TL;DR
In one week, the Chrome side panel became the most contested real estate in consumer software. On July 9, 2026, OpenAI announced it's folding the shutting-down Atlas browser into a ChatGPT Chrome extension. On July 14, 2026, Google expanded Gemini in Chrome to UK desktop users, with multi-tab comparison, Calendar/Maps/Gmail/YouTube integration, memory of past conversations, Nano Banana 2 image editing, and stated prompt-injection defenses.
Three kinds of software now want that panel: ChatGPT, Gemini, and small single-purpose tools like Convert: Web to PDF. They are not the same kind of software, and the difference isn't quality — it's scope.
An AI side panel is broad by design. It needs page context, memory, and cross-service reach, because that's what makes it good. A converter is narrow by design. It needs the tab you're on when you click it, and nothing else, ever.
That distinction stops being philosophical on August 1, 2026, when Chrome Web Store enforcement begins on a policy requiring that collected data be strictly necessary to an extension's disclosed single purpose. That's an easy bar for software that does one thing, and a genuinely hard one for software designed to do many.
This post is not a hit piece on AI panels. They're useful, Google shipping prompt-injection defenses is real engineering against a real problem, and we're going to say so.
What just happened, in one week
July 9: OpenAI unbundles Atlas into an extension
OpenAI is shutting down the Atlas browser on August 9, 2026, about nine months after launch. The stated rationale — "the browser is a feature, not the destination" — comes with a plan, not an exit.
Atlas's agentic features are being redistributed into:
- A ChatGPT Chrome extension — page context, summarize, ask, and launching longer tasks. Aimed squarely at Google's Gemini side panel.
- An upgraded ChatGPT desktop app with a built-in browser — login, downloads, page interaction.
- A server-side cloud browser for agent tasks.
OpenAI concluded that owning the whole browser was the wrong shape, and that sitting inside the browser you already use is the right one.
July 14: Gemini in Chrome expands to the UK
Five days later, Google expanded Gemini in Chrome to UK desktop users, with iOS slated for "next month." The feature list is the real story:
- Multi-tab comparison — reasoning across several open tabs at once
- Calendar, Maps, Gmail and YouTube integration
- Memory of past conversations
- Nano Banana 2 image editing
- Stated prompt-injection defenses
Look at that list as an engineer rather than a shopper. Every single item is a scope expansion. Multi-tab means it reads more than the tab you're on. Integration means it reaches services outside the browser. Memory means state persists across sessions. Each one makes the product better. Each one widens the surface.
That's not a criticism. It's a description — and the AI panels would be useless without it.
Three tenants, two kinds of software
Here's the frame that actually explains the week.
AI side panels are broad by design. Not by carelessness, not by greed — by design. An assistant that can't see your page can't help with your page. One that forgets everything between sessions is worse. One that can't reach your calendar can't tell you if you're free. Every capability that makes an AI panel worth installing requires it to see more. Breadth is the product.
Single-purpose tools are narrow by design. Convert: Web to PDF needs the tab you're on, at the moment you click, and nothing else — not your history, not your calendar, not last Tuesday. Narrowness isn't a limitation we're apologizing for. It's the spec.
These are different categories of thing that happen to live in the same real estate. Comparing them on "which is better" is a category error. The useful comparison is what each one costs you and what happens when it fails.
The comparison table
| AI side panel (ChatGPT / Gemini) | Single-purpose local tool | |
|---|---|---|
| Scope | Broad by design — page context, multi-tab, memory, cross-service | Narrow by design — one tab, on click, one job |
| Data reach | Page content, and increasingly calendar / mail / maps / video; memory persists across sessions | The tab you're on, only when you click it. Nothing retained |
| Where processing happens | Server-side models. Content leaves your machine — that's how inference works | On-device. Chrome's own print engine, zero network requests during conversion |
| Account required | Yes | No |
| What breaks if it shuts down | The capability goes. Memory and anything account-held goes with the account | The tool stops. Every PDF it made is already on your disk and unaffected |
| Works offline | No — it's a model call | Yes. Local rendering of a loaded page |
| "Strictly necessary to a single purpose" | Genuinely hard bar — breadth is the purpose | Easy bar — the purpose is one sentence |
| Prompt-injection exposure | Real and inherent; Google ships stated defenses | Not applicable — nothing reads the page to decide what to do |
That last row deserves care rather than a victory lap. We don't have prompt-injection exposure the way a shop with no restaurant doesn't have food poisoning risk. It's not virtue. It's just the absence of the mechanism — nothing in a print engine interprets page content as instructions.
The August 1 policy is where this gets concrete
On July 1, 2026, Google announced updated Chrome Web Store Developer Program Policies, enforced from August 1, 2026. Four changes; two matter here:
- Collected data must be strictly necessary to the extension's disclosed single purpose.
- All collection must be prominently disclosed, with proactive notice if practices change post-install.
Now notice the timing. OpenAI announced a ChatGPT Chrome extension on July 9. Google expanded Gemini in Chrome on July 14. The single-purpose data policy starts being enforced on August 1. Same summer, same surface.
"Strictly necessary to a single purpose" is a much easier bar for a tool that does one thing. Watch how short the argument is for us:
Purpose: convert the current webpage to a PDF. Data strictly necessary: the current webpage. Data collected: none — conversion is local, with zero network requests.
Done. There's no interpretive work because there's nothing to interpret.
Now try writing that paragraph for an assistant with multi-tab comparison, cross-service integration, and persistent memory. It's not that an honest answer is impossible — it's that the answer is long, and every clause is a place where "strictly" is doing contested work. When your purpose is "help with anything on the web," almost everything is arguably necessary. That's exactly the elasticity the single-purpose requirement exists to squeeze.
We're not predicting enforcement outcomes and we're not accusing anybody of anything. Both OpenAI and Google are large, well-lawyered organizations that will file disclosures fully aware of what August 1 says. The observation is narrower and harder to argue with: the policy is structurally easier to satisfy the smaller your software is. Which is our whole manifesto line one — software should be small — showing up in someone else's rulebook.
The practical move for you is ninety seconds long: open any extension's Privacy practices tab on its Chrome Web Store listing, read the single-purpose sentence, read the declared data categories, and ask what one has to do with the other. Do it to the AI panels. Do it to us. That gap — between the stated job and the declared appetite — is what August 1 is aimed at, and it's the same gap that has burned people for years in the malicious-extension cases and in PDF tools that quietly upload your files.
Being fair to the AI panels
The lazy version of this post sneers. We're not doing that, for three reasons.
They're genuinely useful
"Compare these four tabs and tell me which flight is actually cheaper once bags are included" is a real task that a converter cannot touch. Neither can a bookmark manager, a scraper, or anything else in our shop. Multi-tab reasoning is a legitimately new capability. Cross-service integration — knowing your calendar while reading a booking page — is a real thing that saves real time. Memory across conversations makes an assistant less exhausting.
If those solve your problem, install one. We'd rather be honest about where we're irrelevant than pretend a print engine competes with a language model.
Prompt-injection defenses are real engineering
Gemini in Chrome shipping stated prompt-injection defenses deserves credit specifically because prompt injection is a genuinely hard, genuinely unsolved problem. When an AI reads a webpage and that page contains text crafted to look like instructions, the model has to distinguish content from commands — and that distinction is the entire attack surface of the agentic web.
A vendor that ships defenses is a vendor that has admitted the problem exists. That's better engineering posture than one that hasn't. It's the hard, unglamorous work, and treating it as a punchline would be both unfair and wrong. We covered the shape of this risk class after ClaudeBleed.
Broad isn't a synonym for bad
An AI panel's breadth is what it's for. Criticizing it for being broad is like criticizing a truck for being large. The right question is never "is this broad?" — it's "is this broad in a way I understood and chose?"
Which is precisely what the August 1 disclosure rule is trying to guarantee: not that software is narrow, but that you can see how broad it is before you install it.
So what should actually live in your side panel?
The answer is probably "some of each," and the sorting rule is simple.
Rent understanding. Own keeping.
Understanding is a service. Summarizing, comparing, reasoning across tabs, drafting — that's model work, it happens on servers, it improves monthly, and renting it is correct. You don't want to own that.
Keeping is a possession. The page you'll need in eight months. The dashboard before your access is revoked. The receipt, the record, the source. That should be a file, on your disk, in a format that opens anywhere, with no service in the path.
Atlas's shutdown is the proof, and it's four weeks old. The understanding survived — it's moving into a ChatGPT Chrome extension and a cloud browser. The container people kept things in is what stops working on August 9. If your setup had rented the understanding and owned the keeping, that's an inconvenience. If you'd let the container hold both, it's a scramble.
What this looks like in practice
Use the AI panel to figure out which of the twelve tabs matters. Use a converter to make sure the one that matters is still readable in 2029.
Convert: Web to PDF makes real PDFs — selectable text, clickable links, embedded fonts, straight out of Chrome's own print engine. Not a screenshot. It works on login-protected pages because it uses the session you already have, which is the category most worth saving and the one cloud tools handle worst. Article Mode strips a page to its main content, with one honest caveat: it can drop <pre> code blocks, so use the default mode on code-heavy pages. Remove Elements deletes a cookie banner or sticky header before export, with undo. On infinite scroll, we capture what's already loaded — scroll first, and prefer single items over feeds.
Free. No watermark. No account. No file limits. Zero network requests during conversion — the only transmission is one anonymous post-conversion ping carrying a random install token, with no URL, no content, and no IP. We say that out loud because a policy about prominent disclosure is a bad week to get shy.
The small-tool side panel
Worth admitting: our own shop ships a side-panel tool. ScrapeMaster lives in the Chrome Side Panel API and extracts structured data from pages. It's narrow — one job, data stored locally in IndexedDB, exports to CSV/XLSX/JSON. It's also not perfectly zero-network, and we'll say so: during auto-detect, the page's HTML structure — not its content — goes to an analysis API to suggest selectors. Extracted data is never uploaded. That's the disclosure, stated the same way we'd want it from anyone. If you want the comparison against agentic browsers, it's here.
Same philosophy elsewhere in the shop: CineMan AI puts IMDb and Rotten Tomatoes ratings onto Netflix, Prime Video and Disney+ cards. It doesn't want to be your entertainment operating system. It wants to put a number on a card so you stop opening four tabs to decide on a Tuesday night. Small tool, one job, done — which is the entire pattern, and why we ship four extensions instead of one suite.
Frequently asked questions
Is the ChatGPT Chrome extension replacing Atlas?
Partly. Atlas stops working August 9, 2026, and its agentic features are being redistributed across three things: a ChatGPT Chrome extension (page context, summarize, ask, launch longer tasks), an upgraded desktop app with a built-in browser (login, downloads, page interaction), and a server-side cloud browser for agent tasks. OpenAI's framing is that "the browser is a feature, not the destination."
What did Gemini in Chrome actually add on July 14, 2026?
It expanded to UK desktop users, with iOS "next month." The capabilities include multi-tab comparison, Calendar/Maps/Gmail/YouTube integration, memory of past conversations, Nano Banana 2 image editing, and stated prompt-injection defenses.
Are AI side panels unsafe?
That's the wrong question, and we'd rather answer the right one. They're broad — by design, because breadth is what makes them useful. Breadth means more surface, including prompt injection, which is why Google shipping stated defenses is a real answer to a real problem rather than marketing. The question worth asking isn't "is it safe?" but "is it broad in a way I understood and chose?" The August 1 disclosure rule exists to make that answerable before you install.
Why does "strictly necessary to a single purpose" favour small tools?
Because the test measures data against a stated purpose, and the narrower the purpose, the shorter the argument. "Convert this page to a PDF" needs this page — full stop. "Help with anything on the web" makes almost anything arguably necessary, which is exactly the elasticity the rule is designed to squeeze out.
Can I run an AI panel and a converter at the same time?
Yes, and that's the setup we'd actually recommend. They don't overlap. Rent the understanding from the AI panel; own the keeping with a local file. Use the panel to find the tab that matters and the converter to make sure it's still there in three years.
Does Convert: Web to PDF use AI?
No. It's Chrome's print engine, run locally. Article Mode uses the Readability algorithm to strip a page to its main content, which is a content-extraction algorithm, not a model — no inference, no server, no page content leaving your machine.
What happens to my PDFs if any of this software shuts down?
Nothing. A PDF is a file on your disk with no dependency on the tool that made it. That's the structural argument, and it applies to us too: if we shut down, our manifesto commits us to publishing the source and walking away. Your files never needed us in the first place.
Where can I get more detail on the converter?
The 117-question FAQ covers paper sizes, Article Mode caveats, login-protected pages, infinite scroll behaviour, and the rest of the awkward questions.
Bottom line
Three things want your side panel now, and only one of them is trying to be everything.
ChatGPT and Gemini are broad by design, and that breadth is exactly what makes them useful. Multi-tab reasoning is real. Cross-service integration is real. Prompt-injection defenses are real engineering against a real problem, and Google deserves credit for shipping them rather than a punchline. If you want an assistant, get one — we don't compete with a language model and won't pretend to.
Small tools are narrow by design, and that narrowness is what makes them durable. When Chrome Web Store enforcement starts on August 1, 2026, "strictly necessary to a single purpose" will be a one-sentence answer for anything that does one thing, and an essay for anything that doesn't. Not because breadth is dishonest — because breadth is genuinely harder to bound.
The sorting rule that survives the whole news cycle: rent understanding, own keeping. Atlas proved it four weeks ago — the understanding moved on, and the container is what stopped working.
Convert: Web to PDF is the "own keeping" half, and it's deliberately unambitious. Real PDFs with selectable text, clickable links and embedded fonts, straight from Chrome's own print engine. Local, with zero network requests during conversion. Works on login-protected pages using the session you already have. Free, no watermark, no account, no limits.
It will never read your calendar, never remember last Tuesday, and never want your side panel for anything except the four seconds it takes to hand you a file.
That's not a limitation. That's the spec.