Search a local product query and get an AI Overview that cites a page from another country entirely — one your users cannot buy from, in a currency they don’t use, sometimes in a store that won’t even ship to them. This is not a broken hreflang setup or sloppy international SEO. It is what happens when a system built to synthesize the most complete answer collides with a business that needs the most usable one.

Google’s AI Overviews mark a real architectural break from classic search. Traditional ranking-and-serving was built to return the single best regional URL for a query. Generative synthesis is built to assemble the most defensible explanation of a topic, pulling from whichever sources answer each piece of that topic most clearly — regardless of whether the user in front of the screen can actually act on what they’re reading. That mismatch produces a specific, recurring failure: geographic leakage, where an AI Overview cites out-of-market sources for a query that clearly has local or commercial intent.

Why this counts as good engineering

From an engineering standpoint, picking an international source is not a bug in the traditional sense — it is the system doing exactly what it was built to do: maximize factual coverage and minimize hallucination.

AI Overviews break a single prompt into several parallel sub-queries, each probing a different facet — definitions, mechanics, legal constraints, comparative attributes. The unit competing for inclusion is no longer a page or a domain; it’s a fact-chunk. If one source explains a sub-facet more explicitly or in a more extractable format, it gets pulled in as a grounding source for that piece of the answer, whether or not it’s the best overall result for the person asking.

Language stops being a boundary too. Modern models don’t translate a page as a separate step — they normalize content from any language into a shared semantic space and generate the answer from the underlying facts, not the visible snippet. An English summary sourced from a foreign-language page is simply cross-language retrieval working as intended.

The deeper issue is a structural one between two different logics. Classic Google Search is built around serving: once relevance is established, signals like IP location, language and hreflang act as directives that determine which regional URL a user actually sees. Generative systems are built around retrieval and grounding, where those same signals become secondary hints inside a retrieval-augmented pipeline — and get overridden the moment a higher-confidence semantic match turns up elsewhere. Once a specific URL has been selected as the source of truth for a fact, geographic logic downstream has very little power to change that.

The vector identity problem

At the root of this sits what’s best described as a vector identity problem. Content gets represented as numerical vectors encoding meaning, and two pages with substantively identical content — even ones serving completely different markets — often collapse into the same or a near-identical vector. To the model, they read as interchangeable expressions of the same underlying concept.

Market-specific realities like shipping eligibility, currency, or checkout availability aren’t semantic properties of the text itself — they live in the URL’s metadata, which the model isn’t weighing when it selects a grounding source from a pool of high-confidence matches. Whichever regional version was crawled more recently, renders more cleanly, or phrases the concept slightly more explicitly can win that selection without any evaluation of whether it’s commercially usable for the actual searcher.

Freshness makes this worse, because retrieval-augmented systems often treat recency as a rough proxy for accuracy. Once markets are already normalized into the same semantic space, even a minor content update to one regional page — a clarifying sentence, a slightly more explicit phrasing — can be enough to tip selection in its favor. Freshness isn’t a neutral signal here; it’s a multiplier on whichever version already has semantic dominance.

Ambiguity makes it worse, and hreflang can’t fix it

Query ambiguity is probably the least understood driver of this behavior. Classic search resolved ambiguity late, at the ranking and serving layer, using location, language, device and behavioral signals — users learned to trust that Google would infer intent and localize accordingly.

Generative retrieval handles ambiguity in the opposite direction. Instead of resolving intent early, ambiguity triggers semantic expansion: the system explores every plausible interpretation in parallel to maximize how complete the answer looks. That’s a deliberate design choice that reduces the risk of leaving something out — but it means the system becomes progressively more willing to violate geographic and commercial constraints that used to be enforced downstream. The question the system is answering shifts from “which result fits this user” to “which sources most completely cover the space of possible meanings.”

That’s also why a technically correct hreflang cluster does not guarantee regional preference. Hreflang was built for a post-retrieval substitution model — once a relevant page is found, the right regional variant gets served. AI Overviews resolve relevance upstream, during fan-out and semantic retrieval, before hreflang ever gets a say. If an international page provides the clearest “first answer” for a given sub-query, it gets pulled in as a grounding source immediately, and a localized equivalent is never considered unless it’s a technically superior answer for that same semantic branch. Hreflang can influence which URL gets served to a user. It has essentially no influence over which URL gets retrieved as a source — and retrieval is where the real decision happens now. Our piece on why international SEO needs a global knowledge integrity strategy goes deeper into fixing this at the content layer.

The diversity mandate adds a second layer of leakage

AI Overviews are also explicitly designed to surface a broader spread of sources than a traditional top-10 result set, and it evaluates URLs as distinct sources rather than business entities. A country-specific subfolder gets treated as an independent candidate even when it’s the same brand and product as the primary URL. Once a brand’s main page is selected, the diversity filter may go looking for a second, different URL to populate additional citation cards — creating a kind of ghost diversity, where the answer looks like it’s drawing on multiple perspectives while actually citing the same entity twice through different market endpoints.

Why this is a real commercial problem

None of the mechanics above are misconfiguration. They’re the predictable output of a system optimized for semantic completeness rather than commercial utility, and that gap has real business cost.

AI Overviews have no built-in concept of commercial harm — no check for whether a cited source can actually be purchased from, acted on, or legally used in the searcher’s market. When a user gets routed to an out-of-market destination, conversion probability collapses, and that failure is invisible to the system: there’s no corrective penalty because nothing in the evaluation loop measures it. The signals that used to protect against this — IP location, language, currency, hreflang — now function as weak hints inside generative synthesis, easily overridden by a higher-confidence semantic match found elsewhere in the fan-out process. And because AI Overviews occupy the most prominent position on the page as organic space keeps shrinking and zero-click behavior rises, whatever gets cited receives outsized attention — so a geographic mismatch here costs more than the same mismatch would have in a ten-blue-links world.

What a generative-era technical audit needs to check

Adapting to this means moving past standard visibility work into what amounts to generative engine optimization, with three specific priorities:

• Semantic parity across markets. Keep fact-chunks aligned at a near-identical level across regional pages — small asymmetries in explanation or specificity are enough to create an unintended retrieval advantage for one market over another.
• Retrieval-aware structuring. Break content into atomic, clearly extractable blocks that map to the sub-queries a fan-out process is likely to generate, rather than writing for a single linear read.
• Explicit machine-readable utility signals. Reinforce market validity and availability in a form the model can actually pick up, since it does not reliably infer commercial constraints from prose alone.

That kind of structured signaling connects directly to the entity-level work covered in our overview of giving AI systems a structured view of your business, and to the broader technical SEO audit built for the AI search era.

Frequently asked questions

What is geographic leakage in AI Overviews?

It’s when Google’s AI Overviews cite international or out-of-market sources for a query that has clear local or commercial intent, sending users to pages they cannot actually purchase from or use in their region.

Can fixing hreflang solve geographic leakage?

Not on its own. Hreflang governs which regional URL gets served once relevance is established, but AI Overviews resolve relevance earlier, during semantic retrieval and fan-out, where hreflang has little influence.

Is geographic leakage a sign of poor international SEO?

No. It’s a structural outcome of how generative retrieval resolves ambiguity through semantic completeness rather than commercial narrowing, even on sites with correct localization and geo-targeting in place.

The bottom line

Geographic leakage isn’t a quality regression — it’s what happens naturally when search shifts from routing transactions to synthesizing information. The engineering is working exactly as intended: ambiguity triggers expansion, completeness gets prioritized, and semantic confidence wins the citation. The business exposure is just as real: the system cannot tell the difference between a factually correct source and one a given user can actually do something with. Until generative systems build in a stronger notion of market validity, the practical response is defensive — visibility now depends not just on being the most complete version of the truth, but on making sure that version is also the one your specific market can use.