This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
A year ago, when Google introduced the first Gemini model, I wrote that it might change who summarizes you first. Yesterday’s launch of Gemini 2.0 suggests the next question: what happens when the system that summarizes you is also the one that acts on the summary?
What Google announced
Google calls Gemini 2.0 a model for “the agentic era.” In his introduction, Sundar Pichai described agentic models as ones that can “think multiple steps ahead, and take action on your behalf, with your supervision.” The first release is an experimental version of Gemini 2.0 Flash, available to developers and, in a chat version, to Gemini users. Google said it would bring Gemini 2.0’s reasoning to AI Overviews, which it says now reach a billion people, for more complex, multi-step questions.
Two other pieces matter more for reputation.
The first is Deep Research, launched in Gemini Advanced. It browses the web on a user’s behalf and compiles a report with links to sources. Google’s own examples include a small business owner gathering a competitor analysis.
The second is Project Mariner, a research prototype that runs as a Chrome extension. It reads what is on the browser screen, including text, images and forms, and can type, scroll and click to complete tasks. Google says it scored 83.5% on the WebVoyager benchmark, is “not always accurate and slow,” acts only in the active tab, and asks for confirmation before sensitive actions such as purchases. For now it is limited to trusted testers.
A shorter chain
Here is the information chain most reputation work has been built around. Something happens. Media and websites write about it. Google indexes those pages. A person searches, scans the results, maybe reads a few, and forms an impression. More recently, an AI summary sits in front of the links and forms part of the impression for them.
Agents compress the chain further. With Deep Research, the output is not a page of links but a finished document: a competitor analysis, a vendor comparison, a briefing on a company. That document can be saved, forwarded and used in a decision without anyone at the company ever knowing it exists. With something like Mariner, the step after reading could be an action: a shortlist, a form submitted, a purchase started.
At each step, the human checkpoint gets thinner. A person scanning ten results might notice that one source is a forum post from 2019. A report compiled by an agent may simply include it, phrased in the same confident tone as everything else.
What changes for reputation teams
The old question was “what does the answer say about us?” The new one is closer to “what will an agent conclude about us, and what will it do next?”
That shifts attention in a few directions. Facts that agents need in order to act, such as pricing, policies, locations, leadership and contact details, should be stated plainly on your own pages and consistent with what third parties say. If your site is hard for software to read, an agent may lean more heavily on other people’s descriptions of you. And the sources that research tools are likely to draw on, from news coverage to Wikipedia to review sites, matter even more, because their claims may now be repackaged into documents you never see.
It is worth being careful about pace. Mariner is a prototype for testers, and Google itself says it is slow and imperfect. Deep Research is a paid feature. Nobody can say yet how many decisions will run through tools like these next year.
But the direction is clear enough. For two decades, reputation online has mostly meant managing what people find. Gemini 2.0 is an early sign that it will also mean managing what software reads, concludes and then does on someone’s behalf.