The agent who visits your website knows the person who sent it.
This is the change under that of Google Gemini Deep Search Maxlaunched on April 21, 2026, in public preview on the paid Gemini API tier. Deep Research Max itself is a restricted deployment. The model he proposes is a preview of what Web Agent will become when the other major vendors follow suit, which they usually do within a quarter or two on features like this. When a mixed recovery agent runs, it arrives with private context: the user’s financial data, their file stores, their connected business data feeds, all merged into the query before the agent hits a page.
For web professionals, this is the next chapter of history of the Web agent. The claim that agents are a new core class of visitors has been going on for months. Since then, the claim has evolved. Agents are a new primary visitor class with a private context. The reasoning that decides whether your page responds to a query runs on a set of inputs larger than your page. How much weight the agent gives to your content depends on whether or not it adds anything that private sources haven’t already provided. It’s time for mixed recovery in the Web agent story, and it’s on the supply side, how agents collect, not the user-facing product layer.
The old AI search optimization posture (writing content that matches the keyword query) was weakening before this. He’s getting even weaker now. The new posture is that of structural predictability: own entity relationshipscanonical identity, live data, independence rendering. The structure is functionally important to the agent. When the agent comes up with the context, the content it selects is the content its model can cleanly merge with whatever it already has.
Mixed recovery provides insight into the next layer of the Web agent
Google’s Gemini Deep Research Max, publicly previewed on the paid API tier starting April 21, can extract four classes of input into a single reasoning loop: the public web, file uploads, connected file stores, and arbitrary remote MCP servers. According to Google’s announcement, the agent “searches the web, arbitrary remote MCPs, file uploads, and connected file stores, or any subset of them.”
The two new classes (filestores and remote MCPs) share a property. They are private by default. The agent reads them only with the user’s consent. Once connected, a financial data provider or enterprise CRM exposes their data to Gemini via the Model Context ProtocolAnthropic’s open standard with more than 97 million installs from March 2026. Google Agent retrieves from these private sources with the same reliability as it reads from the open web, in the same reasoning process.
This is the structural change that everyone looking at Agent Web has been waiting for from a major vendor: the public web and private context, merged by the agent, into a single request. Gemini is first.
The model is not there yet for most operators. Deep Research Max is a public preview behind a paid API, not a feature of the Gemini consumer app. Most websites will not be read by a mixed recovery agent this quarter. What Google announced on April 21 is direction, not arrival. Treat it as a leading indicator: if this architecture is evolving, and the main suppliers generally copy each other in a quarter or two days, with capacities like this, the operator’s work becomes real before the traffic.
Signal sharing breaks down when the agent has better alternatives
In a mixed fetch query, each connected source competes for signal sharing: the open web, the user’s file stores, and any private MCP servers. The weight of a single source is proportional to how well the agent can extract and merge its signal with whatever the agent holds.
For public websites, this changes the competitive terrain in two ways.
First, machine-driven websites gain more citation shares. A page with clean structured data, unambiguous entity relationships and rendering that does not hide content behind JavaScript It is easy for the agent to merge with the user’s private context. The merged answer refers to the first page of the machine, because that page provided usable, mergeable material.
Second, poorly structured websites lose the signal share they previously got for free. In the web-only era, even a messy page could appear in a citation because there was no better alternative to the public web. In the age of mixed recovery, the alternative may be user-uploaded documents or a connected MCP with cleaner data. The messy content page loses the quote sharing it used to split with clean sources.
This is a different competition from classic SEO. Classic SEO ranks pages relative to each other. Mixed retrieval classifies pages based on the user’s own context. You cannot see competing sources. You can only ensure that when the agent reaches your public page, the page brings something extractable and unambiguous.
Diagram of structured products and offers is cited more often than unstructured descriptions when the user’s private context touches on something related to it. Canonical identityclean entity relationships and the rendering of independence all become more leveraged when the agent merges the signal between sources. THE Adobe AI Traffic Reversal Q1 2026 was demand-side proof that structured commerce wins in AI research; mixed recovery is the supply-side mechanism that produces the same effect on the rest of the web.
The honest counter-read: some queries scan your entire website
Not all mixed fetch queries will end up citing a public website. Some queries will be able to be answered entirely from the user’s connected sources. A financial analyst running Deep Research Max on an internal MCP server, along with downloaded quarterly reports, may never need the public web for this answer. The traffic for this request is not going anywhere; the answer is satisfied within the boundary of the private context.
This is a true subset. Most queries still mix public and private sources, because most analytical questions touch on both.
Mixed recovery does not mean that every website gets less traffic. This means the agent is more selective about what it uses. The bar goes up for sources chosen by the agent. Deep Research Max is a glimpse into what agent web is about to demand. Machine-Focused Websites will take its share when this scale arrives. Unstructured content will continue to lose it. Google showed us the model on April 21, but scale is where the real work of web professionals begins, and it’s time to do that work before traffic catches up.
More resources:
This article was originally published on No hacks.
Featured image: RobinRmD/Shutterstock
