The Browser Became an Answer Engine
The shift is easy to state: the browser is turning from a window onto pages into an answer engine. Three products define the category as of this writing, and each earns its place for a different reason.
ChatGPT Atlas
Atlas is OpenAI's browser: Chromium-based, macOS-first, free to use, with ChatGPT built into the frame rather than bolted on as a sidebar extension. Its headline capability is agent mode, gated behind the paid ChatGPT tiers, which performs multi-step tasks inside the browser on your behalf, researching a topic across sites, filling forms, making bookings, the kind of work that used to mean twenty minutes of your own clicking. Atlas also ties into ChatGPT's memory, so the browser knows what you were researching last week and answers accordingly. Used for what it is good at, it is genuinely impressive: the gap between wondering something and having a defensible answer has never been smaller, and for task-shaped browsing, book this, compare these, fill that in, agent mode is the closest thing yet to the browser doing the errand for you.
Perplexity Comet
Comet is the research specialist. Free since October 2025 and available across platforms, it comes from a company whose whole identity is answering questions with sources attached, and it shows. Comet's strength is working across your open tabs: ask it to compare the four vendors you have open, summarize the disagreement between two papers, or pull the pricing out of everything on screen, and it does so quickly and with citations. For anyone whose work is reading and comparing, an analyst, a student, a journalist, a founder evaluating tools, Comet turns an afternoon of tab-switching into a conversation, and it threads its memory through those research sessions so a project picked up on Thursday remembers what you established on Monday. It is the AI browser that most clearly understands research as its core use case.
Dia
Dia comes from The Browser Company, the team behind Arc, now part of Atlassian, and it is the most opinionated of the three about what a browser should feel like. It is macOS software in the best sense, fast and considered, and it is built around the idea that your tabs are a workspace to think in, not just a row of pages: you chat with a tab, with several tabs, with your history, and Dia builds its memory around your tabs and its skills around your repeated workflows. Where Atlas leans on the power of ChatGPT and Comet leans on the discipline of search, Dia leans on design, on making cross-tab work feel native rather than clever. For people who live in their browser all day, it is the one that feels least like a chatbot wearing a browser costume.
Three different bets, one shared conviction: reading the web is now the machine's job as much as yours. On that, all three deliver. The problem this article is about starts after they deliver.
The Reading Half and the Keeping Half
Research has always had two halves. The reading half: find, skim, compare, understand. And the keeping half: decide this mattered, and put it somewhere you will find it again. AI browsers compress the first half brilliantly. Nobody notices what happened to the second half until about three months later.
Three months later sounds like this: that pricing table the browser summarized for me in March, where is it now? The comparison of the four frameworks, the paragraph that finally explained the regulation, the spec you based a decision on, all of it was read, understood, used, and never kept. In the old browser this failure was at least visible: you either bookmarked the page or you watched yourself not bookmark it. In an AI browser the failure is invisible, because the moment of understanding no longer happens on the page at all. It happens in the answer, and the answer is a chat message. Chat scrolls. Answers evaporate. The synthesis that felt so permanent while you read it has the retention properties of a conversation, which is to say, none.
There is a second, subtler loss. When the browser answers across your tabs, what you take away is the conclusion, and the conclusion arrives already detached from its evidence. You know the verdict, vendor B is cheaper at your scale, but the table it came from is behind the answer, on a page you never actually visited, in a tab you closed an hour later. That is the answer-not-the-source problem: you kept what you decided and lost what would let you re-check, re-argue, or re-use the decision. For casual questions this costs nothing. For real research, the kind you will be asked to defend in a meeting or revisit in a quarter, an answer without its source is a memory of having known something.
Browser Memory Is One More Vendor Cache
The obvious objection is that the browsers themselves solve this now. Atlas remembers through ChatGPT, Dia builds memory around your tabs, Comet threads memory through your research. If the browser remembers, why keep anything? Because browser memory is a cache with an owner, and we have seen this movie before.
We made this argument at length about assistants in the memory wars piece: every AI vendor's memory is genuinely useful and, at the same time, a retention mechanism, because context that lives inside one product is context that keeps you inside that product. AI browsers extend the same dynamic to a new surface, and arguably a more intimate one, since your browser sees not just what you asked but everything you read on the way to asking it. Atlas's memory of your research is ChatGPT's memory, held on OpenAI's side, in OpenAI's format. Dia's understanding of your working life is Dia's. Comet's picture of your ongoing projects is Comet's. None of this is sinister, and all three deserve credit for building memory that makes daily browsing better. But not one of these memories is a record you can open in full, hand to a different tool, or take with you.
And you will want to take it with you, because the category is a year or two old and churning hard. The Browser Company has already been through one acquisition. Pricing tiers, features, and gating have shifted across all three products since launch, and the honest expectation for 2027 is more movement, not less. Switching browsers used to cost you some bookmarks and muscle memory. Switching AI browsers costs you your research memory: every session the browser summarized, every project thread it was quietly maintaining, stays behind in the old product. If your research trail lives inside whichever browser's memory you happened to use this year, then your research trail has the life expectancy of a browser choice, and browser choices, as of this writing, are not looking long-lived.
Tab Hoarding, Now With a Chatbot on Top
AI browsers have also done something unexpected to the oldest bad habit on the web: they made tab hoarding feel rational. If the browser can reason across forty open tabs, then keeping forty tabs open starts to look less like clutter and more like maintaining a working context.
It is a seductive upgrade to the old excuse. Tabs used to stay open because closing them felt like losing them; now they stay open because they are, in a real sense, the AI's working set, and every open tab makes the answers a little better. Some people have begun treating the tab bar as their research database and the chat box as its query language. The trouble is that everything that made a wall of tabs fail as a system still holds. Tabs die with a crash, a restart, an accidental window close. They pin your research to one machine and one browser profile. They mix the three pages that mattered with the thirty-seven that were just passing through, and no amount of intelligence layered on top changes what the pile underneath is. A pile with a chatbot on top is still a pile. The chatbot just makes it more comfortable to keep stacking.
We wrote a whole guide on why tab hoarding happens and how to finally close them all without losing anything, at /blog/how-to-organize-browser-tabs, and its core line survives the AI era intact: a tab is a reminder with no message. An AI browser can now read the reminder aloud and even summarize it, which is genuinely helpful, but the reminder is still not a kept thing. It is a page you have not yet decided about, and the deciding, this mattered, keep it, was always the actual work.
The Workflow That Survives Browser Churn
None of this argues against using an AI browser. It argues for a clean division of labor: let the browser own the reading half completely, and refuse to let it own the keeping half at all.
The workflow is short enough to state in a breath. Read and interrogate in whatever AI browser you like, Atlas for agent-shaped tasks, Comet for sourced research, Dia for living in, and switch freely as the category evolves, because none of your research depends on the choice. Then, the moment something proves valuable, and only then, save it out: the source itself, the link, the PDF, the screenshot of the pricing table, together with the conclusion you drew from it, into a library that no browser owns. Not the answer alone, because an answer without evidence cannot be re-checked. Not the tab, because a tab is not a copy of anything. The source plus the conclusion, in a home outside the browser.
- Read anywhere Use the AI browser at full power: summaries, cross-tab comparisons, agent runs. This half of research is genuinely solved, and there is no virtue in doing it the slow way.
- Keep deliberately When something crosses from interesting to load-bearing, a page you will cite, a table you will act on, a document you will need again, save the source and your conclusion at that moment, while both are in front of you.
- Keep it outside every browser The library must survive switching from Atlas to Comet to whatever ships next year, which means it cannot live in any of their memories, sidebars, or clouds. A library you own outlives every browser choice you will ever make.
What this buys you is independence from the churn. The AI browsers can keep competing, shipping, merging, and repricing, and you can keep choosing the best one each year the way you choose the best model each year, without ever asking how much accumulated research you are willing to abandon at the door. The keep step is one deliberate act per valuable source. In an AI browser that act matters more than it ever did, not less, precisely because the browser has made it so easy never to perform: the reading feels complete, the memory feels safe, and neither feeling puts the pricing table where you can find it in March.
Where Mindly Fits
The workflow above stands or falls on one number: how much the keep step costs. If saving a source takes six clicks and a filing decision, you will stop doing it inside a week. Mindly's answer is to make it one shortcut from any browser, with the filing done for you.
Press ⌘M in Atlas, Comet, Dia, Safari, or anything else on your Mac, and whatever is in front of you goes into your library: the page as a link, the PDF as a file, the pricing table as a screenshot, your conclusion as a note, even a spoken voice memo if talking is faster. Mindly then does the librarian work that the browsers do invisibly, except here it is done in the open, on items you own: AI tags each save by topic, summarizes long content, transcribes audio, reads the text inside screenshots, and links related items together, so the vendor comparison you saved from Comet sits connected to the notes you took in the meeting about it. When March arrives and you need that table, you search in plain language, the pricing comparison for the email tools, and semantic search finds it by meaning, not by whichever title the tab had. You can open any saved item and chat with it, asking questions answered from your own kept source rather than from a model's recollection, and a mind map shows how a season of research actually connects.
And the library sits where this whole article has been arguing it should: in a folder on your Mac, local, outside every browser vendor's memory. AI processing runs over encrypted channels and your content is not retained on Mindly's servers after the request completes. There is no account to create. The free tier holds 25 items, enough to run the keep step on a real research project and feel the difference, and Mindly Pro removes the limit at €7.99 a month or €44.99 a year. Use whichever AI browser wins your affection this year. Just give the research it helps you do a home that will still be yours when you change your mind.
Free for macOS, no account needed. Let the browser do the reading, and give the keeping half a home you own. Download Mindly →