Price Anomalies: Telling a Real Cut From a Glitch
Published on October 22, 2026 by Niccolò
The Alert You Believe Is the One That Costs You
A monitoring system that alerts on every large price move trains its owner to ignore it. Give it a month and the fake 40% drops arrive alongside the real ones, nobody opens either, and the setup meant to shorten your reaction time has quietly lengthened it.
Most extreme moves turn out to be something other than a competitive decision: a pricing error, a currency or unit mix-up, a regional price leaking into your feed, a variant swap on the page, a bot-served stub that never contained a product, or a promotion that only applies at checkout. Each draws the same shape on your chart as a competitor going to war on price.
The skill worth building is triage. Given a drop, you decide within a couple of minutes whether it is real, and therefore whether it deserves a response at all. That is a separate discipline from setting up competitor monitoring, and almost nobody documents it.
What follows: eight checks ordered cheapest first, five patterns you will meet repeatedly, how long bad prices survive, and where the legal position sits.
Nobody Has Measured How Often a Big Drop Is an Error
There is no published measurement of what share of extreme price moves are errors rather than decisions. No academic study, no trade body, no vendor with a large enough panel has put a number on it.
That absence matters. Every "alert me on drops over 25%" rule implies a belief about the base rate of errors at that magnitude, and that base rate is unknown. Everyone tuning a threshold is guessing, including the people selling you the tool.
What has been measured is how fast normal prices move. Alberto Cavallo found that price-change frequency at large multi-channel US retailers rose from about 15% of products per month in 2008 to 2010 to nearly 30% per month in 2014 to 2017, with average regular-price duration falling from 6.7 months to 3.7 (NBER, October 2018). A price change on its own carries less information than it did fifteen years ago.
So you cannot filter errors out with a threshold. Thresholds decide what reaches a human, which the guide to price drop alerts covers, but they work as a volume control rather than a truth test.
Two Kinds of Wrong, and Only One Is a Glitch
The first kind of wrong is an artifact of your collection method: the price you recorded was never shown to anyone. Your collector was blocked and handed a placeholder, parsed a strike-through reference price, followed a redirect to a different child SKU, or wrote a fallback value when parsing failed. Nothing happened in the market. Something happened in your pipeline, and whether you built that pipeline or bought it changes how you debug it, which is the practical half of the scraping versus monitoring API decision.
The second kind is a real price that is real only for someone else. Personalized pricing means a number your monitor sees and your browser does not can be entirely genuine. Hannak and colleagues found evidence of steering or price discrimination on 9 of the 16 e-commerce sites they measured (Northeastern University, 2014).
In a log, a glitch and a segment price look identical. Both are numbers your buyers may never see. The checks below separate them, and the response differs: you fix the first, and you decide whether the second is aimed at your customers.
Eight Checks, Cheapest and Most Decisive First
Run these in order and stop when one gives a clear answer. The sequence fits inside two minutes on a single listing.
1. Re-fetch the exact URL now, by a second method
Real if the same price returns on two independent fetches. Artifact if the price differs, or if the second fetch returns no product data at all.
This kills most bad readings, because the common failures are transient and method-specific. Use a genuinely different route: another client, another network path, or your own browser. Refreshing the same broken pipeline twice proves nothing.
2. Look for product signal in the raw response
Real if the response carries a product title, a SKU and an add-to-cart control. Artifact if you got an HTTP 200 containing challenge markup, an interstitial, or no product node.
A study of 10,000 sites across four browser configurations found a soft-block rate of 15% for headless Chromium against 7% for other configurations, and attributed 82% of blocks to bot detection (arXiv, June 2026). Be honest about its scope: the authors capture 403, 429 and 503 responses and state that content degradation inside HTTP 200 responses falls outside their observation. The 200-with-no-product-data case, the one that writes a fake price into your history, is exactly the case nobody has counted.
3. Compare the JSON-LD price with the rendered price
Real if the structured data offer price and the price on the page agree. Suspect if they disagree, because you have probably captured a minimum offer across sellers, a stale feed value, or a "from" price for the cheapest variant.
Trust the rendered page. It is the number the shopper is asked to pay, which makes the structured data the side more likely to be wrong.
4. Check the currency and the resolved URL
Real if the currency, the country and the child SKU are the ones you have been tracking. Artifact if the currency symbol changed, if you were redirected to another locale, or if the final URL carries a variant parameter that differs from the tracked one.
A 30% drop and a currency swap look alike. So does a price per unit replacing a price per pack.
5. Check the siblings
Real if other variants of the product, or other listings from the same seller, moved together. Artifact, or at most one clearing size or color, if a single variant collapsed while its siblings held.
Identical percentage moves across a whole range deserve their own suspicion, since that signature belongs to a repricing rule or a feed error rather than a human strategy. This is why repricing guardrails matter for reading competitors as much as for running your own engine.
6. Take it one step toward checkout
Real if the price holds once the item is in a cart. Never an offer if it reverts at the cart, or if the discount appears only at checkout.
This also catches the expensive inverse: a rock-bottom price on an out-of-stock listing gives you no competition to answer, and the stock signal underneath it has a calculable value of its own.
7. Read the shape of the history
Real if the price stepped to a plausible new point and stayed there across two or more consecutive checks. Artifact if one observation spiked and the next returned to the old price.
Persistence is the cheapest confirmation available, and it costs only time. On an hourly or two-hourly checking floor, two confirmed observations arrive well inside the window in which you could implement a response anyway.
8. Only then, test it against plausible landed cost
If the price survived the first seven checks, ask whether anyone could sell at it. Below plausible landed cost plus fees, treat it as a pricing error however well it verified. Do not match it, and expect it to be corrected or the orders canceled.
Five Patterns You Will See Again
| Pattern | Most likely cause | Correct response |
|---|---|---|
| One variant at 90% off, siblings unchanged | Tracked URL resolved to a different child SKU, or a genuinely clearing size or color | Confirm the SKU, then ignore. Never reprice a range off one variant |
| Whole catalog moves by an identical percentage | Repricing rule, feed error, or a currency or unit mix-up | Do not match. Wait one cycle. Identical deltas come from a machine, not a strategy |
| Price drops to an exact round number such as 1.00, 0.01 or 9.99 | Default or fallback value written when the real price failed to parse | Treat as an artifact until a second independent fetch confirms it |
| Drop visible only from one geography or IP class | Geo or currency page, or datacenter versus residential divergence | Re-check from the market you actually sell in. Otherwise there is no signal |
| Drop appears only in your tool, price unchanged in a browser | Strike-through or "was" price captured as the current price | Fix the extractor. Do not act |
The last two rows rest on practitioner observation rather than measurement, and that is worth saying plainly instead of attaching a percentage to them. Audit reference price capture before any promotional period, because a monitor that reads a "was" price as the current price will misread the whole deal-eligibility lookback window.
How Long a Wrong Price Stays Up
Correction speed has not been studied either, but two documented incidents bracket the range.
On Friday 12 December 2014 a third-party repricing tool drove Amazon UK marketplace listings to GBP 0.01 for roughly an hour. Amazon canceled the vast majority of the resulting orders, though items already dispatched by FBA could not be recalled (Washington Post, December 2014). Item and seller counts varied between outlets and were never company-confirmed.
The opposite case ran for ten days unnoticed. In April 2011 two sellers listed the same used biology book and each priced off the other: one at 0.9983 times its rival, the rival at 1.270589 times the first. The listing peaked at USD 23,698,655.93 on 18 April 2011 and collapsed to USD 106.23 the next day (Michael Eisen, April 2011).
The arithmetic explains why nobody caught it. One full cycle multiplies the higher listing by 1.270589 x 0.9983, which is 1.26843, a rise of about 26.8% per day. Ten cycles compound that to 10.78 times. When the loop was spotted the lower listing sat near USD 1.73 million, putting its mirror at 1.270589 times that, about USD 2.198 million. Multiply by 10.78 and you land on USD 23.7 million, matching the recorded peak ten days later. A 27% daily rise is unremarkable in any single day's log. Ten in a row is a 10.8x move.
Your response window therefore depends on who else is watching. On a high-traffic marketplace listing, assume an error is fixed within minutes to hours, so speed buys exposure rather than advantage. On a long-tail listing nobody monitors, a wrong price can sit for days, which makes persistence weaker evidence there.
Do They Have to Honor a Mispriced Offer
One legal question comes up every time, and it is jurisdiction-specific enough that this is orientation rather than advice.
Whether a mispriced offer binds the seller depends on when the contract formed. In common-law jurisdictions a displayed price is an invitation to treat: your order is the offer, and the contract exists only once the retailer accepts, so an error caught before acceptance can be refused, with a refund owed if money was already taken (Citizens Advice). Where a contract can instead form at the point of order, an obvious error, meaning a discount so extreme no buyer could reasonably believe it, still tends to let the seller cancel the order, and the seller is expected to correct the page promptly (ECC Netherlands, March 2018).
Questions Sellers Ask About Price Anomalies
Does a retailer have to honor a mispriced item they sold me?
It depends on the jurisdiction and on whether a contract was formed. In common-law jurisdictions a displayed price is an invitation to treat, so the seller can refuse the order before acceptance and must refund money taken. Where a contract can form at the point of order, an obvious error still tends to let the seller cancel. Take legal advice before relying on either commercially.
Why does my monitoring tool see a different price than my browser?
Usually one of three things. Bot mitigation served your collector a degraded page with no real product data, your collector landed on a different country or currency version of the listing, or the price is personalized and your browser sits in another segment. Compare the structured data price with the rendered price, then re-fetch from the market you actually sell in.
Is a price that only some shoppers see a glitch or a real price?
Often a real price, just not one offered to you. Hannak and colleagues found evidence of steering or price discrimination on 9 of the 16 sites they measured. Personalized pricing means a number your monitor sees and your browser does not can be entirely genuine. It is a competitive signal only if the segment seeing it overlaps with your buyers.
What alert threshold stops false positives without missing real cuts?
No number is derivable, because nobody has published a base rate for how often extreme moves are errors. Set the threshold at the size of move that would actually change your price, then require confirmation across two consecutive checks before anything reaches a human. Confirmation removes far more noise than a higher threshold does.
Where a Tracker Does Part of the Work
Several of these checks are things a monitoring setup should be doing for you. Respot tracks price and stock per variation rather than at parent-product level, which makes check 5 readable without opening five tabs: one variant collapsing while its siblings hold shows up in the data instead of being averaged away.
Recorded history is what makes check 7 work. The free plan keeps 7 days of price history across 5 trackers, enough to see whether a drop persisted. Paid plans extend that to 30, 90 or unlimited days across 100, 400 or 2,000 trackers, and longer history turns "this looks odd" into "this seller does this every quarter".
The Decision Worth Making Before the Next Alert
The next time a competitor's price falls off a cliff, ask whether anything actually happened before asking how fast you can match it.
Write the eight checks down where your team can reach them, and agree that no price move gets a response until it has survived the re-fetch, the sibling check and one repeat observation. That removes most of the false alarms from your week at the cost of a short delay on the genuine moves, which were never going to be won or lost in the first ten minutes.
The damage from an anomaly comes from acting on it, rarely from missing it. Start tracking your competitors with a procedure for deciding what to believe.
13 min read
Catch competitor price cuts before they cost you sales
Respot monitors any competitor product page and emails you within minutes of a price or stock change. Free plan, no credit card.
Catch competitor price cutsRelated articles
What to Do With Your Pricing Data After Black Friday
The post-Black-Friday analysis only works if you recorded the right things beforehand. What to capture before peak week, and what to run in December.
Peak Season Monitoring FAQ
Eight questions sellers ask in late October, answered: peak-week check cadence, repricer risk, stockout value, the deal lookback and what to record now.
Brand vs Retailer: What Each Should Actually Monitor
Brands monitor their own products on other people's storefronts. Retailers monitor other people's products. Different objects, thresholds and legal risk.