Why Automated Domain Appraisals Overvalue Names

Automated domain appraisals often overvalue names because they learn from data that skews high: asking prices instead of sales, the small share of sales that get reported, record deals that dominate the headlines, and retail prices mixed with wholesale ones. Without the comparable sales behind the number, a buyer or seller cannot see the drift, and a typical name ends up priced like an exceptional one.

This is not an argument against automation. Models are fast, consistent and useful for sorting large lists. The problem is the black box: a single number with no comps, no venues and no dates. This article explains where the upward drift comes from, using recorded sales from May 2021 to May 2026, and what to look for in any estimate you rely on.

Start with what the market actually looks like

Domain prices are extremely uneven. Across 128,867 recorded domain sales in that period, the median sale was 1,900 USD. Of those sales, 86.8 percent were below 5,000 USD. Only 297 sales, fewer than one in four hundred, reached 100,000 USD or more.

The average recorded sale was 4,094 USD, more than double the median. That gap alone is a warning. A handful of very large sales pull the average up, and any model that leans on averages, or on the sales people talk about, will drift above what a typical name fetches.

Reason 1. List prices are not sales

The easiest data to collect about domains is what sellers ask. Marketplaces display millions of listed prices. Few of those names ever sell at the listed price, and many never sell at all.

If a model treats listed prices as a signal of value, it learns what owners hope for. Owners are, understandably, optimistic. The result is an estimate that matches the asking price of a similar name on a marketplace, which may have been sitting unsold for years.

A sound estimate should rest on completed sales, where money changed hands at a known venue on a known date. If an appraisal cannot tell you which sales it used, you cannot tell whether it is built on sales or on hopes.

Reason 2. Survivorship in reported sales

Most domain sales are never reported. Public sales reports typically list sales above a floor (the reports our data comes from start at 1,000 USD), and private deals are reported only when a party chooses to share them. The many names that sell for a few hundred dollars, or that never sell, leave no trace in the data.

This bias applies to every dataset built from reported sales, including ours. A name type that often sells for 300 USD can appear in the data only through its rare 1,500 USD sales. A model trained on that data sees the 1,500 USD and not the many 300 USD trades below it.

The honest response is to say so. When no comparable recorded sale above 1,000 USD exists, the right conclusion is that names like this usually trade below that floor. Producing a figure anyway, extrapolated from unrelated sales, is how low-value names receive four-figure estimates.

Reason 3. Record sales anchor the model

Headlines are made by outliers. icon.com sold for 12,000,000 USD at Hilco Digital Assets on 2025-04-24. help.com sold for 3,000,000 USD at Above.com on 2023-01-10. commerce.com sold for 2,200,000 USD at GetYourDomain.com on 2025-02-18. These sales are real and worth studying.

But they are not representative of one-word .com names in general. Many one-word .com names in the data sold for a few thousand dollars. If a model gives heavy weight to features like "one word" or "dictionary term" because they appear in the biggest sales, it will lift the estimate for every name that shares those features, whether or not buyers ever paid those prices for anything similar.

The fix is to weight the cluster of close comps, not the famous ones. A good estimate shows you the cluster.

Reason 4. Missing venue context

The same name can sell for very different prices depending on where it is sold. Expired auctions and investor venues, such as DropCatch, NameJet, Park.io, Dynadot and Namecheap, are wholesale markets where investors buy from each other. End-user marketplaces, such as Afternic, Sedo, Atom.com and BuyDomains, are retail markets where businesses buy names to use.

In the recorded data, wholesale venues show a median sale of 1,725 USD across 30,021 sales, and retail venues show a median of 2,288 USD across 23,816 sales. The averages differ more: 3,312 USD wholesale against 6,437 USD retail. In .ai the gap is wider, with a retail median of 9,995 USD against 2,000 USD at wholesale venues.

Single names show how far the gap can stretch.

A model that blends both markets into one number produces a figure that is too high for a quick sale to an investor and possibly too low for a patient sale to the right end user. When an owner sees only the single blended figure, they tend to treat it as what they can get now. That is usually the wholesale market, where the number is too high.

Reason 5. Old comps and trend peaks

Prices move. Some segments rise quickly, and some cool after a trend. In .ai, recorded sales went from 152 in 2021 to 2,070 in 2025. Individual names rose sharply: stack.ai sold for 4,449 USD at Whois.ai on 2021-07-06 and for 72,999 USD at Dynadot on 2024-06-03.

Timing can push a model in either direction. Using 2021 sales for .ai today would undervalue. Using peak sales from a hot quarter for a segment that has since cooled would overvalue. Without dates on the comps, the user cannot judge which way the error runs. An estimate that shows when each comp sold lets you make that call yourself.

Reason 6. Features credited too generously

Automated systems score names on features: length, extension, dictionary status, keyword popularity, number of words. Each feature is reasonable. The trouble comes when the scoring gives credit for features that buyers do not pay for in practice, or adds credits that overlap.

Plural forms are a good example. Sometimes the plural sells as well as the singular: rebate.com and rebates.com each sold for 500,000 USD at Sedo on 2022-12-01. But that is a rare case. A model that ignores the plural discount entirely will overvalue the many plural names that do not have a matching buyer. Length works similarly. In .ai, two-character names had a median recorded sale of 5,049 USD, while three-character names had a median of 1,826 USD. Being short is not enough by itself.

Comparable sales check the features against reality. If the closest sold names with the same features sold for 2,000 USD, the features are worth about 2,000 USD in that context, whatever the model believes.

Reason 7. No confidence, no admission of thin data

A single number looks equally precise whether it rests on fifty close comps or on none. That uniform confidence is part of the problem. A name with few comparable sales deserves a wide range and a low confidence label, and a name with no comparable sales above the reporting floor deserves no specific figure at all.

When a tool always produces a number, the hardest cases, which are usually the weakest names, get the same confident presentation as the easiest ones. Owners of weak names see a number and assume it is as solid as any other.

What a trustworthy estimate shows

You do not need to avoid automated tools. You need to ask more of them. A trustworthy estimate shows:

• The comps. Each comparable sale by name, price, venue and date, so you can check it.
• Why each comp is relevant. Same extension, same length, shared word or meaning.
• The venue split. A wholesale figure and a retail figure, not one blended number.
• A range and a confidence level. Low, likely and high, with a label that drops when comps are thin.
• What would move it. What evidence would raise or lower the estimate.
• An honest floor. When nothing comparable sold above the reporting floor, it says so instead of inventing a figure.

If you are comparing approaches, our pages on the EstiBot alternative and the domain value checker describe how a comp-based estimate differs from a model-only score. Check each provider's own site for their current features and pricing.

How SoldComparables approaches it

SoldComparables builds every estimate from recorded sales: 128,867 of them between May 2021 and May 2026, each with its venue and a link to the public report it came from. It does not use asking prices. Every comp gets a relevance score from 0 to 100 with up to three plain reasons, every sale is labelled wholesale, retail or mixed venue, and every estimate has a low, likely and high value, a wholesale and a retail figure, a confidence level and a rationale that cites the comps by name, price and date.

When there are fewer than three good comps, the estimate says so, widens the range and lowers confidence. When no comparable recorded sale above 1,000 USD exists, it says that names like this usually trade below that level and does not produce a number. It is a data-backed estimate, not a certified appraisal and not a guarantee of sale price. To see the comps behind a name you own or want to buy, run a domain appraisal search.

Frequently asked questions

Are automated domain appraisals useless?

No. They are quick and consistent, and they help sort long lists. The risk is treating a single number without visible comps as a price. Use automated scores as a first filter, then check the actual comparable sales before you set a price or make an offer.

Why is my appraisal so much higher than the offers I get?

Offers usually come from investors paying wholesale prices, while many appraisals blend wholesale and retail data or lean on asking prices and record sales. The offer reflects the investor market. A retail buyer may pay more, but finding one takes time.

What is a typical domain sale price?

Among 128,867 recorded sales from May 2021 to May 2026, the median was 1,900 USD and 86.8 percent were below 5,000 USD. Many more names sell for less than 1,000 USD, but those sales are generally not reported, so the true typical price is lower still.

How can I tell whether an estimate is reliable?

Look for the comparable sales behind it, with venue and date for each, a range rather than a single number, separate wholesale and retail figures, and a confidence label. If none of that is visible, treat the number as a rough guess.