Everything so far in this book assumed one implicit asker: no particular place, no particular language. This chapter drops that assumption. When the answer depends on where the asker is or what language they ask in, the sources feeding the answer change, and so does the work. Local and international visibility share this chapter because they share that lesson: context changes the sources, and the sources decide what you optimize.
Local prompts run on retrieval
"Best plumber near me" cannot be answered from model memory in any useful way. Training data thins out fast below the national level: a local service business generates a fraction of the web text a consumer brand does, and the model's memory of a neighborhood's businesses is sparse, stale, or empty. So assistants answering local and service prompts lean almost entirely on what they can fetch: the maps stacks and their business listings, review corpora, and directory pages that rank for the query. OppAlerts' cited-versus-recommended analysis points the same direction: local and service categories sit at the read-driven end of the spectrum, where recommended brands were fetched during the answer, while famous-brand consumer categories run on memory.From the cited-versus-recommended analysis in the OppAlerts study; the memory-versus-retrieval split it measures is explained in Model Memory and Live Retrieval, with method in The Research Behind This Guide.
That skew is good news operationally. Retrieval-side work pays in weeks, not model generations, so local visibility is one of the most responsive arenas in the book. The current, crawlable state of your listings and pages is nearly the whole game.
Local entity hygiene
The work itself is the least novel in this book, because the sources are the ones local SEO already serves. What changes is the consumer of the data: an answer engine composing one recommendation, rather than a map pack displaying three pins.
- Profiles: claim and complete the business profiles on Google and Bing, since their maps stacks feed both their own AI products and, through their indexes, others. Categories, services, hours, service area, photos: complete and current.
- Consistency: name, address, and phone consistent across profiles, directories, and your site. Conflicting facts in retrieved sources give the model a reason to drop you for a competitor whose facts agree, or worse, to state the wrong facts confidently.
- Reviews: volume, recency, and text. Review text is retrievable evidence about you, and "steady and recent" reads better to a system composing an answer today than a heap from three years ago. The community mechanics are the same ones in Reviews, Reddit, and Communities.
- Pages: a crawlable location page per location, with the entity facts in text and in LocalBusiness structured data, per Structured Data and Machine-Readable Facts.
Multi-location brands run the same checklist at fleet scale, where the failure mode is rot: hundreds of profiles drifting out of date. Assign ownership, audit quarterly.
Language is a training-data boundary
International visibility starts from an uncomfortable fact about the models: their training corpora are heavily weighted toward English. Providers do not publish exact language splits, but the weighting shows in behavior, and the practical consequence is that a model's memory of your brand can differ sharply by language. A brand well represented in English text may be barely represented in the German or Japanese slice of the corpus, so the memory answer that names you in English omits you in German. Visibility earned in one language does not automatically transfer.
Retrieval adds its own version of the boundary. A prompt in Spanish does not guarantee Spanish sources: systems sometimes retrieve English pages for non-English prompts and translate while composing, and sometimes retrieve locally. Treat the behavior as measurable rather than assumed: run your prompt panel per language and market, the same sampling discipline as Measuring AI Search Visibility, and see which sources the models actually read in each. Where English sources dominate answers in your non-English markets, your English visibility is doing double duty. Where local sources dominate, you need local coverage, local reviews, and local links in that market's language.
The technical layer carries over from international SEO intact. Publish real localized pages and use hreflang to tell Google which pages are variations of the same content for which locales; note that Google detects a page's language from the visible content itself, not from hreflang or lang attributes.Google, localized versions documentation: hreflang maps equivalent pages across locales, while language detection reads the content. Machine-translated boilerplate does not build visibility in a language; it thins the very corpus you need to be well represented in.
Market prioritization with a limited budget
Each market also runs its own assistant mix: which AI products dominate, and which search indexes those products draw on, differ by country, and the mix decides where your effort lands. Nobody can afford full-stack visibility work in every language. Prioritize with three questions per market. How much revenue does the market carry now, or credibly soon. What does the assistant mix there retrieve for your category, measured, not assumed. And how far does your existing English or home-market visibility already reach there, which your per-market sampling shows you directly.
Sequence the work like the rest of the book: retrieval-side fixes first in every market you enter, because they pay within weeks; memory-side investment, local press and community and links, only in the markets you commit to for years, because that is how long the memory clock runs. A market you cannot fund past the retrieval layer is a market where you compete on current pages alone, which is a defensible position, just a ceiling on one.