If your company sells in Germany, Japan or Brazil, the working assumption is that an AI answer engine reaches into that market's web to answer a buyer's question. It does not. Across 15,883 observed answer runs and 22,213 cited domains, domains registered under a national country code take 2.73% of citation volume — and three of the 1,462 of them have accumulated enough evidence to carry a published rate. All three are .us.
Meanwhile two country codes that are sold worldwide as brandable strings — .ai, delegated to the Government of Anguilla, and .io, delegated in the British Indian Ocean Territory — take 8.56%, which is 3.13 times every national domain on Earth combined.
The instrument cannot see a market it was not pointed at, and that fence is stated in full at the end. But inside a fixed basket of English-language commercial questions, the answer layer is not a multinational layer. It is a .com layer with a .ai fringe.
What was measured #
Every figure below comes from one artifact: the public Machine Relations Index release mri_score_v2.0+2026-09-19+0cad03121f60, 59,716,928 bytes, SHA-256 0cad03121f60591aac78475994512dda09e70a5463f37d430d83c3230fe887d7, matching its own release manifest. The observation window runs 2026-05-10 to 2026-09-19, 126 days. Six engines were healthy for the whole window: Perplexity, ChatGPT, Gemini, Claude, Google AI Mode and Google AI Overviews.
Two denominators do the work, and they are not the same number.
- Domains: 22,213 distinct domains were cited at least once.
- Citation pairs: 111,510. This is the sum, across all domains, of the answer runs in which that domain was cited. One domain cited in one run is one pair. It is the volume denominator, and it is the honest one for a share question, because counting domains alone treats a site cited once the same as
reddit.comcited 2,012 times.
Every domain is classified by the last label of its name, which is what a resolver and a registry both use:
| Class | Domains | Share of domains | Citation pairs | Share of volume |
|---|---|---|---|---|
Generic (.com, .org, .net, .dev, …) |
17,876 | 80.48% | 94,971 | 85.17% |
| Two-letter codes sold as brandable strings | 2,611 | 11.75% | 11,651 | 10.45% |
| National country codes | 1,462 | 6.58% | 3,046 | 2.73% |
US-restricted (.gov, .edu, .mil) |
264 | 1.19% | 1,842 | 1.65% |
The second and third rows are the same kind of object under RFC 1591 and ICANN's own definition: every two-letter top-level domain is a country-code top-level domain, derived from the ISO 3166-1 list. Splitting them is an editorial judgment, and the strict reading is given alongside the split one in the method note. It is made here because .ai and .io carry no national meaning in practice, and merging them would hide the finding rather than report it.
Three of 1,462 #
The Index publishes a citation rate for a domain only once it clears an evidence floor: at least 10 observations across at least 7 distinct run dates. 503 of the 22,213 cited domains clear it and carry a rank.
| Class | Graded | Of | Rate |
|---|---|---|---|
| US-restricted | 7 | 264 | 2.65% |
| Generic | 440 | 17,876 | 2.46% |
| Brandable two-letter | 53 | 2,611 | 2.03% |
| National country code | 3 | 1,462 | 0.21% |
The three are market.us at rank 226 (cited in 49 of 15,883 runs, 0.31%, grade C, five engines, 30 distinct days), bark.us at 333 (39 runs, 0.25%) and canopy.us at 482 (30 runs, 0.19%). All three are American businesses using .us as a short brandable string, which means the count of non-US national domains that clear the floor is zero — out of 1,410.
Ninety-one national country codes appear in the cited set at all. Here is the whole leaderboard that matters:
| Code | Territory | Domains | Citation pairs | Share of volume |
|---|---|---|---|---|
.uk |
United Kingdom | 337 | 669 | 0.600% |
.au |
Australia | 177 | 366 | 0.328% |
.us |
United States | 52 | 254 | 0.228% |
.in |
India | 136 | 245 | 0.220% |
.ca |
Canada | 98 | 234 | 0.210% |
.eu |
European Union | 59 | 161 | 0.144% |
.de |
Germany | 96 | 148 | 0.133% |
.ch |
Switzerland | 28 | 68 | 0.061% |
.fr |
France | 29 | 55 | 0.049% |
.nl |
Netherlands | 27 | 52 | 0.047% |
.ie |
Ireland | 19 | 46 | 0.041% |
.sg |
Singapore | 24 | 42 | 0.038% |
.nz |
New Zealand | 27 | 39 | 0.035% |
.jp |
Japan | 22 | 34 | 0.030% |
.it |
Italy | 22 | 32 | 0.029% |
Germany's DENIC administers one of the largest ccTLD registries in the world. Ninety-six .de domains were cited across 126 days, for 148 citation pairs out of 111,510. Japan contributed 22 domains and 34 pairs.
The single most-cited national domain in the release is not a publisher. It is europa.eu, cited in 29 runs across 18 days by all six engines — and still marked collecting, because the strata it appears in have not accumulated enough observations to publish a rate. Behind it: nhs.uk at 26 runs on two engines, herba-medica.si at 24, point-up.co.uk at 24, cdhf.ca at 18, sqmagazine.co.uk at 17, ncsc.gov.uk at 13, and independent.co.uk — a national newspaper with a large newsroom — at 12 runs on two engines.
For scale, the top of the same release: reddit.com 12.67%, youtube.com 9.02%, linkedin.com 6.79%, medium.com 5.08%, forbes.com 4.14%, nih.gov 3.07%.
Anguilla outranks every nation #
.ai was delegated in 1995 and its IANA record names the ccTLD manager as the Government of Anguilla, a Caribbean territory of about 16,000 people. .io was delegated in 1997 and its record still gives an address on Diego Garcia, British Indian Ocean Territory. Neither string signals anything about where a company is or who it sells to.
In this release:
.ai— 1,178 domains, 5,210 citation pairs, 4.67% of volume.io— 954 domains, 4,337 pairs, 3.89%.co(delegated to Colombia) — 340 domains, 1,544 pairs, 1.39%- All 1,462 national domains, 91 territories — 3,046 pairs, 2.73%
Three brandable strings account for 2,472 of the 4,073 two-letter domains in the set and 11,091 of their 14,697 citation pairs, 75.5%. The highest-ranked two-letter domain of any kind is dev.to — Tonga's country code, managed from an address in Burlingame, California — at rank 22 overall, a 1.32% citation rate and a grade of B. No national domain ranks within 200 places of it.
This is not an accident of the measurement. It is what happens when the source layer of an answer engine is assembled from the global commercial web rather than from any country's web, and the global commercial web buys its domains from whichever registry has short names left.
The news shape is the least international #
The obvious hypothesis is that news questions pull in wire services and national papers while buying questions stay on .com review sites. The measurement says the reverse, and the gap is not small.
| Question shape | Citation pairs | National-ccTLD pairs | National share |
|---|---|---|---|
problem_first |
8,856 | 403 | 4.55% |
how_choose |
11,242 | 442 | 3.93% |
top_list |
12,774 | 459 | 3.59% |
best_x |
12,713 | 416 | 3.27% |
x_vs_y |
9,915 | 294 | 2.97% |
is_x_worth |
10,794 | 303 | 2.81% |
news_topic |
45,216 | 729 | 1.61% |
A problem-shaped buying question ("our warehouse pick rates are dropping, what do we do") is 2.8 times more likely to reach a national domain than a news question about the same industry. The news shape is the largest single bucket in the release — 45,216 of 111,510 pairs — and the most concentrated on the generic layer.
That is the second time this week the Index has separated the news answer from the buying answer along an axis nobody expected. The first was the publication identity axis: the outlets that hold an industry's news source list are largely not the outlets that hold its buyer source list. This is the same divergence measured by country of registration, running in the opposite direction from intuition. Anyone planning international earned media on the assumption that news coverage is the globally-distributed surface and product content is the local one has it backwards on both counts.
Consumer categories reach further than B2B categories #
National-domain share by subject category splits the release cleanly in two.
| Category | Citation pairs | National share | Brandable two-letter share |
|---|---|---|---|
| Emergent prosumer | 3,229 | 7.15% | 4.30% |
| Consumer health | 6,062 | 6.91% | 1.68% |
| Deep tech and hardware | 4,112 | 6.40% | 4.38% |
| Consumer products | 5,757 | 6.27% | 1.23% |
| Family software | 3,424 | 6.07% | 3.59% |
| iGaming and betting | 1,050 | 4.38% | 11.05% |
| AI visibility and GEO | 5,939 | 2.44% | 26.17% |
| AI security and privacy | 5,358 | 2.33% | 24.47% |
| AI infrastructure | 4,300 | 2.07% | 25.23% |
| Cybersecurity | 9,573 | 1.93% | 10.32% |
| Enterprise software | 8,206 | 1.80% | 8.75% |
| Fintech | 8,375 | 1.64% | 10.63% |
| HR and talent | 7,047 | 1.63% | 11.34% |
| Martech and advertising | 4,602 | 1.52% | 12.30% |
| Healthcare services | 8,131 | 1.24% | 6.67% |
| Consumer finance | 5,430 | 1.18% | 0.81% |
Consumer-health questions reach a national domain 5.9 times as often as consumer-finance questions. The pattern is legible: where the answer depends on a regulator, a health service or a physical product sold in a place — nhs.uk, europa.eu, cdhf.ca, vic.gov.au — the engines reach into national infrastructure. Where the answer is about buying B2B software, they do not, and the space that national domains would occupy is taken by .ai and .io vendor sites instead. In AI visibility and GEO, brandable two-letter domains take 26.17% of citation volume against a national share of 2.44%, a ratio of nearly 11 to 1.
Why national domains do not accumulate #
The three supporting distributions all say the same thing: national domains are cited, but they are cited once, by one engine, on one day.
- Engines that observed the domain. 1,265 of 1,462 national domains (86.5%) were cited by exactly one of the six engines, against 14,985 of 22,213 overall (67.5%). Mean engines observing a national domain: 1.19, against 1.64 across the set.
- Run count. 900 of 1,462 (61.6%) were cited in exactly one run out of 15,883, against 46.7% across the set.
- Persistence. Mean distinct days cited: 1.72 out of 126, against 2.91 across the set.
That is why the grade line reads three. The evidence floor asks for 10 observations across 7 distinct days. A domain cited once, by one engine, cannot clear it no matter how good the page is — and the failure mode is not quality, it is accumulation. A national domain that does get cited tends to be an institution rather than a publisher, which matches the source-role mix: 90.4% of national domains fall into the Index's uncategorized_source role against 84.0% across the set, and only 4.9% are classified editorial media against 5.5% overall.
What this changes for an operator #
Three consequences, in the order they cost money.
A national domain is not a citation strategy on its own. If the goal is to be cited in AI answers for an English-language commercial question, moving content to example.de does not help and measurably hurts: the entire .de zone contributed 148 of 111,510 citation pairs. Google's own guidance treats a ccTLD as the strongest signal of country targeting, which is exactly the property that makes it invisible to an engine answering a question with no country in it. The correct read is that country targeting and answer-engine citation are pulling in opposite directions, and a brand that needs both needs localized versions of pages that also live on a domain the global layer already cites.
Your international PR list and your AI citation list are different lists. independent.co.uk at 12 runs on two engines is the shape of the problem: a real newsroom with real reach, sitting below the Index's evidence floor for every segment it appears in. A placement there is worth what it is worth for human readers. It is not, on this evidence, a route into the answer layer.
Category matters more than country. If you sell consumer health or physical products, national infrastructure is already in the answer and worth pursuing. If you sell B2B software, the measurable path runs through the generic layer — the aggregators, the community platforms, the review databases — which is where the shape of the question, not the category, picks the winner.
Methodology and sources #
The methodology is a single-pass classification of one public release; every figure in this report is computed from it and from nothing else.
The dataset is the public Machine Relations Index release mri_score_v2.0+2026-09-19+0cad03121f60, read over plain HTTP on 2026-09-19 at 12:39 UTC and verified against the manifest by SHA-256. It holds 22,213 cited domains, 15,883 answer runs, 125,115 source events, 912 eligible queries, 157 strata (88 published, 63 collecting) and a window of 2026-05-10 to 2026-09-19.
Classification is by the final label of each domain name. .gov, .edu and .mil are counted separately because registration in all three is restricted to United States entities. Two-letter labels are split into "national" and "brandable" using a fixed list of twenty codes that registries market worldwide without a local-presence requirement: io, ai, co, me, tv, cc, ly, to, fm, gg, sh, ws, im, st, vc, so, gl, la, nu, pw. The strict reading, with no split at all: every two-letter domain is a ccTLD, which makes it 4,073 domains, 18.34% of the set, and 14,697 citation pairs, 13.18% of volume — of which .ai, .io and .co alone are 75.5%. Both readings are reported so the judgment is inspectable rather than embedded.
Delegation facts are taken from the IANA root zone database, which is the authoritative record: .ai (Government of Anguilla, registered 1995-02-16), .io (Internet Computer Bureau Limited, British Indian Ocean Territory, registered 1997-09-16), .to (Kingdom of Tonga, registered 1995-12-18), .us, .uk, .de, .eu, .ca, .au and .in. The ccTLD concept and its derivation from ISO 3166-1 are defined in RFC 1591 and by ICANN. European ccTLD registry volumes are published by CENTR and, for Germany specifically, by DENIC. Global registration totals across all TLD classes are tracked in Verisign's Domain Name Industry Brief.
The fence, stated plainly. The release publishes no locale, language or region field, and the Index observes a fixed basket of English-language commercial questions. This measurement therefore answers one question exactly: when an answer engine responds to an English commercial question with no country in it, where are the sources registered? It does not measure what a German-language query returns in Germany, and nothing here should be read as a claim about that. The distinction between language and region is a real one and this instrument only sees one side of it. A locale-varying basket is the obvious next measurement, and it is not this one.
Per-domain evidence for every figure above is public at the Index domain profiles and the per-category leaderboards.