About NameFew

NameFew pits a robot against people on one question: what is this domain worth?

You type a .com. The robot answers in a second with a price range and the real sales it leaned on. Then people (appraisers, we call them) rank that name against two names that really sold. Their combined answer shows up beside the robot's. Later we check both against what the name actually sells for. That last step is the point.

Why this exists

I built NameFew in 2020 as a crowdsourced appraisal experiment. It never recorded a single vote. I rebuilt it in October 2026 with a sharper question: the machines are mediocre at this, and nobody had measured how mediocre, so can people do better? We don't know yet. The site exists to find out.

How the robot works

The robot is a statistical model trained on 19,329 reported .com sales (2025-12-31 to 2026-10-21). It knows the name's length, whether it's made of real words, and a small language model's reading of it (what kind of name it is, its category, brandability, commercial appeal, memorability, trademark risk). That reading never sees a price. The range it shows is wide on purpose: calibrated so that about 80% of comparable sold names fall inside it.

An AI opinion sits next to it, from a language model that sees the name and the real comparable sales.

How accurate is the robot?

Modest, and we'd rather say so.

• The statistical model. We trained on January to June 2026 sales and tested on 6,443 unseen sales from July to October. Rank correlation (0 is a coin flip, 1 is perfect) was about 0.31, up from 0.26 before adding the language-model reading. It flagged the top 7% of sales (those at $8,000 or more) with an AUC of 0.78, where 0.5 is chance. About 80% of those unseen sales fell inside the stated ranges. All-digit and very short names get wider ranges, because some of them sell for ten times the usual price.
• The AI opinion. We tested it on 300 unseen .com sales from July 1 to October 21, 2026. Its ranking was no better than the simple model's (rank correlation about 0.21, with a margin of about 0.11 either way). Its ranges were honest (the real price fell inside 85% of the time) but wide, about five times from low to high. And it can't see the rare high-value sale: names that sold for $24,000 to $40,000 were guessed at $1,350 to $4,000.
• Full write-up, with the test sets: Are AI domain appraisals accurate?

How the humans work

When you judge, you see three names. Two of them really sold, for prices we know. The third is waiting for appraisal. You tap them from most valuable to least. We grade you on the two that sold, and where you put the third helps work out its value. A name needs several appraisers before we show a human number, and then we show a range, not a verdict.

Not every vote counts the same. We weight each appraiser by how well they rank names with known prices, because a vote from someone who gets those right says more than a vote from someone clicking at random. A new appraiser's votes count for little until they've shown some skill. And a group of accounts that arrive together and vote alike (the same network, the same device) counts as closer to one voice, so a brigade can't push a name's value around. If we think an account is a duplicate or is judging a name it has an interest in, we discount or remove it. The rules are in the terms.

We haven't yet measured whether the humans beat the robot. In a simulation with only five appraisers per name, the human middle figure barely beat the robot. With ten it was clearly better, and a name now needs seven. When there's real data we'll publish it, whichever way it goes.

Tools

• Appraise: one name, with comparable sales, the robot's range, an AI opinion and the humans' range once there is one.
• Bulk appraisal: up to 25 names in a sortable table you can copy as CSV, and a button to request human appraisals for them.
• Judge: be an appraiser.
• Game and the daily round: rank real names, see real prices, find out whether you beat the robot.
• For sale and Sell a name: free listings with no commission. You prove you own a name (a DNS record, or a file or tag on the name's own site), and every price sits beside the robot appraisal and the human appraisal. Buyers write to you in the site (offers and messages). NameFew isn't a broker and holds no money.
• Guides: what names sell for, with our own numbers.

What it can't do

• Tell you whether a name will sell. Most domains never sell.
• Value names below roughly $1,000. The data only contains sales above that.
• Know a buyer's reasons, keyword demand, or brand conflicts.
• Handle non-.com names yet.

Where the data comes from

Reported domain sales from multiple sources. We show a handful of comparable sales per appraisal and publish summaries (medians, counts, spreads), never the full list.