Ask ChatGPT or Google's AI for the "best fertility clinic" without naming a city, and you'll often get an answer tailored to your city anyway. The engine inferred where you are and quietly localized the recommendation. The way Google Search has always done. But how it knows, how precisely, and which engines do it are widely misunderstood. Getting this right is the foundation of local AI visibility.
The short version: most AI engines localize, but only through their web-search layer, only coarsely (your city/region from your IP), and not in the model itself.
How can an AI "know" my location at all?
Every AI engine has two layers, and only one of them can know where you are:
- The trained model. Frozen at a knowledge cutoff, a large language model has no awareness of the person asking. No location, full stop. With web search off, the answer is location-blind.
- The live search layer. When the engine runs a web search to ground its answer, it behaves like a search engine, and this is where your location enters. It reads your approximate location (usually from your IP) and rewrites your question: "best clinic near me" becomes "best clinic in [your city]" before it searches.
So the right question isn't "does the AI know my city?" It's "does this engine run a location-aware web search, and does that change the answer?" That reframing is the whole game.
Location enters through the search layer, never the model. Which means local AI visibility is won exactly where local search is won. In the listings, reviews, and citations an engine retrieves.
Which engines localize, and how much?
All of the consumer engines localize coarsely through search; the APIs do not unless you tell them to. As of mid-2026:
| Engine | Localizes implicitly? | How |
|---|---|---|
| Google AI Overviews / AI Mode | Yes | Inherits Google Search's location signals |
| ChatGPT (search mode) | Yes | IP → coarse city/region; rewrites "near me" into your city |
| Perplexity | Yes | IP-based by default, even logged-out |
| Gemini | Yes / partial | Google account + IP; saved Home/Work & device location with permission |
| Claude (web search) | Yes, coarse | IP city/region; opt-out available; no GPS collected |
| Any of the above via API | No, unless passed | OpenAI & Perplexity accept an approximate user_location; none by default |
Two caveats worth stating plainly. AI Mode is a separate ranking system from Google's local 3-pack: one 2026 study found 28.5% of businesses in the local pack were absent from AI Mode for the same query, so winning the map pack does not guarantee inclusion in the AI answer. And these figures come from studies run in the US, UK, Canada, and Australia; localization behavior in other markets, including India, is unverified and should be tested directly rather than assumed.
Why is it only "coarse"?
Because the default signal is your IP address, which resolves to a city or region, not a precise spot. Engines round further, and IP geolocation is imperfect (testers have been placed in a city 45 minutes from where they actually were). Precise, GPS-level location is opt-in, not default: ChatGPT added an explicit precise-location toggle in March 2026, and Anthropic states Claude collects no precise or GPS location at all. Signed-in and paid accounts tend to produce sharper local answers than logged-out sessions, because the account adds context beyond the raw IP, but the baseline everywhere is coarse.
Does this happen if I never mention a city?
Yes, that's the key point for benchmarking. Implicit localization (the engine infers your city from IP and localizes silently) is different from explicit localization (you type "in City X"). They don't always return the same answer, and the implicit path is the one most real users trigger without realizing it.
This is also why testing local AI visibility by hand is unreliable: you can't easily fake a city-level IP inside a consumer app, engines round your location, and the same prompt in the same place can return different answers run to run because these systems are non-deterministic. The dependable approach is to benchmark both: explicit-city prompts across a matrix of cities, sampled repeatedly, and, where possible, use the APIs' location parameter to set the city deterministically. (That methodology gets its own part in this series.)
What this means for your brand
If location lives in the retrieval layer, then local AI visibility is won in the local content ecosystem an engine retrieves from, not by "training the model." Studies of local AI answers find the majority of citations go to third-party directories and review sites, with the rest to the businesses themselves. So the levers are familiar: a complete, consistent presence across local listings and review platforms, structured data that states who and where you are, and earned placement in the "best [X] in [city]" lists engines lean on. For regulated, high-trust categories like healthcare, the bar on credibility and corroboration is higher still.
The takeaway: AI is not inventing a new local game. It's reading the existing one through a new surface. The brands that show up are the ones whose local footprint is clear, consistent, and trusted enough for an engine to repeat. Measuring that footprint city by city, across every engine, is exactly what Buffy Intel is built to do.