation is arguing with its own past self.

7. Make MAP a hard stop

If you sell brands with minimum advertised pricing, the MAP check is absolute: it blocks publication rather than adjusting the target. As of August 2026 there is no named, publicly documented case of a MAP breach traced to a specific automated repricing rule. The exposure is structural and the enforcement letters are real, but nobody has published the incident report, so treat the guardrail as insurance against a documented policy rather than a documented disaster. Pair it with MAP violation detection across your retailers to see a breach from the brand side too.

8. Count discount days against Amazon's 90-day window

From 18 May 2026, if more than half the days in a product's 90-day history sit below the non-promotional median, Amazon recalculates Typical Price using all sales including promotional ones (EcommerceBytes, April 2026). An always-on aggressive rule that holds a promotional price for more than 45 days in any 90-day window quietly destroys the strike-through it needs in November. Count the days as a rule input, not as a quarterly report.

The Order the Rules Run In

Every article about repricing lists guardrails. Almost none specify the sequence, and the sequence is what makes them binding. A MAP stop that runs before the Buy Box rule is not a stop, it is a suggestion the next rule overwrites.

A defensible order:

1. Freshness and failure gate. If the observation is missing, stale or implausible, hold the last price and exit. Nothing downstream runs.
2. Reference set construction. Drop out-of-stock and unbuyable listings, then compute the competitive picture from what remains.
3. Candidate price. Run the competitive rule: match, beat by a margin, target the Featured Offer, whatever your strategy is. Amazon-specific mechanics are covered in the Amazon repricing and Buy Box guide.
4. Clamp to floor and ceiling. Absolute bounds beat relative logic. If the candidate falls outside them, the bound wins.
5. Cooldown and movement cap. Reject or trim the move if the SKU moved recently or the daily budget is spent.
6. MAP hard stop. Last. Anything that ran earlier can be undone by a later rule, so the contractual constraint goes at the end of the chain.
7. Publish and log the inputs that produced the price, not just the price.

Decide two edge cases in advance rather than at 6am on the Friday. If MAP forces a move larger than your daily cap, publish and flag it, because a contractual obligation outranks a prudential limit. If MAP sits above your ceiling, or your floor sits above MAP, the rule has nothing valid to do: stop, alert a human, keep the last price. Silence is a safe answer. Guessing is not.

Test the Rule Set Before the Traffic Arrives

Rules are software, and software that has never failed in rehearsal will fail in production.

Run the rule set in shadow mode against recorded prices from last peak week and read the price path it would have produced, hour by hour. Look for three things: how often the movement cap binds, whether the floor was ever the active constraint for more than a few hours, and whether any SKU oscillated. A rule that spends peak week pinned to its floor has stopped pricing and started donating margin.

Then break the feed on purpose. Withhold observations for an hour on a test SKU and confirm the system holds its price and alerts instead of doing something clever. That one test would have prevented December 2014. The replay only works if the history is honest: a record that starts in November is already discounted throughout, which makes September the moment to fix monitoring depth. On the demand side, price elasticity work tells you which SKUs deserve an aggressive rule at all.

The Data Your Rules Are Only as Good As

Every guardrail above depends on observations that are fresh, honest about stock, and specific to the variation you actually sell. Respot tracks price and stock per variation, so a rival selling out of the one size that competes with your bestseller registers as a stockout instead of being averaged away at parent-product level. History depth is the other input: the free plan keeps 7 days across 5 trackers, enough to test a setup but not to run the 90-day discount-duration counter, while paid plans reach 30-day, 90-day or unlimited history (plans and pricing).

Decide Your Failure State Before November

If you change one thing this month, make it the answer to a single question: what does the rule do when the data does not arrive? Most rule sets have no answer, which means they have a default, and defaults chosen by accident are how a repricer ends up selling inventory for a penny.

Write the order down. Put MAP last. Compute the floor from your own costs. Then break your own feed on a test SKU and watch what happens, while the stakes are still a rehearsal. Start tracking the prices your rules depend on with enough runway to have a baseline worth trusting.

Frequently Asked Questions

Should I turn my repricer off during Black Friday?

Turning it off trades one risk for another: you stop the runaway loop and you also stop responding for the busiest week of the year. The better move is to keep the rules running with a daily movement cap, a cooldown after every cut, and a hold-on-failure state, so the worst case is a price that stands still rather than one that walks to the floor.

What is a safe maximum price change per day during peak week?

A net cap of 5 to 10 percent per SKU per 24 hours is defensible for most catalogues, with anything larger requiring a human to release it. The cap is on net movement, not per adjustment, because the 2011 runaway was built from steps that each looked reasonable and compounded. A cap does not cover the other failure mode: the 2014 penny incident was a single instantaneous drop by software whose degraded behaviour was to price to the bottom instead of doing nothing, and only a hold-on-failure state stops that one.

How do I stop a competitor's stockout from dragging my price the wrong way?

Exclude out-of-stock and unbuyable listings from the reference set before the rule computes anything. A stale in-stock flag imports a price nobody can transact at, which either holds your price down against a phantom offer or, once the flag flips, hands your rule a sudden jump it should not chase.

Will heavy Q4 discounting damage my Amazon strike-through price next season?

It can. From 18 May 2026, if more than half the days in a product's 90-day history sit below the non-promotional median, Amazon recalculates Typical Price using all sales including promotional ones. Holding a deep price for more than 45 days in any 90-day window can erase the reference price you wanted for November.