Quick Answer: Not every AI assistant searches the live web before answering. Some ground every response in real-time retrieval, while others rely largely on static training data and only occasionally incorporate live sources. Testing by Andre Alpar found this distinction changes the actual content of an answer rather than only whether it cites a source, since assistants that searched live provided correctly localized information that assistants relying on training data got wrong. For GEO strategy, this means content freshness and optimization only reach the platforms that are actually retrieving live content in the first place.
Ask an AI assistant a question in English from Germany, and depending on which one you asked, you might get an American emergency number handed to a German user, not because the assistant is broken, but because it answered from memory instead of checking the live web, and its training data defaults to the United States.
That is the finding at the center of a recent test by Andre Alpar, who ran the same 50 questions through 16 AI assistants in both German and English. The assistants that searched the live web found the correct German information even when asked in English. The ones answering purely from training data defaulted to American answers, regardless of who was actually asking.
Alpar's test asked the same 50 questions in German and English across 16 consumer AI assistants, then checked which ones returned correct, localized answers. One question involved a country-specific emergency number. Asked in German, all 16 assistants gave the correct German numbers. Asked in English, most defaulted to American numbers, even though the underlying question had not changed.
The split lined up cleanly with whether an assistant searched the live web before responding. The assistants that pulled in real sources during the German-language test kept doing so in English and found German results regardless of the language used to ask. The assistants that answered from memory carried their training data's default assumptions into the English-language responses, and that default was American. Grounding, in other words, did more than add a citation. It changed what the assistant actually said.
The emergency number example is a clean illustration, but the underlying mechanic applies directly to B2B buyer research. If an AI assistant answers a category question from static training data rather than a live search, then updating your content, refreshing your data, or publishing a new benchmark does nothing for that platform this quarter, since the assistant is not looking so much as recalling whatever it learned during training, which may be many months old.
That reframes a basic assumption behind most GEO advice. Freshness, schema updates, and new content only reach the AI platforms that retrieve from the live web in the first place. For any platform that still answers primarily from training data, a brand's most recent work is invisible until the next training cycle, whenever that happens.
Alpar's testing also surfaced something that complicates AI visibility tracking more broadly. Several consumer AI assistants are not a single, stable model behind the scenes. In his test, Perplexity's default setting routed questions to Claude rather than its own proprietary model, and other assistants would not confirm which underlying model was answering at all.
That matters for anyone measuring AI Share of Voice by platform name. A dashboard reporting visibility inside a named assistant may actually reflect visibility into whichever model that assistant happened to route the query to that day, which is a moving target rather than a fixed system to optimize against.
Do not assume live retrieval applies uniformly across the platforms your buyers use. Test directly, with your own category queries, on each platform you care about, and check whether the answer references anything published recently. An answer that only reflects older, well-established information is a signal that the platform may be responding from training data rather than a live search for that query type.
Run the same test in every language your buyers actually use. Alpar's finding held for a functional, factual question with a clear right answer, and the language-dependent default appeared consistently across the assistants tested. A B2B brand serving multiple markets should confirm that an assistant's answers hold up in the languages its actual buyers use, not only in English.
GEO advice tends to assume a single, consistent AI to write for, but the platforms your buyers actually use do not behave that way. Some are grounded in live search and will reward exactly the freshness and citation work most GEO guidance recommends. Others are answering from memory and will not notice any of it until their next training update.
Knowing which is which, for the specific platforms and languages your buyers use, is the starting point. If you want help mapping where your brand actually shows up across AI platforms and which ones are grounded versus static, the AEO Index tracks that by platform.