>_TheQuery
// Reading nowStart
← All Articles

Google IO 2026 Built a Dependency Trap in Three Moves. Developers Noticed by Thursday.

By Addy · May 27, 2026

On May 19, Google held its annual developer conference and announced products across five surfaces simultaneously. The headlines wrote themselves: Gemini Spark, the 24/7 persistent agent. Gemini Omni, the unified multimodal model. Gemini 3.5 Flash, faster than every comparable frontier model. AI Mode moving from experimental to the core of Google Search. A Google I/O for the ages.

By Thursday, May 21, the developer community had read the fine print.

The Hacker News thread on the Gemini CLI shutdown had a comment that cut through everything: "developers will never adopt our products if we keep killing them." It collected hundreds of upvotes. It did not get a response from Google.

The three announcements that generated the most enthusiasm on Tuesday generated the most anger by Thursday. Not because the products were bad. Because developers had been through this before with Google, recognized the pattern in real time, and said so loudly and specifically.

Move One: Google Search Stops Sending Traffic

The first move was the broadest and the slowest to land.

Google announced the most significant redesign of Search in over 25 years. AI Mode, the conversational, Gemini-powered interface that has been in experimental testing, is now central to the product rather than optional. The classic list of blue links is being surrounded by AI-generated summaries that trigger conversations. Agentic capabilities let users request multi-step tasks. The search box now accepts images, files, videos, and Chrome tabs alongside text.

The user-facing product is genuinely better for the person searching. Google says AI Mode has surpassed one billion monthly users, and that queries in AI Mode have more than doubled every quarter since launch. The product is stickier, more capable, and more deeply integrated into how people actually want to research things.

The publisher-facing consequence is the part Google did not headline. AI Mode already has over one billion monthly users. Third-party trackers disagree on the exact share of searches that trigger AI Overviews, but the direction is clear: more answers are being resolved inside Google's own interface. The zero-click rate, searches where the user gets their answer from Google's AI summary and never visits a website, is increasing. Research published in May 2026 found that AI Overviews give Google unprecedented editorial control, sometimes include unsupported claims, and can suppress publisher click-through even when publisher pages are cited.

For publishers, this is not a theoretical future problem. It is a present one. The traffic that used to flow from Google Search to the websites that produced the content Google's AI is now summarizing is declining. The publisher who spent years building Google-organic traffic as their owned distribution channel is discovering that "owned" was always conditional on Google continuing to send the traffic.

TheQuery, this publication, was built on the premise that Google traffic is owned and Reddit traffic is rented. That premise requires a revision: Google traffic is owned until Google decides it would rather summarize your content than link to it. The revision does not invalidate the strategy. It clarifies the risk.

The new skill the shift creates is called AEO: Answer Engine Optimization. Where traditional SEO optimized content to rank in a list of links, AEO optimizes content to be cited inside an AI-generated answer. The differences are specific. AI systems favor content that is structured, sourced, authoritative, and organized around questions rather than keywords. Content that answers "what is X" clearly and specifically gets cited. Content that ranks for "X" by volume and backlink density may not.

Publishers who built on Google organic traffic need to be doing two things simultaneously right now. Publishing content that answers specific, citable questions with clear sourcing and structured prose. And tracking whether their content appears in AI Overviews, not only whether it ranks in blue links. The metric has changed. The content strategy needs to follow.

Move Two: Six Thousand Contributions, Then the Gate

The second move was the one that made developers genuinely angry.

In June 2025, Google released Gemini CLI, a TypeScript-based AI coding agent under the Apache 2.0 license. The open-source release was real: full access, community contributions welcome, free tier with 1,000 requests per day. Over the following year, the tool accumulated over 100,000 GitHub stars and 6,000 merged pull requests. Google's own announcement at I/O cited this community investment as evidence of the project's success.

On May 19, 2026, Google announced that Gemini CLI would stop serving requests for all free users, Google AI Pro subscribers, and Google AI Ultra subscribers on June 18, 30 days from announcement. The replacement is Antigravity CLI, a tool built in Go that by Google's own acknowledgment does not yet match Gemini CLI feature-for-feature.

The reaction from the developer community was not confusion. It was recognition.

The GitHub discussion thread was direct: developers explicitly preferred Gemini CLI and asked for it to stay alive. The engineering Slack channels that had recommended Gemini CLI to new team members spent the week walking those recommendations back. On Hacker News, the reaction was not "how do we migrate?" so much as "Google did this again."

The specific complaints were precise rather than vague. The Antigravity CLI binary, invoked as agy, was not available on npm or Homebrew as a named package as of the first wave of developer testing after the announcement. The weekly quota on Antigravity drew immediate complaints from users who said they hit limits much faster than they had on Gemini CLI. Quota refreshes weekly rather than daily, meaning a developer who hits the limit on Monday waits until the following week rather than the following day.

The reading that circulated in engineering communities was not conspiratorial. It was structural: Google accepted open-source labor to build a capable tool to 100,000 stars, then withdrew the free consumer backend and closed the successor. Whether that reading is fully fair is less important than the fact that it informed toolchain decisions made by engineering teams across the week following I/O. Several teams publicly announced they were evaluating Claude Code, Cursor, and Codex as replacements for workflows that had been running on Gemini CLI.

The most damaging detail did not come from critics. It came from Google's own product manager, who wrote that the Gemini CLI repository remains available as Apache 2.0 code. What the announcement did not solve: without Google's backend serving API requests, the repository becomes much less useful to the people who built workflows around it. A working fork requires independent access to a compatible frontier model, or a paid API path. That is a non-trivial engineering undertaking. The Apache 2.0 license kept the code available. Google kept the backend. The practical community use ended.

Enterprise accounts were unaffected throughout. Organizations with Gemini Code Assist Standard or Enterprise licenses kept access. The gate was built specifically at the boundary between community users who contributed and enterprise customers who pay. The community who built the 100,000 stars were on one side of it. The enterprise customers who benefited from the ecosystem those stars represented were on the other.

Move Three: Your Life on Google's Servers

The third move was the most ambitious and received the least critical scrutiny during the week.

Gemini Spark, announced at I/O and beginning beta rollout to AI Ultra subscribers this week, is a 24/7 persistent agent that runs on Google Cloud infrastructure. It has its own Gmail address. It reads your inbox. It checks your calendar. It monitors your documents. It executes tasks, drafting responses, scheduling meetings, surfacing information, while your phone is in your pocket and your laptop is closed.

The product is genuinely compelling and the architecture is genuinely impressive. The Gemini Spark article this publication ran earlier this week described the shift from passive to persistent AI as one of the significant transitions in computing history. That assessment is accurate.

The dependency it creates is also significant, and the I/O week context makes it legible in a way it would not be without the Search and CLI stories alongside it.

A publisher who built on Google organic traffic is now watching that traffic route through Google's AI summaries instead of to their website. A developer who built workflows on Gemini CLI is now migrating to a replacement with lower consumer-tier certainty on a 30-day clock. A user who activates Gemini Spark is giving Google's servers continuous access to their email, calendar, documents, and the agent loop that runs across all of them, on a service that Google can modify, gate, price, or deprecate on the same 30-day notice it gave Gemini CLI users.

These are not the same kind of dependency. Search traffic loss is a business problem. Gemini CLI shutdown is a toolchain problem. Gemini Spark dependency is a data and continuity problem. But they are all expressions of the same structural condition: Google builds toward your dependency, then exercises discretion over the terms.

Sam Altman described the vision of persistent AI as setting intentions rather than issuing commands, telling your AI what you want to accomplish, and having it run in the background on your behalf. The vision is real. The question Altman did not answer, and that Gemini Spark does not answer, is whose servers your intentions run on, and what happens when the terms of that service change.

The Pattern That Predates This Week

Google did not invent this pattern. It has executed it more consistently than most.

Google Reader launched in 2005, accumulated tens of millions of users who reorganized their information consumption around it, and was shut down in 2013. The shutdown created a specific kind of loss: not just the product, but the infrastructure of habits and workflows built on top of it. Feedly and The Old Reader absorbed the refugees, but the trust gap was real and has persisted in how developer communities evaluate Google product launches.

Google Plus launched in 2011 with forced integration into every Google service, accumulated 500 million registered accounts through that forced adoption, and was shut down in 2019 after the user engagement data showed that most of those accounts were not meaningfully active. The lesson was that forced integration produces account creation, not genuine use.

Google Stadia launched in 2019 with promises of platform longevity, attracted game publishers who built titles for the platform, and was shut down in 2023. Players who had purchased games lost access. Publishers who had invested in Stadia-specific development received compensation but not the platform stability they had signed on for.

Gemini CLI is not Reader, Plus, or Stadia. The shutdown is a migration rather than a deletion. The repository survives, enterprise access continues, and a replacement exists. But the pattern of building community investment and then changing the terms at the enterprise boundary is legible to any developer who has been paying attention, and the Hacker News community has been paying attention for twenty years.

The reaction was not surprise. It was confirmation of a prior.

The Other Side: Google Has Reversed Course Before

Before treating any of these three moves as permanent, one counterpoint deserves its own section.

Google reverses decisions. Not always, not predictably, but with enough frequency that writing off a deprecated product or a changed policy as final is not always correct either.

The clearest recent example is third-party cookies in Chrome. Google spent three years announcing, delaying, and preparing the industry for the deprecation of third-party cookies, a change that the entire ad-tech ecosystem had been building alternatives toward. Engineers rewrote attribution systems. Publishers restructured data strategies. Agencies restructured how they measured campaign performance. Then Google abandoned the plan entirely, citing regulatory scrutiny and industry opposition, and instead introduced a user-choice model that kept cookies largely intact.

Three years of industry investment in alternatives. One reversal. The companies that had moved furthest and fastest away from cookie-based tracking found themselves ahead of a transition that did not happen.

The AI Mode redesign of Search is the current version of that kind of bet. If AI-generated summaries produce enough inaccurate results, enough publisher backlash, enough regulatory scrutiny about anticompetitive reduction of referral traffic, Google may walk back the zero-click experience, reintroduce blue links more prominently, or add publisher citation requirements that restore some traffic flow. The EU is already investigating whether AI Overviews violate competition law by reducing traffic to the publishers whose content trained the models. That investigation has teeth and a track record of producing concessions.

The Gemini CLI shutdown has a more specific reversal mechanism. The Hacker News thread that collected hundreds of upvotes on "developers will never adopt our products if we keep killing them" is exactly the kind of signal Google's developer relations team reads and acts on. Google has restored deprecated developer tools before when the community reaction was sufficient and the enterprise replacement was not yet capable of absorbing the workflow load. The 30-day migration clock is aggressive enough that the backlash may not be fully legible to Google's decision-makers until after June 18. If Antigravity CLI's quota limits and feature parity gaps produce a sustained drop in developer adoption metrics, a Gemini CLI extension or a more generous Antigravity free tier is not an implausible outcome.

The honest read is neither "this is permanent" nor "this will definitely reverse." It is: Google makes decisions based on signals, and the signals from this week's developer community are loud enough to register. The third-party cookie reversal happened because the industry signal was sustained, organized, and backed by regulatory pressure. The Gemini CLI signal is loud but not yet organized and not yet regulatory.

Watch the Antigravity adoption metrics. Watch the EU investigation into AI Overviews. Watch whether Google's developer relations team publishes a response to the Gemini CLI GitHub discussion. These are the signals that will tell you whether May 19 was a permanent pivot or an opening position in a negotiation.

History says it might be either. Planning as if it is permanent, while tracking whether it reverses, is the only defensible approach.

What This Means for How You Build

The practical question underneath the anger is the useful one: given this pattern, how should developers and publishers think about building on Google surfaces?

The answer is not to avoid Google. The distribution advantages of Google Search, the technical capabilities of Gemini models, and the infrastructure of Google Cloud are real and significant. A publisher who ignores Google Search optimization because Google changed the rules is making the same mistake as a developer who ignores Gemini models because Google killed a CLI. The mistake is treating the relationship as stable rather than managed.

The AEO framework for publishers is specific and actionable right now. Structure content around questions that AI systems are being asked. Write clearly, source explicitly, and make the authoritative answer obvious in the first paragraph. Track AI Overview appearances, not just blue-link rankings. Publish content that a language model can cite confidently rather than content that keyword-matches a query. The metric that matters has shifted from "rank" to "citation."

For developers building on Google tooling, the operational lesson is the same one every dependency relationship teaches eventually: own the abstraction layer above the vendor, not the vendor's implementation. Teams who built workflows around Gemini CLI's specific commands are migrating those commands. Teams who built workflows around the capability, AI-assisted coding in the terminal, are evaluating which tool currently delivers it best and will switch again when that changes. The former is a 30-day crisis. The latter is a 30-day decision.

For the persistent AI wave that Gemini Spark represents, and that Codex's locked-screen feature and Claude's computer use capabilities also represent, the dependency question is worth naming before the workflows are built, not after the terms change. An agent that runs on your machine and calls external APIs maintains a separation between the execution environment and the service provider. An agent that runs on a cloud VM maintained by the service provider collapses that separation. Both architectures can deliver the same experience. They create different risks when the service provider changes the terms.

Google built three new dependencies in one week. One for publishers. One for developers. One for users. Each one is backed by genuine capability. Each one is backed by a track record that the Hacker News thread summarized more concisely than anything else written about I/O 2026:

"Developers will never adopt our products if we keep killing them."

Google heard that. Google has heard some version of it for fifteen years. The products keep getting built. The dependencies keep getting created. The gates keep getting moved to the enterprise boundary when the community investment is sufficient.

The pattern is not a secret. It is the business model. Understanding it is not a reason to avoid Google's products. It is a reason to understand exactly what you are building on, and plan accordingly.

Sources:

Previously on TheQuery: Gemini Spark and the End of the Session: How AI Is Shifting From Passive to Persistent and Microsoft Just Killed Its AI Sidekick. The Story Behind That Decision Is Worse Than the Decision.