Feature · published on September 9, 2026

The price of closed sales

Most price estimators read live listings. We read the sales that went through. It is the only difference that matters, and it changes everything else — including our ability to say “I don’t know”.

The three rules of the corpus

• Eighteen months. Past that, a sale no longer describes the market: trends move, new models land, values shift.
• One sale per seller. An account shifting a hundred identical pieces should not count a hundred times; it counts once, like everyone else.
• Two distinct sellers, minimum — except on your own items. Sales from other accounts only count once two different sellers have made one. Yours always come back to you: your history belongs to you, and you should not have to wait for a stranger to sell the same thing before it pays off.

The second rule is worth explaining, because it runs against instinct. A seller moving high volume on one model has a pricing strategy — usually aggressive, sometimes the opposite. Letting them count in proportion to their volume would measure their commercial policy instead of the market. One sale per seller gives a smaller corpus and a far more representative one.

The third rule is the one that costs us something. It stops us answering on rare pieces, which are exactly the ones where an answer would be worth the most commercially. We keep it because a figure built on a single sale is not an estimate, it is an anecdote — and an anecdote presented as an estimate makes people buy badly.

Why asking prices tell you nothing

Listings still visible are, by construction, the ones that did not sell: the ones that find a buyer disappear from the results. An estimator that feeds on them is not measuring the market, it is measuring what sellers hope for — and passing it on to you, high, every time.

This is selection bias, and it is a nasty one because it is invisible: the worse a category sells, the more listings stay visible, and the more optimistic a listing-based estimate becomes. The error peaks exactly where it costs the most.

So why read unsold items at all?

Because an asking price proves one thing and one thing only: nobody paid it. That is weak information, but it is not nothing — it is a ceiling. When closed sales are missing, an item that sat unsold at €60 at least tells us €60 is too much, which beats silence.

So we use them, under three strict conditions that follow directly from the bias above.

• They are labelled as such. The model reads “listed at €60, unsold to date”, never “sold for €60”. The instruction attached to those lines tells it to treat them as a probable ceiling, not a price reached.
• They weigh less. At equal similarity and freshness, a closed sale always comes ahead of an unsold item. The ranking is never ambiguous.
• They are never counted as sales. When we write “based on 4 sales”, those are four closed sales. An unsold item never adds one to that counter, and it never enters the time-to-sell figure either — an item that never sold has no time to sell.

We have not changed our minds about asking prices: they still do not tell you what an item is worth. They tell you what it is not worth, and that is the only reason they get in.

What you get back

• A range, never a single number. A fair price is an interval.
• A confidence level, which tells you how much data it rests on.
• A “see comparables” link, which opens the real sales so you can check for yourself.

The third point is there for a reason: we are not asking you to take our word for it. An estimate you cannot cross-check is an opinion, and you have no reason to trust ours over anyone else’s.

One useful detail about the range: it aims at a realistic selling price within one to two weeks, not the theoretical maximum. Those are two different numbers, and the second almost never happens — or only after months of waiting, which is a legitimate choice but one worth making knowingly. The width of the range is itself information: narrow means a readable market; wide means the comparables disagree, usually because condition or the exact model moves the price more than usual.

Where the corpus comes from

From users’ closed sales, anonymised. Yours feed other people’s estimates, as theirs feed yours. No identity, no email address and no photos are ever sold on — the detail is in the privacy policy, which is in French.

It is a commons, with the usual property of commons: it is worth what its contributors put into it. A well-represented category gives tight, confident estimates; a thin one gives low confidence, and says so.

That is also the honest limit of the whole thing, and we would rather write it here than let you find out: a corpus gets built. On categories that are still thin, the tool will be cautious, and that caution is the intended behaviour — not a fault.

One consequence that matters if you sell outside France: your own sales are always returned to you, without waiting for anyone else. Import your selling history and the estimates fit your market from the first day, even where our corpus is thin. That is not a workaround, it is the same rule as everywhere else — your history belongs to you.

What the estimate does not do

• It does not guarantee a sale. A realistic price is still just a price: you still need a decent listing and readable photos.
• It does not know your local market. On a face-to-face sale, demand where you live can differ from the average.
• It does not replace your judgement on condition. That is the single biggest factor, and you are the one holding the item.
• It does not predict trends. It describes what has sold, not what will.

None of those four is a bug, and none of them is going to be fixed. They mark the edge of what a corpus of past sales can tell you. Knowing where that edge runs is what lets you use the number well — and know when to overrule it.

Free account — 10 estimates a month

Read next: Listing titles and descriptions · Pricing