Google's Gemini API combines Google Search and Google Maps in one agent request

Google has quietly rolled out a feature called Grounding with Google Maps for the Gemini API, letting developers plug live location data straight into their AI apps. Instead of Gemini guessing at nearby restaurants or hallucinating opening hours, the model can now query Google's Maps database — reportedly covering more than 250 million places worldwide — and cite its sources with real URLs and place names.

The mechanics are straightforward. A developer sends a query like "best Italian restaurants within a 15-minute walk from here," optionally attaches latitude and longitude coordinates, and the model decides whether the question has geographic intent. If it does, it fires off the Maps tool, pulls in reviews, addresses, and photos, then weaves that into a normal text response with inline citations pointing back to Google Maps. It's the same trick Gemini already does with Google Search grounding, just pointed at a different dataset — and Google says the two can be combined in a single request.

Google is fairly specific about when this should and shouldn't fire. The tool is off by default, and Google's documentation nudges developers to only flip it on when a query has obvious local intent — "near me," a named neighborhood, that sort of thing — rather than leaving it running constantly, mostly for cost and performance reasons. Pricing also splits along model generation lines: Gemini 3-series models get billed per individual search query the model runs internally, while Gemini 2.5 and older models get billed per prompt that successfully returns a result. That's a meaningful difference for anyone building at scale, since a single Gemini 3 request could quietly trigger several billable searches under the hood.

There are real strings attached, too. Google requires apps using this feature to display Maps attribution exactly as specified — no reformatting "Google Maps" into another language, no splitting it across lines, no letting a browser auto-translate it — and sources have to appear immediately next to the content they support, visible in a single interaction. English-only prompts for now, and availability varies by region. Still, for anyone building a trip planner, a food delivery assistant, or a local-guide chatbot, this closes a gap that pure LLM knowledge never handled well: knowing what's actually open, nearby, and any good, right now.