Tony Wang9 min readThe State of AI Browser Extensions 2026: The Labs' Own Extensions Are the Worst-Rated
We measured 39,241 AI Chrome extensions. The AI labs' own — Claude, ChatGPT — rate far below third-party aggregators like Sider and Monica. The data.
If you searched the Chrome Web Store for "AI" today you'd get tens of thousands of results, and the loudest brands in the category — the frontier labs — would be nearly invisible in them. We pulled the whole AI-extension shelf to see what's actually installed, what it's rated, and what it can see. The headline is counterintuitive: the companies that build the models build the worst-reviewed extensions, and most of them don't build one at all.
The shelf: 39,241 AI extensions, and what happens as they grow
A keyword pull for "AI" returns 39,241 extensions. Filter to the ones with real reach — over 100,000 users — and you're left with 215. That thinning is the first finding: the AI-extension category is a very long tail with a small head, the same shape we found in the AI app graveyard.
The second finding is what changes as you climb toward the head. Extensions get more invasive as they get more popular.
| Metric | All 39,241 AI extensions | The 215 with 100k+ users |
|---|---|---|
Request broad host access (<all_urls>) | 28.2% | 63.3% |
| Declare data collection | 54.6% | 63.3% |
The most-installed AI extensions are roughly twice as likely to read and change data on every site you visit as a random one from the tail. That isn't automatically sinister — a browser assistant that acts on your behalf genuinely needs page access — but it means the permission cost of the category concentrates exactly where the users are.
The ranking: who's actually installed
Here's the head of the shelf — the most-installed AI extensions, with their ratings and permissions. Writing assistants, PDF tools, and translators dominate the very top; the frontier-lab assistants show up lower and rate worse.
| # | Extension | Developer | Users | Rating | Ratings |
|---|---|---|---|---|---|
| 1 | Grammarly | Grammarly | 35,000,000 | 4.50 | 43,097 |
| 2 | Claude | Anthropic | 11,000,000 | 2.71 | 1,306 |
| 3 | WPS PDF (AI Chat) | www.wps.com | 7,000,000 | 4.49 | 183 |
| 4 | Zoom | zoom.us | 7,000,000 | 3.88 | 1,032 |
| 5 | AnyDoc Translator | www.wps.com | 6,000,000 | 4.64 | 504 |
| 6 | Sider (all-AI sidebar) | ChatGPT Sidebar | 5,000,000 | 4.92 | 113,374 |
| 7 | Quillbot | quillbot | 5,000,000 | 4.70 | 5,583 |
| 8 | DeepL | team-browser-extensions | 4,000,000 | 4.73 | 12,505 |
| 10 | vidIQ | vidIQ | 3,000,000 | 4.49 | 10,961 |
| 11 | Monica (all-in-one AI) | Monica Team | 3,000,000 | 4.89 | 32,114 |
Claude sits at number two by installs and dead last by rating. It's the second-most-installed AI extension in the store and the lowest-rated one above a million users. The two third-party AI sidebars near it — Sider and Monica — rate a full two stars higher.
The split: labs vs. aggregators
That contrast isn't a Claude quirk. Line up every first-party lab extension against the biggest third-party AI assistants and the two groups fall into completely separate bands.
Anthropic · first-party
OpenAI · first-party
OpenAI · first-party
first-party
third-party
third-party
third-party
third-party
| Group | Average rating | Range |
|---|---|---|
| First-party lab extensions (5) | 3.45 | 2.71 – 3.83 |
| Frontier-lab chat/agent only (Claude + both ChatGPT) | 3.26 | 2.71 – 3.74 |
| Third-party AI assistants (Sider, Monica, Merlin, MaxAI, HARPA) | 4.81 | 4.69 – 4.92 |
The worst-to-best single comparison — Claude 2.71 against Sider 4.92 — is a 2.21-star gap between two products that both clear the top install tier. And there is no overlap between the distributions: the highest-rated first-party lab extension (Perplexity, 3.83) still sits below the lowest-rated third-party aggregator (HARPA, 4.69).
Why would the specialists lose to the generalists on their home turf? The store reviews point at the same thing our earlier studies found: the complaints are about auth loops, forced changes, permission creep, and how fast the agent burns tokens — not about the model. A third-party aggregator like Sider is a thin, mature wrapper that lets you pick any model and mostly stays out of the way; a first-party browser agent is newer, heavier, wired into a subscription and a quota, and it carries the friction of being the thing that enforces the plan. We took that argument apart in detail in the AI coding usage-limits piece and the Claude app-review study.
The dog that didn't bark: most labs ship nothing
The more surprising half of the picture is who isn't here at all. Of the major AI labs, only three ship a verified first-party Chrome extension:
| Lab | First-party extension? | Best available |
|---|---|---|
| OpenAI | Yes | ChatGPT search (4M) + ChatGPT browser agent (2M) |
| Anthropic | Yes | Claude (11M) |
| Perplexity | Yes | AI Companion (400k) + AI Search (100k) |
| Google (Gemini) | No | — |
| Microsoft (Copilot) | No | No consumer Copilot sidebar; closest is Power Automate |
| xAI (Grok) | No | — |
| Mistral | No | — |
| DeepSeek | No | Only impostor extensions on a spoofed developer name |
Searching the store for "Google" or "DeepSeek" AI extensions turns up utilities and outright impostors — third parties publishing under a spoofed developer name — not the labs themselves. We verified the three genuine first-party publishers by developer email domain (@anthropic.com, @openai.com, @perplexity.ai) and by known extension IDs, and treated everything else as unverified. So the true state of the category is narrower than the marketing suggests: three labs ship a browser extension, and all three underperform the independent tools built on top of their own models.
What the popular ones can see
Because permissions concentrate at the top, it's worth naming what the most-installed lab extensions actually request. Among the top 30 AI extensions by installs, 24 request broad host access and 25 declare data collection. The two frontier-lab browser agents are firmly in that group:
- Claude (Anthropic) requests
<all_urls>host access and declares collection of personally identifiable information, personal communications, location, web history, user activity, and website content. Its permission list includesdebugger,nativeMessaging,downloads, andidentity. - OpenAI's ChatGPT browser-agent extension declares the heaviest profile in the entire set: on top of the categories above, it lists health information, financial and payment information, and authentication information, with permissions including
debugger,bookmarks,history,topSites, andsessions.
Both are opt-in and both ask before sensitive actions, and an agent that drives your browser structurally needs this reach. But it's a real trade: the extensions doing the most on your behalf are also the ones declaring access to the most sensitive slice of your browsing. By contrast, OpenAI's plain ChatGPT search extension and both Perplexity extensions are scoped and declare no data collection — so the invasive profile tracks the browser agents, not lab extensions as a class.
The honest caveat: thin review counts
One number deserves a warning label. These ratings rest on very different sample sizes. Claude has 11,000,000 installs but only 1,306 ratings; OpenAI's browser agent has 2,000,000 installs and 327. Sider, by contrast, has 113,374 ratings and Monica 32,114. A harsh score built on a small, self-selected, mostly early-adopter sample is not the same statistical object as a 4.9 built on a hundred thousand reviews. The direction of the finding is robust — every lab extension trails every aggregator, across five independent products — but read the lab scores as an early, loud signal from the users who bothered to review, not as a settled verdict from the full install base.
The other caveats are the usual ones: user counts are Chrome's bucketed values (so extensions inside a bucket are tied), and the 39,241 "AI" universe is a keyword match that includes writing, PDF, translation, and scraping tools, so the clean apples-to-apples comparison is the lab-vs-aggregator cut, not the raw shelf.
The takeaway
The AI browser-extension shelf in 2026 is not what the model-quality headlines would predict. The category is a long tail with a small, permission-heavy head; most of the famous labs aren't on it; and the three that are ship the lowest-rated assistants in it, all beaten by independent wrappers built on the same models. If you're choosing an AI sidebar, the store is telling you — loudly, and against brand intuition — to look past the logo on the model.
Sources
Methodology
Every number here was pulled through Crawlora's own Chrome Web Store dataset endpoints — the datasets/chrome-extensions search and facet APIs for the universe counts, rankings, permissions, and data-collection flags, and the item details endpoint to confirm host permissions and declared data categories on the lab extensions. User counts are Chrome's bucketed install values and ratings are the store's mean star scores as of the July 14, 2026 crawl. First-party publishers were verified by developer email domain and known extension IDs; spoofed and impostor developer names were excluded. The aggregate dataset behind this post is reproducible from the same public endpoints — test them in the Playground; pricing covers the free tier.
For more of this series, see how to choose a web scraping API and web scraping with AI.
Frequently asked questions
What is the best-rated AI Chrome extension in 2026?
Among AI extensions with over a million users, the highest-rated are third-party AI aggregators, not the labs' own products: Sider (5M users, 4.92 stars) and Monica (3M, 4.89) lead, followed by Merlin (4.79), MaxAI (4.75), and HARPA (4.69). Every first-party lab extension — Claude, ChatGPT, Perplexity — rates below all of them.
Why is the Claude Chrome extension rated so low?
Anthropic's Claude extension has 11,000,000 installs but a 2.71-star rating — the lowest of any AI extension above roughly a million users. Its recent reviews complain about authentication and account-switching loops, permission creep, and how quickly the browser agent burns through usage limits, rather than about the model's quality. Note the sample is thin: the score rests on 1,306 ratings against 11M installs, so it reflects loud early-adopter feedback more than the full install base.
Does Google Gemini or Microsoft Copilot have a Chrome extension?
Not a verified first-party one, as of the July 2026 crawl. Only OpenAI, Anthropic, and Perplexity ship a genuine first-party AI browser extension. Searches for Google (Gemini), Microsoft's consumer Copilot sidebar, xAI (Grok), Mistral, and DeepSeek return either unrelated utilities or impostor extensions published under spoofed developer names — not the labs themselves.
What data do AI browser extensions collect?
It scales with popularity. Broad host access (reading and changing data on every site) rises from 28% of all AI extensions to 63% of those with 100k+ users; among the top 30 by installs, 80% request broad host access and 83% declare data collection. OpenAI's ChatGPT browser-agent extension declares the heaviest profile in the set — including health, financial, and authentication data — while its plain search extension and Perplexity's extensions are scoped and declare no collection.
How many AI extensions are in the Chrome Web Store?
A keyword search for "AI" returns 39,241 extensions, but only 215 have more than 100,000 users — a long tail with a small head. The category is dominated by writing assistants, PDF tools, and translators (Grammarly alone has 35M installs); the frontier-lab chat assistants are a small, lower-rated slice of it.