how accurate are domain appraisals?.

an automated domain appraisal is a sorting key wearing the costume of a price. genuinely useful for ranking a hundred names against each other, near-useless as an expectation of what any one will sell for.

that is the whole answer. the rest explains why, because the distinction between a relative signal and an absolute one decides whether the tool helps you or costs you a sale.

what these tools actually compute

every automated valuation is a scoring function over features of the string, plus a lookup against reported sales. the features are broadly the same everywhere:

• length, with a steep premium on very short names.
• character composition — hyphens and digits mark a name down sharply, as do letter combinations awkward to say aloud.
• dictionary membership, and whether the word is common, a brandable coinage, or a compound of two known words.
• keyword demand signals, usually borrowed from search advertising: what traffic the phrase attracts and what advertisers pay for it. a proxy for commercial intent, not a measurement of the domain.
• the TLD, with .com weighted far above everything else and other extensions discounted by rule of thumb.
• comparable sales — the model's record of what similar names reportedly changed hands for.

none of that is unreasonable, and those features do correlate with value. the problem is not the features. it is what the model is fitted against.

the training data is structurally broken

three problems, and they compound.

only sales are recorded. the datasets underneath these models are lists of domains that sold. there is no record of the far larger population of names listed for years and never sold, so the model has never seen the most common outcome. a valuation trained on completed sales tells you what a name might fetch conditional on it selling, and omits the probability of that condition.

reporting is skewed by size. large sales get announced because announcing them suits everyone involved. small ones frequently are not, and private transactions never are. the visible record over-represents the top of the market.

declined offers are invisible. a seller who turned an offer down has generated no data point, and neither has a buyer who walked away. the model sees prices people agreed to, and nothing of the more numerous ones they refused.

you can build a competent model on that data and it will still be wrong in a specific direction, because the data is a survivorship sample.

why realised prices diverge so far

the deeper issue is that domains are not a liquid asset class, and valuation methods assume liquidity. a share has a price because thousands of people trade identical units continuously. a domain has one unit and, ordinarily, either zero interested buyers or one. its value is not a clearing price — it is the number a single buyer, with a budget and a reason to want that exact string, agrees to on a particular afternoon.

that produces enormous dispersion between names a model cannot tell apart. two four-letter .coms with similar keyword profiles can differ by orders of magnitude because one is the acronym of a company that just raised money. no feature can see that, and no model improvement fixes it, because the information is not in the string.

time is the other missing variable. an appraisal is quoted as a number; the honest form is a number and a horizon, because what a name might fetch given an unbounded wait is not what it fetches this month.

the incentive problem

worth naming carefully, without accusing anyone of anything. several of the best-known appraisal tools are operated by the marketplaces that then list, broker and take commission on the names being valued, and that also earn the registration and renewal revenue on names their customers hold. an estimate encouraging a listing, or a renewal, benefits the operator; one saying 'this is worth nothing, drop it' does not.

that is not a claim that anyone inflates numbers. it is an observation that the incentive gradient points one way, that these models are audited by nobody outside the company running them, and that the result is not disinterested in the way an independent index would be. both things are true at once: they are built by people who understand this market extremely well, and by people who make money when you act on the output.

the practical consequence is narrow. do not quote an automated appraisal to a buyer. they know which tool produced it and that anyone can generate one free in seconds, so it carries no information — and leading with it signals you have no better basis for your price.

what appraisals are genuinely good for

relative ordering. rank is far more robust than magnitude.

run two hundred domains through one tool and the model's errors are largely shared across the list, so the ordering is broadly meaningful even where the absolute figures are not. the top of the list really does hold your better names. that makes it a triage instrument, and triage is the job most people actually have — deciding which twenty of two hundred names deserve attention at renewal.

it also works as a floor-setter in one direction. a name ranked at the very bottom against every comparable in your own portfolio is a real signal, because being consistently unremarkable is the one thing these models detect reliably.

how to sanity-check a number

• look at comparable sales yourself, same TLD and same pattern, rather than trusting internal comparables you cannot inspect.
• apply the reacquisition floor. if you dropped this name tomorrow, what would an acceptable replacement cost? usually a registration fee plus the total cost of ownership of holding it — the honest lower bound.
• run the name through more than one tool. wide divergence is common and is itself the finding: it tells you the model has low confidence where the interface displays none.
• check the TLD assumption. extension weighting is where these tools are crudest, and where perception shifts faster than models update. TLD trust rankings is a different lens on it.
• ask whether you would pay it. if not, you have learned something about the number rather than about the domain.

how we use it, and what we say about it

owndle attaches an estimated value to every domain in a portfolio for one purpose: so a large list can be sorted by whether each name is worth renewing. it sits beside the renewal price, which is the comparison that matters — a name valued below what you pay to hold it every year is worth a conversation.

we call it a heuristic in the product and in the methodology, and we do not present it as an appraisal, because we have no more access to the counterfactual sales data than anybody else. portfolio valuation is a column to sort by, next to a renewal price from the live catalogue. the sort is the product; the number is scaffolding.

if you are pricing one specific name for one specific sale, no automated tool will give you the answer. a broker with recent private comparables in your niche might. otherwise the price is discovered the old way, by asking and finding out.