**TL;DR** — Across 21 client sites with traffic in more than one country, through May 2026, we audited a question that single-locale rank tracking is blind to: when you run the same query string from different countries, how much does the AI Overview citation set actually change, and what decides whether your page survives the border crossing? Across 4,820 query–country pairs (a 200-query basket run from up to eight country/language settings each) the citation set was not stable across borders. The median Jaccard overlap between a query's citation set in the US and the same query's citation set in a second English-speaking country was 0.38 — fewer than two of five cited domains were shared — and it fell to 0.11 once the second market used a different language. The strongest predictor of a page surviving a border crossing was local corroboration: a page cited in a market's top-10 organic results for the query was retained in that market's AI Overview 4.7× more often than a page that ranked organically only in its home market. The second was locale-explicit content — a page that named the country, used local currency/units, and carried a country-appropriate hreflang alternate was retained 2.9× more often than a globally-generic page. The third was ccTLD/subfolder structure: country-targeted URLs (a /uk/ path or a .de domain) were retained in their target market 2.2× more often than a single global URL serving every country. One change — building a genuine local-corroboration footprint (a localised page that earns local organic placement) for the markets that matter, rather than relying on one global page to be cited everywhere — lifted multi-country citation retention by 44% on the affected sites over a 45-day follow-up.
Why we ran this audit
Almost all of our AI Overview measurement through early 2026 ran from a single locale — usually a US English setting — because that is how rank trackers are configured by default and because it is cheap. But several clients with real revenue in two or three countries kept reporting a contradiction we could not explain from single-locale data: our dashboard showed them cited and healthy, and their team in-market showed them absent from the same answer card. Both were right. The dashboard was reading one country and the in-market team was reading another, and the AI Overview was citing a different set of pages in each. We had been treating citation as a property of a query when it is actually a property of a query-in-a-place, and that gap was invisible in every report we produced.
The second motivation was strategic. If the citation set re-shuffles at every border, then a global content strategy — one canonical page, served worldwide, expected to be cited everywhere — is structurally fragile in exactly the markets a client is trying to expand into, because the page has no local corroboration to anchor it there. We needed to know how big the border effect is, which structural choices survive it, and whether the fix is more localised pages or merely better hreflang on the pages that already exist.
How we ran the measurement
21 client sites — 8 SaaS, 6 DTC, 4 publisher, 3 B2B services — each with verified organic traffic in at least two countries. For each site we built a fixed 200-query basket of its real in-market queries and ran every query from up to eight country/language settings (combinations of gl and hl, plus a clean IP in-region where we could provision one) covering the client's actual markets: typically US, UK, Canada, Australia, plus one to three non-English markets (Germany, France, Japan, or Singapore depending on the client). We captured each query's AI Overview answer card and its full citation set, twice daily, through May 2026. For each query we computed the citation-set overlap between every pair of markets (Jaccard over cited domains), and for each client page we recorded whether it was cited in each market, its local organic rank in that market's top-20, whether it carried locale-explicit signals (country name, local currency/units, hreflang alternate for that market), and its URL targeting structure (global URL, subfolder, subdomain, or ccTLD). The full cohort was 4,820 query–country pairs; the cross-market comparison set was 14,300 market-pair observations.
Two normalisation moves matter. We excluded queries that returned no AI Overview in one or more markets — coverage itself varies by country, and a query with a card in the US but no card in Japan is an availability difference, not a citation-variance one; those were 22% of query–country pairs and are a separate question. We also separated language borders from country borders: a US-vs-UK comparison is a same-language country border, while a US-vs-Germany comparison is both a country and a language border, and lumping them would have blamed geography for what is partly a translation effect. The 0.38 same-language and 0.11 cross-language overlap figures are reported separately for that reason.
The shape of the border effect
The flat headline first. Citation sets are not stable across borders, and the instability is large. For same-language country pairs (US/UK/CA/AU among each other) the median citation-set Jaccard overlap was 0.38 — for a typical query, fewer than two of every five cited domains were shared between two English-speaking markets, and the rest were market-specific. For cross-language pairs the median overlap collapsed to 0.11, which is close to "different answer entirely." The reader in each country sees one confident answer card and has no way to know it was assembled from a largely different set of sources than the same query would surface one border over.
The most decision-relevant finding was the asymmetry of who survives the crossing. Big global institutional domains (major reference sites, large international publishers) were retained across borders at high rates — their median cross-border retention was 71% — so the shared 38% is disproportionately made of a few global authorities, and the market-specific 62% is where every commercial site actually competes. For a normal client page the question is never "am I as stable as Wikipedia"; it is "do I survive into the specific second market I sell in," and there the base rate without local signals was low: a page cited in its home market was retained in a second same-language market only 34% of the time by default.
Driver one: local corroboration decides most border crossings
The dominant predictor of whether a page kept its citation across a border was whether the page ranked organically in that market. A page that sat in a market's top-10 organic results for the query was retained in that market's AI Overview 4.7× more often than a page that ranked organically only in its home market, and the effect was close to monotonic with local rank: top-3 local organic retained at 81%, top-10 at 63%, top-20 at 41%, outside top-20 at 14%. The composer behaves, across borders, as though it re-derives the candidate pool from each market's local index rather than carrying a global citation set from country to country — so a page with no local organic presence has almost nothing to be cited from in that market, regardless of how strongly it is cited at home.
We ran a structural test on 18 pages across 9 clients, each cited reliably in its home market but absent in a target second market where the client had real sales. We could not manufacture local rank overnight, so we did the durable version: for each page we created or strengthened a genuinely localised counterpart for the target market — local examples, local pricing, local regulatory framing, an in-market case reference — and built the internal and external links to earn it local organic placement. Over the 45 days that followed, 11 of the 18 localised pages entered the target market's top-20 organic, and of those 11, eight began appearing in that market's AI Overview citation set on at least one target query. The lever for crossing a border is earning local corroboration, not asking one global page to be cited in a market where it does not rank.
Driver two: locale-explicit content survives better than generic content
When local organic rank was comparable, the deciding factor was how explicitly the page declared its locale. A page that named the target country in its answer-bearing content, used the local currency and units, and carried a correct hreflang alternate for that market was retained across the border at 2.9× the rate of a page with identical organic strength but globally-generic phrasing ("$" with no currency named, no country reference, one hreflang-less URL for all markets). The composer appears to read locale-explicit signals as evidence that the page is genuinely about the market the query came from, and to prefer it for that market's answer over a page that is merely relevant in the abstract. Generic global phrasing reads as no-particular-market, and no-particular-market loses to a page that is unmistakably about the asker's country.
We ran a structural test on 15 pages across 8 clients. Each ranked locally but used globally-generic phrasing. We localised the answer-bearing content — named the country, switched to local currency and units, and added the missing hreflang alternate — while leaving the page's core argument intact. Over the 60 days after the change, 10 of the 15 pages improved their retention in the target market, and on 4 of those the page moved from intermittently cited to consistently cited across the twice-daily captures. The two clearest gains were on pages where a single bare currency symbol had been ambiguous across markets; naming the currency and country removed the ambiguity and the page started being treated as that market's answer rather than a generic one.
Driver three: country-targeted URL structure
The third driver was URL targeting structure. Country-targeted URLs — a /uk/ or /de/ subfolder, a country subdomain, or a ccTLD — were retained in their target market 2.2× more often than a single global URL expected to serve every country, holding organic rank and content roughly constant. The effect was strongest where it compounded with hreflang: a country subfolder with a correct reciprocal hreflang cluster was retained at 2.6×, while the same subfolder with broken or missing hreflang fell back toward the generic-URL base rate. The structure appears to help the composer disambiguate which URL on a domain is the right answer for a given market — the same disambiguation problem that decides subpage-vs-homepage citation within one market, now playing out across markets.
The implication is about disambiguation discipline, not about spinning up dozens of thin country pages. We tested it on 12 page-clusters across 7 clients by giving each a clean country-targeted URL structure with correct reciprocal hreflang for the markets that actually mattered — not every country, only the two or three with real revenue — and consolidating duplicate near-identical global pages into the targeted structure. Over the following 45 days, 8 of the 12 clusters improved cross-market retention, and the failures were instructive: two clusters got worse because the new country pages were too thin to earn local organic rank, which sent us straight back to driver one. Structure helps the composer pick the right URL only once a URL is good enough to be a candidate at all.
What changed in our content checklist
Three changes. We replaced single-locale rank tracking with multi-market capture for every client that sells in more than one country: the AI Overview citation report now runs from each of the client's real markets and reports retention per market, because a single-locale dashboard was systematically lying to multi-country clients about where they were cited. We added a "local corroboration" gate to international expansion work: before expecting a page to be cited in a new market, we check whether it ranks organically in that market, and if it does not, the first deliverable is a localised page built to earn that rank rather than an hreflang tag bolted onto the global page. And we made locale-explicit content a requirement rather than a nicety on any page targeting a specific market — named country, local currency and units, correct reciprocal hreflang — because generic phrasing was costing retention even on pages that ranked.
We dropped one assumption. Through 2025 we had treated hreflang as the primary lever for international AI visibility — get the hreflang cluster right and the correct page would be served everywhere. The audit shows hreflang is necessary but nowhere near sufficient: it disambiguates among URLs the composer is already considering, but it cannot put a page into a market's candidate pool if the page has no local organic presence there. So hreflang dropped from "the international fix" to "the second step after local corroboration," and our international briefs now lead with earning local rank and treat hreflang as the disambiguation layer on top of it.
- 01Measure citation per market, not per query. Same-language citation-set overlap was a median 0.38 and cross-language 0.11 — a single-locale dashboard systematically misreports where a multi-country client is actually cited.
- 02Earn local corroboration before expecting a border crossing. Local top-10 organic raised cross-border retention 4.7×; 8 of 18 audited pages entered a new market's citation set after a localised page earned local rank.
- 03Make the page locale-explicit. Named country, local currency/units, and correct hreflang retained pages across borders 2.9×; 10 of 15 audited pages improved after localising the answer-bearing content.
- 04Use country-targeted URLs with reciprocal hreflang — for the markets that matter. Country-targeted structure retained at 2.2× (2.6× with correct hreflang), but only once each URL is strong enough to rank locally at all.
Where this argument breaks
For genuinely global, language-neutral facts (a scientific constant, a historical date) the border effect is small — those answers are dominated by global institutional domains whose 71% cross-border retention swamps any local signal, and no localisation work meaningfully changes a citation set you were never going to be in anyway. For markets where the client has no realistic path to local organic rank, the honest answer is that AI Overview citation in that market is not yet a reachable goal and the budget belongs elsewhere; local corroboration is a prerequisite, not an optimisation. For very small or low-volume markets the AI Overview may not render at all, which is an availability question prior to the citation-variance one. For Chinese-language AI search (文心一言, 元宝, 通义) the relevant border is not country but index and language entirely — those engines draw from a different corpus, and the lever there is native Chinese-language pages with local hosting and ICP filing rather than hreflang on a global site, a problem our cross-border framework does not address. Our window was 60 days and the cohort was 21 sites across their specific markets; the per-market numbers are point estimates that will differ by vertical and by how contested each market is. Outside those carve-outs the lesson holds: in 2026 the AI Overview citation set is a property of the query-in-a-place, fewer than half of cited domains survive a same-language border and far fewer survive a language border, and a page crosses into a market by earning local corroboration and declaring its locale — not by being cited at home and hoping it travels.