VOL. I · NO. 01 · POSITION PAPER · DOMINIUM · MAY MMXXVI
The Permanence
Doctrine
In an era of fleeting platforms and algorithmic volatility, the digital land remains the only sovereign constant.
A field guide to digital real estate in the era of the agent — and the case for why names are entering their era of greatest significance.
§ THESIS
The agentic web does not abolish the domain. It exalts it. When humans delegate navigation to agents, the question is no longer which URL will a person type — but which name will the agent trust.
PART ONE
The Short Version
If you only have five minutes, this is everything you need to know.
A domain is digital real estate. That's not a metaphor.
It's the closest thing the internet has to actual property. And like real estate, the rules are simple:
Location matters. A short, memorable, category-fit name is Fifth Avenue. A long, awkward one is a dirt road off the highway.
Scarcity drives value. There is exactly one Bank.com. There is exactly one Insurance.com. ICANN does not build new beachfront — the supply of premium names is fixed and shrinking, because every great name registered today is one fewer left for tomorrow.
Age compounds. A domain held continuously since 1998 with a clean record is fundamentally different from one registered yesterday. You cannot manufacture history.
What people keep getting wrong.
Every few years, someone announces that domains are dying.
First it was apps that were supposed to kill them. Then social media. Then voice assistants. Now AI agents. Every prediction has made the same mistake — confusing the interface on top with the infrastructure underneath.
Apps run on domains. Social platforms link to domains. Voice assistants look up domains. AI agents read, rank, and recommend domains. The address layer hasn't moved in thirty years. It isn't going to.
A DOMINIUM POSITION PAPER / DOMAIN STRATEGY
— Issued by the Dominium portfolio
Why AI actually makes domains more valuable.
This is the part almost everyone is still getting wrong.
When you type a URL, you can type anything. You're free to be curious. You can visit a sketchy site if you want.
When an agent picks a URL for you — ChatGPT booking your flight, Claude finding you a contractor, Perplexity researching a purchase — it has to choose. And it can't afford to be wrong. So it leans hard on signals that scream legitimacy.
Old domains. Clean records. .com extensions. Names that match real companies. Names with Wikipedia pages and press coverage.
The same names that mattered when humans typed are the names AI picks now. Except there's a twist:
Traditional search used to show you ten results and let you decide. AI search shows about four. Fewer winners. Bigger winners. The names with authority capture more attention than before, not less.
Trust signals, in plain English.
When an AI agent evaluates a website, it checks a handful of things — fast:
1. Is the domain old? Older is better. New is suspicious.
2. Is the extension respectable? .com beats .xyz beats .top. Always.
3. Does the name match the company? Mismatches are a red flag.
4. Does anyone respectable mention this brand? Wikipedia. Reddit. Major publications.
5. Is the technical setup clean? SSL history, DNS records, no abuse complaints.
A name that looks shady to a human looks shady to an agent. The signals are shared. The systems agents inherit from — search engines, certificate authorities, trust databases — were built over twenty years to reward authority. Agents just enforce them harder.
The extension hierarchy, in plain English.
Not all extensions are equal. The order matters more than most people realize:
.com — The global default. Embedded in every AI's training data. The only true universal address.
.ai — The new category address for AI businesses. Registrations grew roughly nine-fold from 2022 to 2025. The companion play to .com — own both.
.org · .edu · .gov — Institutional trust. Gated or near-gated. Cannot be bought casually.
.net — Heritage commercial. Still respected, paired with .com in the registry that runs both.
.io — Tech-startup credibility, but slowly losing ground as serious companies graduate to .com.
Country codes (.uk · .de · .ca · .jp · .au) — Strong regional trust when used by local operators.
.top · .xyz · .shop · .click · .loan — Used so heavily for spam, phishing, and toll-road scams that legitimate operators on these extensions pay a permanent reputation tax that no amount of good behavior can erase.
The takeaway.
A great domain on a great extension does three things at the same time:
- 1. It tells humans you're real.
- 2. It tells AI agents you're trustworthy.
- 3. It compounds in value every year because the supply is fixed.
The internet has cycled through five interface revolutions in twenty years. Search bars. Social feeds. App icons. Voice assistants. Agents. Each one new. Each one temporary.
The address layer hasn't moved. It never will.
"Names are permanent. Interfaces are temporary. The portfolio compounds."
That's the short version. If you want the data, the studies, and the deeper market analysis — keep reading.
PART TWO
The Long Version
For those interested in the more complicated and deeper understanding of the industry and market in 2026 — well, here you go.
I
§ FOUNDATION
MOD. 01 / DOCTRINE
Digital real estate is the cleanest asset class on earth.
Real estate works because of three properties: scarcity, location, and durability. Domains satisfy all three more cleanly than any physical property does, because the ledger is global, indivisible, and enforced by infrastructure rather than law.
Scarcity is absolute. There is exactly one CapitalOne.com. There is exactly one Bank.com. ICANN does not zone new beach frontage; the .com namespace cannot be expanded. The meaningful subset — short, pronounceable, semantically loaded — is vanishingly small and shrinking, because every premium name registered today is one fewer for tomorrow.
Location, online, is semantic. A name's "address quality" is measured by category fit, character economy, and the cognitive ease with which a person — or an agent — resolves it to a meaning. A category-fit dictionary .com is a corner lot on Fifth Avenue. A coined wordmark is a private estate. A junk-TLD typosquat is a sketchy storefront on a service road. The hierarchy is not opinion. It is reflected in price, in renewal behavior, and increasingly in how agents weight the names they encounter.
Durability is the most underappreciated property. A domain held continuously since 1998 with consistent WHOIS, clean abuse history, and an unbroken certificate chain is structurally different from a name acquired yesterday. Authority signals compound. They do not transfer easily. They cannot be manufactured retroactively.
The market does not agree the asset class is in decline.
The industry's own numbers contradict the narrative that domains are dying.
386.9M
Total domains registered globally · End Q4 2025
173.5M
.com and .net registrations combined
+6.2%
Year-over-year growth in 2025
1,265
TLDs delegated to the root zone
SOURCE · VERISIGN DOMAIN NAME INDUSTRY BRIEF, Q4 2025
Behind these numbers sits a parallel market in disclosed sales that has accelerated through the AI era. The premium aftermarket is not behaving like a dying asset class. It is behaving like a consolidating one — fewer transactions at the top, higher prices, longer holds.
PREMIUM SALES · DISCLOSED AFTERMARKET
$185M / +32.8% YoY
Public disclosed domain sales totaled roughly $185 million in 2024 — a 32.8% jump over the prior year. Headline transactions included Rocket.com at $14M, Gold.com at $8.5M, and the April 2025 sale of Icon.com at $12M. Public domain sales in Q1 2025 jumped 121% over the prior quarter. The all-time leaderboard — Voice.com at $30M, Chat.com at $15.5M, and the legendary Cars.com valuation of $872M — continues to be dominated by single-word .coms with broad brand surface.
SOURCES · VERISIGN DNIB Q4 2025 · NAME.COM · VERPEX 2024 SALES REVIEW
And one TLD outside the .com namespace has moved with extraordinary velocity: the .ai namespace, which has gone from a curiosity to an institutional category in three years.
.AI NAMESPACE · THREE-YEAR TRAJECTORY
60k → 551k
.ai domain registrations grew from approximately 60,000 in 2022 to 551,000 by January 2025 — roughly a nine-fold expansion in three years. The keyword "AI" was the most-searched term on the Sedo aftermarket in eight of twelve months across 2024. A category TLD is being built in real time, and the institutional play — owning the .com and the .ai of the same name — has emerged as the dominant defensive posture for AI-adjacent brands.
II
§ THE INVERSION
MOD. 02 / DOCTRINE
Agents concentrate domain authority. They do not dissolve it.
The standard objection runs: if agents do the browsing, who needs the domain?
This argument inverts the truth. When a human typed a URL, the domain was an interface — a string the human had to remember. The pessimistic case treated the domain as a relic of that interface. But agents do not abolish the address. They concentrate it.
An agent resolving a query needs a ground-truth endpoint. It cannot synthesize a new endpoint. It must pick from the names that exist, and weight that choice using authority signals that converge on the same names humans already trust. The result is not the disappearance of brand value. It is its consolidation into the names that already carry institutional weight.
The shift is happening at scale, and quickly. The agentic web moved from research demo to mainstream consumer surface in 2025. Traffic from autonomous browsing agents — Perplexity Comet, ChatGPT Atlas, Gemini in Chrome, Microsoft Copilot Mode, Opera Neon, and a thickening lineup of dedicated agent runtimes — went from a rounding error to a measurable share of commercial intent within a single calendar year.
+1,300%
Agentic browser traffic growth, Jan–Aug 2025
87%
Of AI agent visits are product-related
SOURCES · SNOWPLOW ANALYTICS · HUMAN SECURITY DATA
Now consider what happens to the address layer under that pressure. Traditional search returns ten links. The user re-ranks, clicks, and reads. The address layer is wide — many sites get a partial click. LLM-based search engines do not behave that way.
THE CONCENTRATION MECHANIC
10 → 4.3
Recent research analyzing 55,936 queries across six LLM-based search engines and two traditional search engines found that AI search systems return roughly 4.3 URLs per response, versus around 10 for traditional search. Fewer slots, harder selection, more weight on each chosen source. And the citation pool is genuinely different — 37% of domains cited by LLM-based engines do not overlap with traditional search results for the same query. A new authority economy is being selected for, in real time, with fewer winners per query.
SOURCE · ZHANG ET AL., "SOURCE COVERAGE AND CITATION BIAS," ARXIV:2512.09483 (DECEMBER 2025)
The risks of pure agent autonomy — recently demonstrated in the so-called Scamlexity research, where commercial AI browsers cheerfully completed purchases on phishing storefronts; in Cometjacking, where prompt-injected URL parameters exfiltrated users' email and calendar data; and in the broader collapse of traditional browser-trust assumptions under autonomous agents — all push in one direction. Agents that cannot afford to be tricked will lean harder, not less, on classical trust signatures: established domains, clean WHOIS history, real legal entities, valid certificates, structured-data discipline, brand-name coherence.
"The user who once typed amazon.com still ends up at amazon.com when an agent shops on their behalf. The domain is no less owned. The agent is simply the new courier."
+131%
Additional MoM growth, Aug → Sept 2025
90M
U.S. adults projected to use AI for search by 2027
III
§ TRUST SIGNALS
MOD. 03 / DOCTRINE
What agents actually evaluate.
Agents do not navigate by aesthetics. They evaluate signals — many of them inherited from search infrastructure, all of them rewarding authority. The signals that mattered to Google still matter to Claude, ChatGPT, and Gemini. They are simply being applied with less tolerance for ambiguity, against a smaller citation pool, with the user no longer in the loop to override a bad pick.
Below is the operational stack. Each line item is not a brand preference. It is something agents can read, score, and weight.
01 — DOMAIN AGE
A long, continuous WHOIS history is treated as evidence of legitimacy. New registrations carry an authority deficit that takes years to repay. Agents inherit this prior from the search and certificate infrastructure they sit on top of — a name created yesterday is structurally suspicious; a name registered in 1998 is not.
02 — TLD AUTHORITY
A .com or a respected ccTLD carries presumptive trust. Junk TLDs — frequently abused for spam, phishing, and toll-road scams — trigger discounts and, in some pipelines, outright filtering. The differential is measurable and growing. (See Section IV.)
03 — CERTIFICATE CONTINUITY
Unbroken SSL history and absence of revoked certificates signal operational discipline. Certificate Transparency logs are public; agents and security tools read them as a ledger of legitimacy.
04 — DNS HYGIENE
DNSSEC adoption, stable MX records, and clean reverse-DNS all factor in. They are quiet signals — never marketed, rarely visible to humans — but they are first-class inputs to the systems agents inherit from.
05 — STRUCTURED DATA
Schema.org markup gives agents typed entities to reason over. Unmarked sites are increasingly invisible to the new browsing layer. As AI retrieval infrastructure matures, the industry is converging on a discipline now sometimes labeled Agentic Interaction Optimization — structuring the page so a non-human reader can extract action, not just text.
06 — BRAND-MENTION GRAPH
External mentions on authoritative sources — review platforms, publications, industry directories — are the strongest single predictor of citation. Recent retrieval research shows brand search demand correlates roughly 0.334 with LLM citation likelihood — stronger than the backlink signal that defined the prior era. Sites cited across four or more AI platforms are 2.8× more likely to be cited in ChatGPT than single-platform sites.
07 — NAME–ENTITY COHERENCE
When the legal entity, the trademark, and the domain string all match, the trust score multiplies. A mismatched name is a friction point an agent will quietly route around. Owning the canonical name is not vanity — it is alignment between the address an agent picks and the entity the user actually wanted.
08 — WIKIPEDIA & EARNED MEDIA
AI search systems display a marked preference for third-party authoritative sources over brand-owned content. Roughly 47.9% of ChatGPT's top-10 citations are sourced from Wikipedia. Reddit dominates Perplexity and AI Overviews citation distributions. A domain without an off-site reputation graph is, increasingly, an unfindable domain.
The compounding insight: agents read the same signals search engines do, with the same biases — and with even less tolerance for ambiguity. A name that looks suspicious to Google looks suspicious to Claude, ChatGPT, and every browsing agent built on top of them. The trust stack is shared infrastructure.
DOMAIN AUTHORITY · STILL ~35% OF THE CITATION DECISION
Across multiple analyses of LLM citation patterns, domain authority signals continue to account for roughly 35% of citation likelihood. The definition has expanded — brand mentions, review-platform presence, and cross-platform visibility now share the weight that backlinks once carried alone — but the directional finding is consistent. Authority is doing more work in the agentic era, not less.
IV
§ TLD HIERARCHY
MOD. 04 / DOCTRINE
The hierarchy, with status notes.
Not all addresses are equal. The hierarchy has calcified.
T
he TLD is the first signal an agent reads, and the ranking is no longer in dispute — even if the secondary market hasn't fully priced it in. The case rests on three converging data sets: renewal behavior (which separates real use from speculative registration), training-corpus composition (which embeds preference into every commercial LLM), and abuse rates (which give junk TLDs a permanent trust tax).
Renewal rates reveal what registrations alone hide.
The headline registration numbers can mislead. The new gTLD category grew nearly 30% in 2025. But registrations are easy to buy. Renewals are not. When the second year of fees comes due, speculation evaporates and only real use remains — and the renewal-rate spread between mature TLDs and the new-gTLD category is brutal.
THE RENEWAL SPREAD · 2025
.org · 76.9% renewal — mature institutional
.com / .net · ~75% renewal — mature commercial
New gTLDs · aggregate · 31.3% renewal — speculative dumping ground
.shop · ~14.5% renewal — almost entirely transient
A renewal rate below 35% means the namespace is, in aggregate, a parking lot — a place names go to be registered cheaply and abandoned. The signal an agent receives from a name in that namespace is, on the prior alone, suspicion.
SOURCE · VERISIGN DNIB Q3 / Q4 2025 REPORTS
The training-corpus effect is structural.
There is a second mechanism — quieter, but more permanent — that locks .com into agent preference. Every commercial LLM is trained, in significant part, on Common Crawl. GPT-3 drew approximately 82% of its raw training tokens from Common Crawl. Roughly 64% of all major LLMs published between 2019 and 2023 used a filtered version of the same source. And Common Crawl, as a corpus, is structurally overrepresented in English-language, .com-domain content.
This means .com is not merely popular in the agent layer. It is linguistically embedded. When an agent reaches for a canonical name to recommend, the prior probability that it surfaces a .com is higher than the prior probability that it surfaces any other TLD — by a margin that no marketing budget can close.
"The flywheel: .com is overrepresented in training data → agents are trained to expect and trust .com → agents recommend .com → users see .com → the next training run reinforces the pattern."
.com
The global default. Dominant in training corpora, indexing weight, certificate volume, and brand convention for three decades. Agent priors favor it on every dimension. The aftermarket continues to set records for single-word .coms.
~159M registrations · ~75% renewal · embedded LLM bias
APEX · DEFAULT
.ai
The fastest-rising serious TLD in the modern era. A category address for AI-adjacent services, accelerated by the AI build-out and the willingness of well-funded operators to anchor there. Complement to .com, not substitute — the institutional play is to own both.
+9× registrations 2022–2025 · top Sedo keyword 8/12 months 2024
RISING · CATEGORY
.org · .edu · .gov
Institutional authority that no commercial TLD can replicate. .gov and .edu are gated. .org carries the highest renewal rate of any major TLD — a signal of serious use, not speculation.
.org renewal ~77% · highest in the major TLD category
INSTITUTIONAL
.net
A legacy commercial TLD with strong durability and certificate hygiene. No longer the rising star it was in the 1990s, but it sits in the same trust band as .com for most agent purposes.
~14M registrations · paired with .com in Verisign registry data
HERITAGE
.io
Retains technical credibility from a decade of developer-culture adoption. Slowly losing ground as software companies mature into more conventional naming. The Chagos sovereignty transition introduces a long-term registry-stability question worth watching.
Strong incumbency · uncertain long-horizon governance
TECH · MATURE
.uk · .de · .ca · .jp · .au
Country-code TLDs with strong institutional bases carry presumptive regional trust. Agents serving local queries will continue to weight them appropriately. The top ten ccTLDs collectively represent roughly 145M registrations — comparable in scale to .com itself.
145.6M ccTLD registrations globally · +3.4% YoY in 2025
ccTLD · REGIONAL
.top · .xyz · .shop · .click · .loan
Vanity and ultra-cheap gTLDs that have become structurally compromised by abuse. Eighteen of the top twenty TLDs ranked by spam-domain score are new gTLDs. Eight of the top twenty highest-threat TLDs are operated by a single registry that aggressively undercuts on price. Across the top ten highest-risk TLDs, roughly 53% of registered names are pure numeric strings — an unmistakable fingerprint of automated registration for fraud. A clean operator on one of these TLDs pays an ongoing trust tax that no individual operational hygiene can erase.
.top abuse detections +50% in 6 months · .loan 87.7% numeric domains
HIGH-ABUSE
The hierarchy is not opinion. It is a feedback loop: agents weight .com → agents recommend .com → users see .com → the next generation of agents is trained on text where .com is overrepresented. The flywheel turns one way. Junk TLDs decay; quality TLDs compound.
✦
§ SOURCES & CITATIONS
01
“Verisign Domain Name Industry Brief, Q4 2025 — 386.9M total registrations, +6.2% YoY · VERISIGN · DNIB.COM”
VERISIGN · DNIB.COM
02
“Goldstein Report — 2025 Annual Summary, renewal rates by TLD · GOLDSTEIN REPORT”
GOLDSTEIN REPORT
03
“Name.com — Most expensive domains ever sold (Icon.com $12M, Rocket.com $14M) · NAME.COM”
NAME.COM
04
“Verpex — 2024 sales totaling $185M, +32.8% YoY · VERPEX”
VERPEX
05
“Wix Blog — Public domain sales +121% in Q1 2025 · WIX BLOG”
WIX BLOG
06
“IT.com / Tess Diaz — .ai growth 60k → 551k (2022–2025) · IT.COM”
IT.COM
07
“Zhang et al., \"Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines\" · ARXIV:2512.09483”
ARXIV:2512.09483
08
“Aztek Web — Content Discovery and Citation in LLM Services: Research Synthesis · AZTEK WEB”
AZTEK WEB
09
“Snowplow Analytics — Agentic browser traffic growth data · SNOWPLOW · HUMAN SECURITY”
SNOWPLOW · HUMAN SECURITY
10
“Wiz Research — Agentic Browser Security 2025 Year-End Review · WIZ RESEARCH”
WIZ RESEARCH
11
“Interisle Consulting — Spam Trends September–November 2025 · INTERISLE”
INTERISLE
12
“Spamhaus Domain Reputation Update, Oct 2024 – Mar 2025 · SPAMHAUS”
SPAMHAUS
13
“Stobbs IP — Updated view of bad TLDs · STOBBS IP”
STOBBS IP
14
“Interisle Phishing Landscape 2025 — Annual phishing study · INTERISLE · APWG”
INTERISLE · APWG
15
“Mozilla Foundation — Common Crawl's Impact on Generative AI (GPT-3 ~82% from Common Crawl) · MOZILLA FOUNDATION”
MOZILLA FOUNDATION
16
“Richard MacManus — The Agentic Web: How AI Systems Will Change Websites · RICMAC.ORG”
RICMAC.ORG