A shopper in India asking Google’s Gemini for a phone recommendation this week saw something that would have been unremarkable in an ordinary store and is new in an AI interface: a buy button. On a small set of Flipkart listings, Gemini and Google’s AI Mode now let a customer complete a purchase without leaving the assistant, a step that moves the company’s AI from finding products to closing sales.
The test covers phones, electronics, and accessories, and TechCrunch observed it on September 26. Google intends to widen the rollout before India’s festive shopping season in October, a period that produces a large share of the country’s annual retail spending, according to the report. Flipkart, the Walmart-owned marketplace on which the feature depends, is the country’s dominant e-commerce player.
For Google, the move is the logical end of a push to make its AI assistant transactional. The company has spent years building product discovery into search and, more recently, into Gemini, and it has argued that the next step for AI is to complete the purchase rather than merely describe it. A checkout that stays inside the AI interface is the difference between a catalog and a store.
The same day, a different Google team described a less flattering view of the AI economy. Google’s Threat Intelligence Group found that dark-web marketplaces are selling access to leading AI models at discounts of up to 97 percent, according to a Financial Times report on September 26, with offerings tied to Anthropic, Google, and OpenAI.
The two developments, a polished buy button on one side of the world and discounted model access on the dark web on the other, describe the state of a market that has matured faster than its controls. The same models that Google is packaging into a shopping assistant are, elsewhere, commodities traded in the open like stolen account credentials.
The dark-web sales follow a pattern security researchers have tracked for more than a year: API proxy servers and resellers repackage access to top models and sell it below the official list price, in some cases by hijacking accounts or aggregating token quotas. For buyers shut out of the major providers or unwilling to pay full price, the gray market offers a way in.
Google’s researchers framed the finding as a threat-intelligence alert, a warning that stolen or resold access can be used to abuse the models, to launder the cost of adversarial use, or to reach capabilities a provider would otherwise restrict. The discounts, as steep as 97 percent, signal how much surplus access is sloshing around the system.
Analysts said the two stories, read together, show Google trying to do two hard things at once: build a consumer business around AI while policing a market in which its own models and its rivals’ are being moved through back channels. The company’s shopping ambitions depend on trust, and the dark-web finding shows how quickly that trust can be undermined from below.
The India rollout is itself a careful, limited experiment. Google has not said how many users see the buy buttons, or how the economics are split with Flipkart, or whether the feature will survive past the festive season. Companies in this space routinely describe tests as early even when the ambition behind them is clearly larger.
The ambition is not subtle. Google has said repeatedly that it sees commerce as one of the few places where an AI assistant can prove durable value, because a purchase is a concrete action with a clear outcome. Search already shapes what billions of people buy; an assistant that completes the transaction would sit one step closer to the money.
That is what makes the dark-web finding uncomfortable for the company’s narrative. The value Google is trying to attach to its AI, a trusted path from question to purchase, is the same value that is being undercut by markets selling access to the models themselves at a fraction of the price.
Google has not said whether the threat-intelligence work has led to any takedowns, or how it intends to stop the resale of access to its models. The company’s disclosure, like its buy button, is a first step rather than a finished product.
What the week showed, in two markets a world apart, is that AI has become a real business in both the clean and the shadow versions of the economy. Google wants to own the clean one. The dark-web listings are a measure of how much of the other one is already running.
The timing of the India test is no accident. The festive season, built around Diwali and the sales that surround it, is the moment when Indian shoppers spend most heavily, and e-commerce platforms and payment firms build their year around it. A buy button that works through Gemini during those weeks would give Google a window into whether consumers will trust an assistant to hold their money, not just their attention.
Google already runs the infrastructure the feature leans on: the Shopping Graph, a dataset of billions of products that feeds recommendations across its surfaces. Flipping a product from a listing into a completed sale is a small step technically and a large one commercially, and the company has been inching toward it for years with features that compare prices and track deals. The Flipkart button is the point where those inches add up to a transaction.
