How we price, and how we know

1. 1

   Read the photo

   Brand, model, maker's marks, edition and condition, from the picture. When a valuable item names its model, a second pass re-reads the legible text and every disagreement is logged, so the misses teach the next version.

2. 2

   Find real sales

   Sold-price sources, asked in order until one answers: specialist eBay sold-price API; an independent backup sold-comp service; archive of recorded sales.

3. 3

   Keep the sales that match

   Lots, sets, bundles and the wrong variant are filtered out. A comp sold for parts counts for less. Best Offer sales are counted below their shown price. When the matches spread too wide, the confidence drops instead of the spread hiding inside a median.

4. 4

   Price it

   A weighted median of the matched sales, corrected by a calibration learned from items that actually sold, per category and per price band. At $50 and up an AI read of those same sales against your item's condition sets the point, kept inside our range. Then three list prices, each with its two-week sell chance once enough listings like it exist.

5. 5

   Say how sure

   A confidence grade, a range built to hold most sales, and the sold listings themselves one tap away, so you can judge the number before you trust it.

• A sale is capped by the ask

  A listing cannot sell above its asking price. When a seller lists below our number, the sale lands below it too and looks like a miss.

  What we do. We grade ourselves only on sales where the seller chose their own price, and read high or low only from sales no ask could cap.

• The photo cannot show everything

  Whether it works, what is missing from the box, a chip on the underside. Two identical photos can be two different items.

  What we do. The scan grades condition from the photo. Tap a different condition and the three prices move, and a detail you add re-runs the sold comps at no charge.

• Some things sell for many prices

  A named maker, a rare colorway, a first edition. Vintage and art run widest, and a median of everything is wrong for the piece in your hand.

  What we do. For pottery, where sold comps made the price worse than the model's own read, the model's number is kept and the comps become a sanity band. The weekly grade says which categories should follow.

• Comps lie a little

  A lot of six listed as one, a part sold as the whole, a Best Offer shown at its asking price.

  What we do. Matched-sale filters, a discount on Best Offer sales, and a ceiling that stops one outlier lifting the number.

• Markets move

  Ninety days is a long time for a phone and a short time for a chair.

  What we do. Recent sales weigh more, and the weekly grade tells us when a category drifts.

1. September 25, 2026

   The same photos put to other AI every week and published beside our number.

2. September 18, 2026

   This page: the weekly grade published, wins and misses.

3. September 14, 2026

   The AI price refiner promoted for items at $50 and up after a shadow test.

4. September 11, 2026

   Comp pollution filters after a drone camera counted as a comparable for a bike trainer; a price now waits for the comps instead of showing a guess first.

5. September 10, 2026

   Realisation factor: the estimate corrected by how much of our number items in that price band actually fetched.

6. September 9, 2026

   Best Offer sales counted below their shown price, since the shown price is the ask.

7. September 8, 2026

   Model anchor for pottery, where sold comps were making the price worse than the model's own read.

8. September 3, 2026

   Per-category calibration learned from items that sold: the price level and the range width now come from measured sales, not a fixed margin.

Sales the seller reported in the last 90 days, with the photo they scanned. A model that gave no number on a sale is left out of its own count, so its row also shows our gap on exactly the sales it answered. Linked means the answer carried an eBay listing link; for us it means the scan showed sold listings. Our own model cannot link, so that cell says n/a. Junk sale prices are dropped by the same rule for every row. Updated October 5, 2026.

• 4 in 10

  sales landed inside the range we showed. 6 in 10 across every sale.

• $8

  on a $100 item is how far sellers who listed at our number ended up from it. They were asking our number, so this proves less than it looks.

• High

  We run 2% high on sales no asking price could cap.

How we measure it

• A sale counts when you mark an item sold and enter the price, or when a listing published through us sells on eBay. We use the price the item actually fetched.
• The gap is the distance between our estimate and the sale price, as a share of the estimate. The typical gap is the median, so half of the sales were closer than the figure shown.
• Sellers who list at our exact number and sell at it are, in part, proving our number by using it. So the headline figure counts only sales where the seller priced on their own, and the anchored figure is shown separately.
• Lean is read only from sales no asking price could cap, such as auctions, best offers and sales you report yourself. A listing that sells at the seller's ask cannot sell above it, so it says nothing about whether we ran high.
• A category is named once it has 8 or more sales in the window. Smaller ones are folded into the totals.
• The window is the last 90 days and the report runs every Tuesday.

What we do with it

When a category runs weak, the result page says so in plain words next to the price, and the comps view shows the sold listings the number came from, so you can judge it for yourself. The weekly report also feeds the calibration that sets our ranges, category by category.