DomainTier
Methodology

How DomainTier values a domain

Every valuation is built from observable evidence (recorded sales, word-frequency data, structural scarcity) and combined through a transparent formula. No black box: the number, the inputs, and the confidence behind it are all shown.

The principle: evidence over opinion

A domain is worth what a buyer will pay, and the best predictor of that is what comparable names have actually sold for. DomainTier anchors every valuation to real transactions wherever they exist, and falls back to a calibrated structural model only when they don't.

The estimate is always presented as three figures forming one band: a liquidity floor (a fast, motivated-seller price), a market value (the realistic mid-point), and a strategic premium (what a category-defining acquirer might pay). A confidence score states how much evidence stands behind the number.

The formula

The displayed value is the strongest of four independent reads (a documented sale, a structural floor, a comparable-sales estimate, and an expert appraisal), then mapped to a retail figure and shown as a band.

$$ V(d) \;=\; \bigl(\,B \circ \Pi \circ S_{T,L}\,\bigr)\!\bigl[\, \operatorname{val}(d) \,\bigr],\ \mathit{where}\ \operatorname{val}(d) \;=\; \max\!\bigl\{\, a(d),\;\; \varphi(d),\;\; \hat{m}(d),\;\; A(d) \,\bigr\} $$
val(d)
Working value
The strongest of four independent reads selects the basis: a max-join, not a fragile product. A documented sale, a structural floor, the comparable-sales estimate, and an expert appraisal each get a vote; the best-supported figure carries. One deliberate exception: when the appraisal wins, it is blended against the name's own market evidence rather than adopted outright, so even the winning read never speaks alone. Nothing else is multiplied through a chain where one weak term sinks the rest.
a(d)
Documented sale
A recent recorded sale of the exact domain, blended against the name's own market evidence — a corroborated sale carries the figure, while a single out-of-line trade is capped and flagged rather than repeated, so a manufactured price can never set the value. An older sale still speaks, but its weight decays with age and it is blended against today's evidence rather than anchoring outright — and a wholesale auction print below the evidence reads as a floor on the name, never a ceiling.
φ(d)
Structural floor
The highest of four floors: scarcity (short pure-letter strings), recognition (domain hacks, famous phrases and idioms, tickers, known marks), class (extension × name-type), and real-entity demand: a company, ticker or product that wants the exact name, read from the entity ledger, recognized live, or evidenced by the site that recently operated it. Evidence from a former operator, read from an archived copy of the site, counts for at most twice the model's own read: a business that let the name lapse is not a buyer.
m̂(d)
Comparable-sales estimate
A calibrated model over the nine-factor vector x(d), fit to 1.3M recorded sales. The name's value from its own structure and what its neighbours have actually sold for, before any display shaping.
A(d)
Expert appraisal
For thinly-traded brandables and compounds the comparable set can't place, a rubric appraisal (length, brandability, sentiment, commercial pull) grounded in real retail sales. It lifts a comp-thin name toward its retail value, blended with the market read rather than adopted outright; it never pulls a name below the model. Names with independently verified high-tier entity demand are appraised regardless of word shape, and every name is appraised as the exact domain being valued. When a recent documented sale anchors the name, the appraisal serves as corroborating evidence for that sale instead. The two roles never stack.
ST,L
Retail step
Maps the wholesale figure toward a market value along two arms, taking whichever the evidence supports: the class's fitted wholesale→retail spread \(v\cdot s_p^{\,p_{\text{ret}}}\) (spread tamed by measured retail sell-through, calibrated per extension tier, length bucket and name class), or a demand-graded spread for names whose measured liquidity supports a stronger retail position. Where retail sell-through is rare the step stays close to wholesale; the full retail figure survives as the strategic ceiling, not the headline. The report names the arm that carried.
Π
Refinement pipeline
Ordered, class-specific adjustments layered on the base in sequence: short-name scarcity, domain-hack floors, digit quality for pure-numeric names (lucky-digit and pattern premiums, and a discount for a 0 or a 4 in a four-digit name, calibrated on real numeric sales), a premium-word blend that credits a genuine dictionary word in proportion to its measured buyer demand (bounded, never a free pass), a per-extension class band that positions a dictionary word on a specialty extension (round one: .ai) into its own measured sale band, placed by where words of its frequency tier actually clear, lift-only, never a canned price, demand-composed gravity for thin-evidence names on secondary extensions (a model read is scaled by the name's measured probability of actually finding a buyer: the correction for the fact that sold-price data only records the winners, applied at full strength on low-value inventory and fading out as the model's own read rises, so an established read stands on its own evidence), a junk guard that holds gibberish down, and live-auction corroboration: sustained competitive bidding on the exact name lifts its value toward the current bid, within a cap anchored to the name's own fundamentals and deepened only by real competition; in the closing hours of a deeply-bid auction whose price is still climbing, the cap widens so the value can track the market’s final consensus. Depth, closeness, and a rising trajectory must all hold, and a bid beyond even that widened ceiling is flagged rather than obeyed. Each step applies only to the class it targets, and each shows its arithmetic in the report.
B
Display band
The final figure as a triple (liquidity floor, market value, strategic premium): a tight band around the market figure, not three separate estimates.

The factor model

Nine factors form the vector x(d) behind the comparable-sales estimate m̂. Each is scored from observable data and weighted by the calibrated model.

Length
Shorter is scarcer. The pool of unregistered short .coms is effectively zero.
Extension
.com leads; tier-2 alt-TLDs (.io, .ai, .co, .net) and ccTLDs are scaled against it.
Pattern
Clean letter strings, premium numeric patterns, and recognisable structures.
Pronunciation
Pronounceable names are more brandable and recall better.
Category
Commercial sector words (finance, tech, health) carry industry-defining demand.
Comparables
Depth and relevance of recorded sales for the name and its neighbours.
Liquidity
Buyer-pool depth: how many industries could credibly brand on the name.
Dictionary
Real words and recognised colloquial terms outvalue coined strings.
Age
Registration history and prior ownership signal.

Comparable sales

The evidence stage draws on a database of 1.3 million recorded sales spanning public marketplaces, auction houses and aftermarket venues. The set is deduplicated across venues, price- and date-normalised to one canonical format, and screened for mis-records before use. A smaller, cleaned corpus is worth more than a larger noisy one. For each query the engine assembles several classes of comparable:

  • Exact same-TLD: a prior sale of the exact domain. The strongest evidence; used as a floor when the sale is representative.
  • Cross-TLD: the same name on other extensions, price-normalised to the queried TLD.
  • Similar-pattern: names of the same length and structural class, recency-weighted so stale sales don't dominate.
  • Semantic neighbours: names with related meaning, surfaced by vector similarity.

A skepticism pass filters dust (drop-catch noise, mis-recorded prices) before the set informs the blend. A recency-weighted median, robust to a single out-of-distribution sale, becomes the anchor that the evidence factor pulls the intrinsic value toward.

Liquidity & sell-through

The liquidity floor and the price-to-time-to-sale read depend on how often a name in a given class actually finds a buyer, and completed sales alone can't reveal that: sold names are only the numerator. So the sell-through signal is trained against the denominator too: hundreds of thousands of names that were listed or sent to auction and drew no bid. Modelling both outcomes, sold and unsold, is what lets a sell-through estimate mean something rather than being read off the winners alone.

The pipeline

A query flows through a modular pipeline. Each stage is independently testable.

01LookupTLD tier, dictionary & frequency data, brand & trademark screening (a local cross-check against millions of live registered marks, plus the domain's own site evidence: protected marks show no open-market price; brand-operated names are marked held, with the estimate reflecting acquisition cost; crowded names are flagged for collision), domain-hack detection.
02AnalysisWord quality, compound detection, category, pronounceability, buyer-pool depth.
03FloorsStructural pattern floor for the TLD, length and name class.
04EvidenceComparable retrieval, skepticism filtering, the evidence blend.
05OutputRetail step, confidence, value range, and the supporting commentary.

Confidence

Confidence is reported alongside the value, never multiplied into it. It rises with the strength of the evidence: a documented sale of the exact domain yields high confidence, decaying as that sale ages; a deep set of close comparables yields a strong-but-lower figure; a purely model-derived estimate is capped, because the engine cannot honestly claim certainty without market evidence. And when two independent reads of the same name disagree widely, confidence drops with the disagreement; the factor breakdown shows that discount as its own row. Live auction activity on the exact name is treated the same way evidence should be: sustained competitive bidding can lift the figure toward the current bid, but only within a cap anchored to the name's own fundamentals and the depth of the bidding. A lone bid, or any bid on a name without fundamentals, moves nothing.

Junk-shape names (long unpronounceable strings, random long numerics) are capped lower still. The goal is calibration: a stated confidence should mean what it says.

What this is not

A valuation is an estimate, not an appraisal or an offer. Thinly-traded categories carry real uncertainty, and the engine surfaces that rather than hiding it. Trademark exposure, registrant intent and live negotiation all sit outside the model. DomainTier flags brand-locked names but does not give legal advice.

And because a methodology is only as good as its results: every valuation filed on a live auction is graded against the price the name later fetches, daily, in public. The running scoreboard is on the track record page.

See it on a real domain

Run any name through the engine: the full breakdown, comparables and confidence.

Valuate a domain →