Google Launches Connected Apps Platform for Gemini and Search

Google just opened Gemini and Search to any business willing to wire up a connected integration, billing it as SEO for the age of AI agents: build the connection or go invisible when an agent does the shopping for your customer.

The technical piece is straightforward. Google adopted MCP (the open connector standard that lets an AI agent call your service) and is cloud-hosting the whole thing, which removes the "run your own server" burden that slowed earlier developer platforms. Fifty-plus apps are live in US beta, with Search rolling out over the next several weeks. Adobe and Experian are among the early partners.

The real decision isn't technical. The integration code is cheap to write. What you're actually deciding is whether to accept Google as the layer between you and your customer again. Google has run this movie before: Actions on Google pulled a wave of developer investment and then quietly starved. Fifty apps in a US beta is a launch-day number. The "agentic web" framing only works if Google wins the front door race, which is wide open. ChatGPT already runs third-party connectors. Apple is pushing on-device actions. You may end up building the same connector three times.

The free console is the hook. Google is cloud-hosting the orchestration compute, the context packaging, the retry logic. At 50 apps that's a rounding error. At 50,000 apps running live agent sessions, that's a real cost center Google is paying for on purpose. The console also ships traffic, latency, and error telemetry, which is genuinely useful because most AI integrations today run blind. Observability is what developers can't easily build themselves, so it's what makes the platform sticky. Sticky is what you eventually charge for.

There's a safety issue nobody has priced. Routing live user tasks through outside connectors opens a prompt-injection risk most teams haven't thought through. If a connector returns crafted text, it can steer the agent mid-task, because the model reads a tool response as instructions. Experian being an early partner makes this concrete: an agent that can pull credit or identity data and then act on it is exactly where a hijacked response does real damage. Google's console monitors speed and error rates. It does not watch for an agent being talked into the wrong action. No published safety test framework exists for connected tools.

The call: Google introduces paid or usage-metered tiers on the Connected Apps Platform before it exits US beta and reaches general availability, no later than mid-2027. Medium confidence. The subsidize-then-charge pattern is consistent across Google's platform history, and the free console telemetry is the specific mechanism that creates switching costs before the meter turns on.

Build the integration if you already serve customers through Google and the cost is a sprint. Don't make it your only distribution channel. Don't route sensitive data through it until someone has red-teamed poisoned tool responses. And before you commit, hammer your auth flow with agent-style retries and run a four-step chained task against your own connector to find where the grounding breaks.

Google just opened Gemini and Search to any business willing to wire up a connected integration, and it's selling the move as SEO for the age of AI agents. Build the connection or go invisible when an agent does the shopping for your customer. The question for anyone who runs a service people reach through Google: is this a real distribution channel, or another Google platform that gets you to invest and then changes the rules?

How hard is this to undo? For Google, easy. For the developer who builds on it, hard in the way that matters. The integration code is cheap to walk away from. The dependency on Google as your front door is not. What's actually being decided here isn't "should I ship an MCP server" (the technical connector that lets Google's AI call your service). It's "do I accept Google as the layer between me and my customer again." No hard deadline. US beta, 50-plus apps, Search rolling out over weeks. The pressure is competitive, not a date.

The Skeptic. Google has done this exact movie before. Actions on Google, the assistant developer platform, got a wave of integrations and then quietly starved. Developers who built voice actions spent real cycles and got stranded. Fifty apps in a US beta is a launch-day number, and launch-day numbers are what PR teams produce. The "agentic web" framing only works if Google wins the race to be where people go to get things done, and that is wide open. ChatGPT already runs third-party connectors. Apple is pushing on-device actions. The integration surface is fragmenting, which means you may have to build the same connector three times for three front doors. That's not a standard winning. That's a tax.

The Safety Lens. Routing live user tasks through outside connectors opens a door most teams haven't thought about: poisoned tool responses. If a connector returns crafted text, it can steer the agent mid-task, because to the model a tool response reads like instructions. Google's console watches speed and error rates. It does not watch for an agent being talked into the wrong action. Experian being an early partner makes this concrete. An agent that can pull credit or identity data and then act on it is exactly where a hijacked response does damage. Google ships no published safety test framework for connected tools. Every app in the ecosystem borrows Google's trust, and nobody has priced what happens when one of them leaks it.

The Enterprise Buyer. The thing a CTO actually buys here is distribution, and Google is giving it away cheap to build density. The console with traffic, latency, and error data is genuinely useful, because most AI integrations today run blind. But look at who signed first. Adobe and Experian have infra teams and legal departments that can absorb the risk of building on a beta. A mid-market SaaS shop is signing up to maintain auth flows that agents will hammer in ways human sessions never did, with no contract, no indemnification, and no commitment that the pricing stays free. That's fine to pilot. It's not fine to make it your only channel. The question in procurement is always the same: what happens when they turn on the meter.

The Compute Pragmatist. Google cloud-hosting the connector is the real giveaway. They're absorbing the orchestration compute, the context packaging, the retry logic, that integrators used to self-host. At 50 apps that's a rounding error. At 50,000 apps running live agent sessions, that orchestration becomes a serious workload Google is paying for on purpose. You subsidize a network until switching costs are real, then you charge. The pattern is older than AI. Watch for tiered pricing on the console once the network is dense enough that leaving hurts. The beta is the acquisition cost. The business model comes later.

The Researcher. Google adopting MCP, the open connector standard, and hosting it for you is the part of this story that actually matters. Removing the "run your own server" burden is how you seed a standard fast, and 50 partners in beta shows it working. The open question nobody has answered: do these connectors hold up when an agent chains four or five steps together? Context degrades as the chain gets deeper, and structured data like Experian's credit records is where grounding failures show up first. The model doesn't quietly fail. It confidently returns the wrong field.

Where they split. The Researcher and the Enterprise Buyer both see MCP adoption as real and useful, but they draw opposite conclusions: one sees a standard finally coordinating, the other sees a free sample with the price tag hidden. That's the live tension. The Skeptic goes further and doubts Google wins the front door at all, which would make the whole integration a sunk cost. And the Safety Lens is pointing at a risk none of the others have priced: the free, convenient, cloud-hosted connector is exactly the thing that makes a poisoned tool response easy to deliver.

What it hinges on. Three beliefs. First, does Google actually become a front door people use to get tasks done, or does the traffic stay on ChatGPT and on-device assistants? Second, does the connector layer stay free, or does the meter come on once you can't leave? Third, do multi-step agent tasks through these connectors work reliably enough to trust with real data? Build the integration if you already serve customers through Google and the cost is a sprint. Do not make it your only distribution, and do not route sensitive data through it until someone has red-teamed poisoned responses. Before you commit, test your auth flow against agent-style retry hammering, and run a four-step chained task against your own connector to see where the grounding breaks.

Prediction: Google will introduce paid or usage-metered tiers on the Connected Apps Platform / Connected Console by the time it exits US beta and reaches general availability, which Google said is coming over the next several weeks to months, no later than 2027-06-30.

Confidence: Medium. The subsidize-then-charge pattern is Google's consistent playbook once a network has density.

Why: Google is cloud-hosting the connector compute itself, which it does not do for free at scale. At 50,000 apps running live agent sessions the orchestration load becomes a real cost center, and the free beta is the acquisition price. The "SEO for the agentic web" framing is the giveaway, because SEO's whole economy is pay-to-stay-visible once organic reach is normalized. The opposite outcome, Google keeping it permanently free, would only make sense if it never achieves the density that creates switching costs, and with Adobe and Experian already live and 50-plus apps in beta, density is exactly what it's building toward.

Revisit by 2027-06-30: We're right if Google announces any paid tier, usage-based fee, or metered pricing on the Connected Apps Platform or Connected Console. We're wrong if the platform reaches general availability with all integration and console features remaining free of charge.

One more thing. The reason this call is worth making over the easy one, which is "Google launches a platform, developers show up," is that the free console telemetry is the hook. Observability is the thing developers can't easily build themselves, so it's the thing that makes the platform sticky, and sticky is what you charge for.