Automated domain appraisals and their limits

An automated appraisal produces a figure from the characteristics of a string. The models behind them are trained on public sale records and on the characteristics that accompany those records, and they perform reasonably at the task they were built for: sorting large volumes of inventory into rough bands quickly.

Problems begin when the output is treated as a price rather than as a sorting signal.

What the models measure

The inputs are the observable properties of a name and of the market around it:

• Character count, and whether the string contains hyphens or digits
• The extension, treated as a categorical variable with its own multiplier
• Whether the string decomposes into known words, and how frequently those words appear in language corpora
• Search and advertising demand associated with the term
• Registration depth: whether the same string is registered in other extensions
• Similarity to previously reported sales, which is where most of the price signal comes from

None of these is unreasonable. Together they capture a real part of what makes a name valuable, which is why the outputs correlate with reality at the extremes: models are usually right that a two word exact commercial term in a strong extension is worth more than a random eleven character string in a weak one.

Where they fail

The failures are structural rather than a matter of tuning.

The model cannot see the buyer. Price in this market is set by whether a specific party needs the name and what their alternatives are. That information does not exist in the string, so the model estimates a market average across buyer types that are separated by a wide range.

The model cannot see how the stronger name is being used. Whether the equivalent in the main commercial extension is unregistered, parked or running a live competitor changes the value of every alternative, as described in the extension effect. Few models check this and none interpret it well.

The training data is biased. The models learn from reported sales, and the public record systematically omits private deals and unsold inventory, as set out in comparable sales. A model trained on outcomes without attempts overstates the probability that any given name is saleable.

Language handling is shallow. Compounds that read as something unintended, terms that are commercial in one national market and meaningless in another, and plural forms that no trade actually uses are all scored as if they were clean.

Trade mark exposure is ignored. A model will happily value a string that reproduces a registered mark, which is a liability rather than an asset, as covered in the trade mark screening notes.

Outputs cluster and anchor. Because similar inputs produce similar outputs, large parts of a portfolio receive near identical figures, and holders anchor asking prices to those figures. Buyers are aware of this and discount an appraisal presented as evidence, particularly one produced by the venue where the name is listed.

Using an appraisal sensibly

• As a filter across a large portfolio, to separate names worth individual attention from those that should be reviewed for renewal.
• As a consistency check: when an estimate differs sharply from a considered manual view, the disagreement is worth investigating, in either direction.
• As a relative measure between similar names in the same extension, where the model's systematic biases apply to both and largely cancel.
• Never as a negotiating instrument. An appraisal presented to a buyer invites the buyer to produce a lower one, and the conversation moves from the name to the credibility of two machines.

A model can tell which names in a portfolio deserve an hour of work. It cannot tell what any of them will sell for.

The manual work the model replaces badly

Everything the model omits is work a person can do in a few minutes for a name that matters: check who is operating on the term, check what the equivalent in the stronger extension is doing, check the registers for conflicting marks, and check whether the term is used commercially by real businesses rather than merely searched for. That sequence, set out in demand signals, produces a defensible view. An automated figure, however precisely stated, produces only a number.