Ask the same AI assistant the same question from London and from Sydney, and there’s a decent chance you get two different answers. Sometimes it’s only the currency or the spelling, but quite often it’s the actual substance, like which product gets recommended, which rule applies or which price shows up. For anyone who publishes content, sells online or just wants to know whether a tool can be trusted, that gap is worth understanding.
Part of it comes from the model, and part of it comes from the web search the assistant runs before it answers. Turning search on fixes a lot of outdated answers, which is good, but search results themselves depend on where the request comes from. So the location of the person asking ends up shaping the reply in ways that are pretty hard to notice from one desk.
Where the differences come from
Most assistants with web access send a query to a search engine or their own index, read a handful of pages and then summarize them. Search engines rank pages differently by country and language, so a user in Germany and a user in Canada asking the same thing in English can easily get a different set of sources behind the answer. Many websites also switch prices, stock or whole sections based on the visitor’s IP address, which means the assistant may be reading a local version of a page that nobody elsewhere would ever see.
How to test answers from other locations
The only honest way to see these differences is to ask from the places you care about, and doing that from one office takes a bit of setup. Teams that run regular checks usually route their test browser or script through residential proxy servers in each target country, so the requests reach the assistant and the sites it reads from the same kind of home connection a local user would have. It’s worth keeping the session sticky for the length of one conversation, because switching addresses mid-chat can change the results halfway through and muddle the comparison.
Beyond the network, try to copy the rest of the local setup as closely as you can. The Accept-Language header tells sites which languages the user prefers, so a test that claims to come from Paris but asks for US English will get a strange mix. Time zone, account settings and whether the user is logged in also change answers more than people expect.
It also helps to fix the prompt wording before you start and keep it the same in every location, because even small edits like adding a polite opener or swapping a brand name can push the reply somewhere else. Write the questions the way a real customer in each market would ask them, and get them translated properly rather than with a quick machine pass, so the comparison stays about location and not about phrasing.
What to record
A single test tells you very little, since the same assistant can give two different replies to the same prompt a minute apart. Run each question a few times per location, save the full reply along with the sources it cited, and note the date, the model and whether web search was on. After a few rounds, the pattern usually becomes clear, like one country getting a newer price or a different brand recommendation.
Pay close attention to answers that sound sure of themselves but cite nothing. Those are the ones most likely to be hallucinations, and in smaller markets, where there’s simply less local text for the model to draw on, they probably turn up a bit more often.
Why it matters for publishers and brands
If your business sells in several countries, the assistant’s answer is slowly becoming a storefront you don’t control. Checking it in each market shows whether your local pages are being found, whether they carry the right signals for multi-regional sites and whether old prices are still floating around somewhere.
None of this needs a big budget, just a short list of questions, a few locations and the habit of checking every month or so. The answers will keep shifting as models and search indexes update, so the useful part is seeing how they change over time rather than any single result.