America.gov puts Gemini and Grok in front of federal services

The US government replaced its front door with a chatbot. America.gov launched Tuesday, and the model answering questions about passports and Medicare is Google's Gemini and xAI's Grok — not a system the government built. Whatever you think of the politics, the AI answer layer just got its most public deployment, and the pattern is the one your customers are already applying to your business.

What actually happened

The administration launched America.gov on September 29 as a single AI entry point to federal services, led by US Chief Design Officer Joe Gebbia, the Airbnb co-founder. Gebbia said the site runs on Gemini and Grok. President Trump framed the target as the more than 10,000 forms scattered across federal agencies.

Version one is retrieval: ask a question, get an answer drawn from federal agency sites, including job listings. The stated roadmap is transactional — renewing a passport, changing a name after marriage, picking a Medicare plan, checking student loan balances, reserving a campsite. Elon Musk, Jensen Huang, JD Vance, Marco Rubio and Mehmet Oz attended the launch.

Two structural details are worth more than the headline. First, the government did not pick one model — it wired in two, from competing vendors. Second, it did not rebuild the 10,000 forms. It put a retrieval layer over content that already existed.

Why the answer layer matters for your business

Your customers stopped landing on your homepage a while ago. They ask a model, the model reads whatever it can parse about you, and it returns an answer you never wrote. America.gov is that dynamic with a flag on it.

Three things to take from it:

Your content is now an API whether you built one or not. The government did not rewrite 10,000 forms into a chat product. It made them retrievable. Do the same: a clean sitemap, real HTML headings, Product and LocalBusiness schema, prices and hours in markup instead of an image, an FAQ page written as question-and-answer rather than marketing prose. A model cannot cite a carousel.

Test what the models say about you. Ask Gemini, ChatGPT, Claude and Grok your five highest-intent customer questions — hours, service area, pricing, lead time, return policy. Write down the wrong answers. Every wrong answer traces back to a page you control, a stale directory listing, or nothing at all. That list is your next month of content work, and it is more valuable than another blog post about industry trends.

Two vendors, not one. The most quietly instructive choice here is the hedge. If the federal government will not route every query through a single model provider, neither should your product. Keep the prompt, the retrieval and the evaluation set in your own repo, and make the model a swappable line in a config file. We build every agent this way, because the pricing and the policy on the other end of that API change without asking you.

The thing being centralized is not the software. It is the point of contact. When a model is the first thing between a person and a service, the winner is whoever made their content easiest to retrieve correctly — not whoever spent the most on the interface.

Sources: CNBC, WGME/Sinclair.