What the model actually uses.

Every estimate on this site is a model output, not an appraisal. This page describes what goes in, how the three channel figures are produced, and where the method breaks down.

The pipeline

  1. 01

    Domain text, TLD, and linguistic features

    The name is tokenized and represented as a vector: word boundaries, length, TLD, and how those words behave in ordinary written English.

  2. 02

    Historical auction records

    That representation is compared against reported sales: roughly three million auction transactions spanning about twenty years.

  3. 03

    Sale-channel model

    A prediction is produced per sale route, because the same name clears at different prices at auction, on a marketplace, and through a broker.

  4. 04

    Three estimates with ranges

    The output is a probabilistic range per channel rather than a single number, which is what the underlying data actually supports.

Data sources

The model is only as good as what it has seen. These are the inputs, in rough order of how much they influence a figure:

  • Historical auction data

    A twenty-year dataset from DNPric.es covering more than three million domain auction transactions. This is what anchors the price scale.

  • Public marketplace sales

    Sales platform data, including GoDaddy Auctions, which widens the comparable set beyond auction records.

  • Dropped domains

    Names that owners chose not to renew. A direct signal about which structures the market has decided are not worth carrying.

  • Active website domains

    Names currently in commercial use, which describe what a working brand domain tends to look like.

  • Open language models

    Pre-trained embeddings from openly published models by Microsoft, EleutherAI, and Hugging Face supply the language understanding layer.

Why there are three numbers

A domain does not have one price. It has a price per route to market, and the spread between those routes is often three to ten times.

ChannelBuyerWhat it means
AuctionInvestor to investorFast, competitive, and usually the lowest of the three. Bidders are buying to resell, so they price in their own margin.
MarketplacePatient retail listingA public listing with no deadline. Higher than auction, and dependent on an end buyer finding the name.
BrokerageBroker-assisted negotiationAn outbound, negotiated sale to a specific buyer. Highest of the three, slowest to realise, and not achievable for every name.

The range shown against each figure widens as the value drops, because low-value names have noisier comparable data.

Known limitations

These are the cases where a HumbleWorth estimate is least useful:

  • Trademark exposure and legal risk are not assessed.
  • Existing traffic, backlinks, and revenue attached to a domain are not measured.
  • A single motivated buyer can pay a multiple of any estimate here, and often does.
  • Non-English names and unusual TLDs have thinner comparable data, so their ranges are less reliable.
  • Training data extends through early 2024, so the most recent market movement is not represented.
  • Estimates reflect the market the model was trained on, not a forecast of where it is heading.

Every estimate on this site comes from one pinned version of the model, not from something that retrains on a schedule. What it is good for is comparing names against two decades of recorded sales; what it cannot do is tell you where the market moved last quarter.

Those are the limits by construction. The accuracy page is the measured version: the auction estimates scored against 2,812 auctions we watched close, including a set where the ordering came out backwards.

Running the model yourself

The estimate model is published on Replicate and can be called directly, in batches of up to 2,560 domains at about $0.10 per 1,000 estimates. HumbleWorth adds no limit of its own there; Replicate's account limits and pricing apply, and both are theirs to change. See the API documentation for request and response formats.

Read next