2026-10-16

Why AI Appraisals Disagree With Your Asking Price

Four of our own names appraised twice and priced a third way: Chauffeurs.si at $199 against a $1,650 mid, GetLogo.si at $1,795 against $7,320, and the thirteen .com names where the median ask was $375 and the median appraisal $1,999. What a comparable set, an extension and a commission ladder each do to the number, and the five steps we use to reconcile them.

Two automated appraisals of four of our own names ran before breakfast this morning, and neither of them agreed with the price on the landing page. Chauffeurs.si is listed at $199 and appraised at a mid of $1,650. PetSitting.si is listed at $1,495 and appraised at $3,580. Excavators.si runs the other way: $1,995 on the page, $1,660 from the model. And GetLogo.si, priced at $1,795, came back at $7,320.

Four names, eight numbers, and not one pair of them matching. That is not a broken appraisal. It is what an appraisal is: a model's opinion about a class of name, built from receipts it has seen. An asking price is a seller's opinion about a buyer who has not arrived yet. They answer different questions, so they produce different numbers — and the useful work is not picking the winner, it is knowing which one is which.

An appraisal prices the class, an ask prices the buyer

Run an automated appraisal and you get one of two shapes back. The first is a single point: a marketplace dashboard prints "Estimated value" next to the buy-now field and the seller copies it, which is how a lot of names end up priced at a number nobody ever argued for. The second is a range with its workings attached — a low, a mid, a high, a confidence label, the number of comparables behind it and every modifier the model applied.

We used the second shape for these four names, through the public .si appraisal endpoint on worthly.si, and the outputs are worth reading line by line because they show exactly where the disagreement is born:

• Chauffeurs.si — listed at $199 · model range $900–$3,450 · mid $1,650 · 6 comparables, median $3,400 · +25% for being registered in .com, .io, .co, .net and .org
• PetSitting.si — listed at $1,495 · range $1,970–$7,520 · mid $3,580 · 6 comparables, median $10,000 · +28% key extensions, then a 0.92 haircut for the name already being taken
• Excavators.si — listed at $1,995 · range $910–$3,480 · mid $1,660 · 6 comparables, median $3,400 · +28% key extensions
• GetLogo.si — listed at $1,795 · range $4,030–$15,380 · mid $7,320 · 6 comparables, median $20,000 · +6% (only .net and .org are taken)

Two of those asks sit well under the model's mid, one sits 20% above it, and one sits far below it. Look at the fourth line again: our ask is $1,795 and the comparable set behind the appraisal has a median of $20,000. Nothing about GetLogo.si changed between the two numbers. What changed is who is speaking.

The comparable set does most of the talking

The method note on Chauffeurs.si's appraisal reads: "Adjusted 35% towards 6 tracked .si comparables (median $3,400)." That one sentence tells you how the number was made — a structural estimate for a ten-letter dictionary word, then a 35% pull toward six real sales. When a class is this thin, six comparables are not a market. They are six events, and one of them can move the mid by hundreds of dollars.

That is the first structural reason an appraisal and an ask disagree, and it is measure-free: the model prices a class, and a class can be three deals wide. Your buyer is pricing one name, one use case, one afternoon of decision-making. On a thin comparable set the two will never meet, no matter how carefully either side is calculated.

Receipts, not claims

The index those comparables come from is public and it draws one line that matters more than any modifier: the sales index excludes asking prices. As of 16 October 2026 it carries 291 tracked .si sales, a median of $2,600, an average of $4,403 and a high of $50,000 — and every entry is either a venue-confirmed deal or a reported one, never a listing price.

Read the median against the average and you have the second reason the numbers diverge. The average is 1.7 times the median, which means a small tail of large deals pulls the mean upward while the typical sale stays low. An appraisal engine anchored near the mean will print your mid-market name high; one anchored near the median will print it low. An asking price has no anchor at all — it has a seller, a renewal date and a bill.

For scale on how thin the whole namespace is: another public .si tracker we read counts 108 reported retail sales in 2026, against 29 in all fourteen earlier years combined, at a 2026 median of $5,000. A country-code market that trades roughly a hundred names a year at retail cannot support a point estimate pretending to be precise. It can support a range, a confidence label and an honest comparable count — which is precisely what the range-shaped appraisals hand back.

Why the two numbers can never be equal

Four reasons, in the order they usually bite:

1. Thin comparables. Six comps for a ten-letter word means the print is an argument, not a measurement. Ask what the comparable count is before you argue with the number.
2. The extension is a separate market. Excavators is registered in .com, .ai, .io, .co, .net and .org, and that breadth is worth +28% to the .si model. The same word in a weaker extension would have appraised lower with the same length, same dictionary status and same demand. An appraiser prices the name it is given; a buyer prices the name they can actually build on.
3. Class versus use. Models are good at categories — dictionary word, compound, invented brandable, two letters, a number. Buyers pay for a sentence: transfers in Ljubljana, logos for a design tool, excavators for a contractor. GetLogo.si is a seven-letter compound with a median comparable of $20,000 behind it. Our $1,795 ask was set for a quick retail flip, not for the use case the model priced.
4. The ask carries costs the appraisal never sees. Our .com names sit inside a retail programme where the commission ladder runs 30% up to $4,998, then 25%, 20% and 15% at the higher bands, with a $100 minimum retail transaction. An ask has to clear that ladder and still leave a floor worth holding for. The model is pricing the name. The ask is pricing the name minus the marketplace, minus escrow, minus five years of renewals.

Thirteen .com names, one direction

The clearest evidence we have is our own repricing of thirteen .com names in a retail programme this month. The old asks had a median of $375. The marketplace's own appraisals had a median of $1,999. The gap between the typical ask and the typical print was more than five times.

Twelve of the thirteen asks sat below the appraisal; exactly one sat above it. Solarell.com was listed at $195 against a $2,399 appraisal. LemonWhite.com was at $100 against $2,799. AlwaysTherapy.com was at $375 against $2,099. At the other end, ParalegalLab.com was asking $2,195 against a $1,499 appraisal — 46% over the marketplace's own opinion, which is the one price a buyer can check from the landing page itself.

The repricing that followed moved those names to between 85% and 90% of the printed appraisal — close enough to read as fairly priced, low enough to read as a bargain. One name went down. That is the whole exercise: the appraisal is a ceiling to argue under, not a target to hit.

The reconciliation we run now

Five steps, in this order, every time a new number lands:

1. Get both shapes. A comparable-based range and a point estimate. One number is not an opinion you can examine.
2. Check the comparable count and the confidence label first. Six comps and "medium" confidence means the mid is the centre of an argument, not a price.
3. Strip the seller's costs before comparing. Commission band, minimums, escrow floor, renewals. A $1,650 mid that nets $1,155 after a 30% band is a different decision than a $1,650 ask.
4. Anchor the floor to receipts. What a comparable actually sold for is the only number in the chain that someone paid.
5. Set the ask inside the band and slightly under it — then leave it alone, and keep one price per name. A second cheap price for the same name on a wholesale thread teaches every buyer your floor.

When the appraisal is simply the wrong number for your name

Three cases where we ignore the print, and say so out loud:

• There is no comparable class. For invented brandables the comp set is thin or borrowed, and a low-confidence print on a name nobody has ever traded is a guess with a currency symbol in front of it.
• A buyer has already spoken. An offer in hand is a price. It beats a model on every axis except convenience.
• The class is re-rating faster than the receipts. The .si namespace went from a reported sale or two a year to more than a hundred in 2026, and appraisals built on earlier receipts lag that shift by definition. We price .si as an emerging extension a tier below .ai, and we say so in the listing rather than hiding behind a number.

An appraisal that disagrees with your asking price is not a verdict. It is a second opinion from someone who has read every receipt in the class and has never met your buyer. Price the receipts as your floor, the band as your ceiling, and your buyer as the only number that settles anything.

Sources and method

• Worthly.si appraisal endpoint, called 16 October 2026 for Chauffeurs.si, PetSitting.si, Excavators.si and GetLogo.si (method label: appraisal v2), with each response's comparable count, comparable median, confidence label and modifiers quoted above.
• Worthly.si sales index, 16 October 2026: 291 tracked .si sales, median $2,600, average $4,403, highest $50,000. The index's own method note states that asking prices are excluded and that entries are labelled verified or reported.
• Appraise.net .si market page, 13 October 2026 snapshot: 108 reported retail sales in 2026, 29 in all fourteen earlier years combined, 2026 median $5,000.
• The thirteen-name repricing figures come from our own Atom Premium sheet, 13 October 2026, and the retail commission ladder quoted is the published Premium band table for a standard listing.
• The four asking prices quoted are the live retail prices on our own portfolio page at dnwai.com/d on the same date.