How a number is made

Written so you can check us. Every figure on this page comes from the last scored day (—, —) and is fetched from the same file the track record renders.

1. What we are actually looking for

The word for it is under-attended. Every auction looks cheap until it closes. An auction only ends below market when price discovery fails: nobody qualified saw it, the title was bad, it closed at 4am, the category is thin. So the thing we rank on is low attention on an item we can independently value. A lot with forty bids is correctly priced by definition, however cheap it looks with an hour to go.

That makes identity, not price, the binding constraint. If we cannot work out what an item is from its title, no amount of quiet tells us anything.

2. The matching ladder

Every listing is matched to a cluster of comparable items. The rung it lands on determines what we are allowed to say about it.

Rung 5 is the important one. A lot we cannot identify is not a cheap lot, it is an unknown lot, and the honest output is silence. The gate is mechanical — rungs 4 and 5 cannot reach an alert regardless of how large the apparent discount is.

3. Why we decline most of them

On the last scored day we declined — of — closes — — of the day — and scored the remaining —.

Most auction titles do not contain enough to identify the item: no model number, no brand, no attributes that pin it to anything priceable. A tool that produced an estimate for all of them would be producing noise for the — above, and false positives here cost you money rather than attention. We would rather miss half the good lots than surface junk.

4. The three prices, kept apart

The two get confused easily, so they carry separate labels everywhere on the site and never substitute for each other.

A Buy It Now price is an ask, not a transaction. Active BIN pools are survivorship-skewed: correctly priced items sell and leave, overpriced ones sit forever, so the median ask drifts above real market value. We therefore use the 10th percentile of the credible BIN pool, not the median. For a buyer that is conservative in the right direction — your real alternative genuinely is somebody’s asking price. For a reseller it is not good enough on its own, so the reseller figure takes the same p10 ask, subtracts eBay’s final-value fee at its published rate, and subtracts what the lot cost to land. That is a proxy and is labelled one everywhere it appears: an active ask is not a sale, and a fee rate is not an invoice. A margin figure built on real sold comps is a different and better number, and it is not the one on these pages.

Where a market has no active BIN pool at all, the reference is instead a leave-one-out comp: the 10th–90th percentile of what the other items in the same cluster closed at. The lot’s own price is never in the pool that priced it. Being under that is a weaker claim than beating a live asking price, so it is labelled under typical close rather than buyer savings.

5. What the interval means

Predictions are published as ranges with a sample size, never as a single number. Two reasons, and neither is stylistic:

• The winner’s curse. You win the auctions where your estimate was highest. Bid guidance taken off a mean is guidance to overpay systematically, so every number we suggest acting on comes off a conservative percentile.
• Sample size changes the action. n=47 and n=3 justify different bets, so n appears beside every aggregate on this site. No n, no number.

On the last scored day, — lots closed under the cheapest credible ask for the same item, a total of — with a median of —. Every one of those figures is measured against an asking price, so read it as what a buyer avoided paying rather than as a profit.

A predicted closing range is published only where the cluster is deep enough to build one, which is a minority of lots. Predicting the hammer needs a pool of comparable closes to draw a range from, and most eBay clusters do not have one — so most rows carry no predicted close at all and are scored against an asking price instead. Where a range does exist it is published with its outcome, hit or miss, and its calibration is published with it: across — scored closes the real price landed inside the range — of the time, against an 80% target. That is a miss rate near two in five, the ranges are narrower than the evidence supports, and it is the reason the predicted close is not what this site leads with.

6. A published score is frozen

A track record that can be rewritten later is not a record. Our comparison pools are built from asking prices that are currently public, so they thin out on their own: an ask nobody sees again for a week is dropped, and once a listing sells it stops being a comparison at all. That means the pool available today is not the pool that existed when a lot closed three weeks ago.

Scoring a past close against today’s pool would therefore let history drift — and in practice it did: re-scoring one day a day later produced noticeably fewer identifications than it had the first time, purely because pools had thinned underneath it. So once a close has settled, its score is written down and never revised. The number you read for a lot from last month is the number we published for it at the time, not a reconstruction, and it cannot improve after the fact either — a lot we declined to price stays declined even if we could price it now.

Days from before we started doing this are reconstructions. We fixed them in place too, so they stop drifting — but they were fixed from comparison prices that had already thinned, so they understate what we published at the time. They are labelled separately on the track record for exactly that reason: a number that has stopped moving is not the same thing as a number we stand behind, and using the freeze to quietly promote one into the other would be the dishonest version of this whole exercise.

7. Shipping is quoted to one reference location

Many sellers price shipping by weight and destination, so there is no single correct figure: eBay cannot quote one without knowing where the buyer is. Rather than leave those blank, we ask for the cost to one fixed reference location in the central United States and publish that. It is the same location for every lot on this site.

Your shipping will differ. If you are far from the middle of the country, a weight-priced item costs you more than the figure shown, and the saving shrinks accordingly. Shipping is quoted at collection time, not for you personally — pricing it per visitor would mean calling eBay on every page view, which the API budget does not allow. Check the real cost on the listing before bidding.

8. The one projected figure, and how it is built

The landing page leads on a daily figure that we have not measured, and this is where it comes from. It is not a claim about eBay’s size — we do not scan eBay, and we have no estimate of its total auction volume. It is our own measured rate, multiplied by the coverage we do not currently buy.

The measured half is on the track record, and it is the narrower of the two totals there: the savings on lots that cleared the alert bar over the frozen window, divided by the days in it — not the larger total across every lot that closed under the cheapest ask. Alerts are what a seat buys, so alerts are what gets scaled. That is a real number about real closes, and it is the number the projection multiplies.

The multiple is 16× — two factors, 5.4 × 3 = 16.2, rounded down:

• 5.4× — auctions we surface but cannot afford to follow. The queries already running turn up roughly 7,600 live auctions a day, and the daily API budget lets us follow about 1,400 of them through to a close. The rest close unwatched. That gap is a spending limit, not a statement about how many under-priced lots exist.
• ~3× — a wider watchlist. Two dozen queries today; the high tens is an ordinary target. The 7,600 above is what those two dozen queries surface in a day; by contrast a single one of them, taken alone, sits on a standing pool of 77,742 live auctions. Those two figures measure different things — a daily flow against a standing stock — and the gap between them is why widening the list is the softer of the two factors.

What would make it wrong. The projection assumes a wider watchlist holds the same gap per lot, and there is a real reason it might not: queries are chosen partly because their items can be priced, and only about 12% of the product clusters we know about currently carry a priced comp pool. A broader list would likely dilute the rate rather than hold it. We publish the figure because the direction is not in doubt and the arithmetic is checkable — not because the number is settled.

9. What this is not

• Bidding. We never do it. The service is read-only and a human places every bid. eBay’s February 2026 user agreement bans automated ordering flows and we are not going anywhere near that line.
• Profit. Neither figure here is any. The buyer’s gap is against an asking price or a cluster of closes, before any fee. The reseller’s figure nets off eBay’s fee rate, which is arithmetic on an estimate and not a sale you have made. Neither is profit.
• History. We keep none of it. Individual listings are not retained past their public-availability window; what persists is aggregates — cluster, median, sample size, timestamp.
• Volume. Still unproven. The rate at which lots clear the full bar is still being measured. See the board, which is empty on most days by design.