2026-10-17

Root Words and Listing Scores: How Marketplace AI Ranks Your Names

Five of our own listings read field by field: ShopUna.com indexes a root word called "Shopu", ImageFixerAI.com indexes Image and Fixer but drops the AI, and Chauffeurs.si indexes one French word. What 164 live names say about root-word overlap, why a listing score is a placement readout rather than a valuation, and the five checks we run before touching a price.

A marketplace knows your domain by about eight fields, and one of them does more work than the price. It is the root-word block: the short list of words printed under Domain Characteristics on a listing page, each one a link into that marketplace's own search. We read that block on our own listings this week, and on one of them the marketplace had decided our name's only root was a word that does not exist.

ShopUna.com is listed at $1,295. Its root words, as Atom prints them today, are: Shopu. Not Shop, not Una — Shopu, a single invented token linking to a query page for a string no buyer has ever typed. The description directly above that block, written by the marketplace's own AI, reads: "a captivating blend of 'shop' and 'una', evoking images of a unique, one-of-a-kind shopping experience." The prose understands the name. The index does not.

That gap is the whole subject of this piece. A listing score is not a verdict on your name handed down from above. It is a readout of how well the marketplace's model can place your name inside its own inventory — and the raw material for that placement is the roots it extracts from your string, plus whatever fields you fill in around them. Understand the extraction and you understand most of what the score is telling you.

Root words are search keys, not descriptions

On Atom, every root word in that block is a hyperlink to the path /premium-domains-for-sale/q/. Open one and you land in the marketplace's result set for that query. That is the whole mechanism in one line: the root word is the query, the query is a shelf, and your name either sits on that shelf or it does not.

So the field is not cosmetic, and four of our own listings show how differently it can resolve:

• LemonWhite.com ($2,495) indexes Lemon and White — two clean dictionary roots, two live shelves.
• ImageFixerAI.com ($2,800) indexes Image and Fixer, and drops AI entirely. The three letters that carry the entire category of the name are not a root word as far as the index is concerned.
• Chauffeurs.si ($199) indexes Chauffeurs — one root, in French, on an extension where the marketplace's own inventory is thin.
• ShopUna.com ($1,295) indexes Shopu — no shelf at all.

Read those four lines together and the pattern is not about quality. Lemonwhite is a two-word brandable with no category at all; it gets two real roots. ImageFixerAI is a name whose entire meaning is a category; it loses the category token. The model is splitting the string it was handed, not reading your positioning. When the split produces a token that is not a word, that is what gets printed, and that is the key your listing is filed under.

The block is derived, not typed

Notice that nothing in our descriptions moved these words. ShopUna's own copy says "shop"; the root says "Shopu". ImageFixerAI's copy is entirely about fixing images with AI; the root list stops at Fixer. The extraction follows spelling and syllable shape, not the sentences a seller writes around the name.

This is the part sellers get wrong most often. We spend an afternoon on price, a paragraph on a description, and nothing on the question of what string the marketplace's parser will actually produce — even though that string is what decides which searches return the name. The description is read by humans who already found you. The root words decide whether they find you.

There is a second-order effect that matters more for a portfolio than for a single name: the shelf is shared. If your root is a live query, you are not only competing with other sellers, you are competing with your own inventory. That is measurable in our book.

Our 164 names, counted by root

Our live portfolio today holds 164 names: 147 .com, 9 .si and 8 .net, with a median buy-now of $488, 35 names at $200 or under and 26 above $1,000. Average label length is 11.1 characters; 53 labels are eight characters or fewer and 33 run to fifteen or more. Eleven names have sold.

Sort the same book by the roots a marketplace would pull out of it and the structure changes:

• 17 names carry "ai" as a token — 10 as a prefix (AIPhotoServices, AIWriterFlow, AILoanCompany) and 7 as a suffix (GenerateImagesAI, ImageFixerAI, SearchJobsAI). Another 5 are filed under our AI/SaaS bucket without the letters in the string at all.
• 6 names share the root "prompt" (DiscoverAIPrompts, ContentPrompting, AIPromptKing, AIPromptBible, PromptDemo, PromptJa).
• 6 share image/pic (GenerateImagesAI, FastImageAI, ImageFixerAI, GeneratePicsAI, AIPictureRestore, PickupCenters — the last one by accident of spelling).
• 5 share "loan", 4 share "home", 4 share "auto", 3 share "shop", 3 share "tech", 3 share "service".

Now the number that explains ShopUna. 41 of the 164 sit in our Branding and Brandable buckets, and 37 of those 41 are single coined strings — Alwam, Tunrio, MyInco, Fruoo, Crypva, Jishoo, Aveedy, Solrco. Fragments stitched into a syllable pattern rather than two dictionary words pushed together. Those 37 names are, by construction, names whose root-word extraction has nowhere clean to land. On a marketplace that indexes by root, that is not a branding advantage. It is a filing problem, and we created it at the registrar.

Three numbers that are not the same number

A listing page shows at least three figures, and sellers routinely conflate them.

The appraisal prices the name as a class. Atom prints one on the listing page: LemonWhite.com carries an estimated value of $2,800 against our $2,495 ask, and Chauffeurs.si an estimated $800 against a $199 ask. Both are the same kind of number — a comparable-set opinion about names that resemble yours — and neither knows what you intend to charge.

The brand score is what the marketplace uses to rank and sort names inside its own surfaces. Atom's own blog describes hand-reg collections ranked using its Brand Score, which is a placement instrument, not a price. A high score puts a name higher in a browse list. It does not make a buyer appear, and it does not move the appraisal.

The listing itself — images, category, possible uses, description, price, payment options — is what a visitor sees once the name has been surfaced by the first two.

The order matters. Root words decide the shelf. The score decides the position on the shelf. The listing decides whether the person standing in front of it acts. Optimising the third while the first is broken is the most common wasted afternoon in this business.

What a score can see, and what it cannot

A model scoring a name can see the string: length, extension, vowel balance, syllable count, whether it splits into dictionary words, how many live extensions the label holds, what comparable sales closed nearby, and whether the name contains a category token it recognises. All of that is in the string and in public registries.

It cannot see the buyer. It cannot know that Chauffeurs.si is the exact word a Slovenian transfer company types into a search box, that Excavators.si matches a contractor's fleet, or that LemonWhite.com will appeal to a skincare brand and to nobody else. It cannot see your traffic, your replies, or the fact that a $199 ask on a name with a $800 appraisal is a deliberate shelf-clearing price rather than a mistake.

So when the score disagrees with your instinct, the useful question is not "who is right". It is: which of the two is answering the question the buyer will actually ask? A buyer asks "is this the name for my business". A score asks "does this name look like the names that sell here". Both are real questions. Only one of them ends in a wire transfer.

The five checks we now run before touching a price

1. Print the root words. Open your own listing and read the block. If a root is not a word, the name is being ranked on a dead query, and cutting the price will not create a shelf that does not exist.
2. Check whether the category token survived. For us, "AI" did not survive extraction on ImageFixerAI.com. If your differentiator is a two-letter token, assume the index dropped it and make sure a real word carries the category instead.
3. Count your own overlap. Seventeen names sharing an "ai" token and six sharing "prompt" means we are bidding against ourselves on those shelves. In a portfolio above a hundred names, that concentration is a decision, not an accident, and it should be deliberate.
4. Write the appraisal and the ask on the same line. $800 estimated against a $199 ask reads very differently from $2,800 against $2,495. One is a name being cleared, the other is a name being held. Know which you are doing before an enquiry arrives.
5. Fix the string before the price. If a name's split is junk, the fix is structural — a hyphenless compound of two real words, a geo plus a service, a category word the model recognises — or it is patience. It is almost never a discount.

None of this makes the score wrong. It makes the score a narrower instrument than it looks: a placement readout built from a string, answering a question about resemblance. Read it that way and it becomes useful. Read it as a valuation and you will spend the next renewal cycle arguing with a parser.

Sources and method

Root-word blocks and appraisal figures were read directly from the live Atom listing pages for ShopUna.com, LemonWhite.com, ImageFixerAI.com and Chauffeurs.si on 17 October 2026. Portfolio counts come from our own master inventory of 164 live names and 11 sold, generated 13 October 2026: extension mix, buy-now distribution, label lengths and category buckets were counted from that file, and the coined-name count is our own classification of the 41 names in the Branding and Brandable buckets. The Brand Score reference is from Atom's public blog on hand-registration collections. No view counts, enquiry counts or revenue figures are estimated anywhere in this piece; where we do not have a number, we say so.