Google’s AI cites its sources. We read them.
Groundmark is a free Chrome extension that checks every sentence of an AI Overview against the page it cites, right on your computer, and shows you the exact passage behind each verdict.
Up to 400 milligrams of caffeine a day is considered safe for most adults. That's about the amount in eight cups of brewed coffee. Children and teens should keep caffeine under 100 milligrams a day. Energy drinks often pack more caffeine than their labels suggest.
Illustrative example. The quotes are verbatim from the cited pages.
A citation isn’t the same as support.
AI Overviews sit above every result and link to sources that look like proof. Independent studies keep finding that the sources often don’t say what the summary claims.
52%
of AI Overviews had at least one claim their cited sources didn't support.
Only 39% were both correct and fully supported by the pages they linked.
1 in 9
claims in AI Overviews weren't backed by the source Google attributed them to.
11% of 98,020 claims across 7,583 Overviews; 4.1% were directly contradicted by their source.
8%
of visits with an AI summary ended in a click on a search result.
Without a summary, people clicked a result in 15% of visits. Fewer clicks means fewer people checking.
From sentence to source in seconds.
Most people never click through to check an AI answer. Groundmark does the clicking, the reading, and the comparing for you.
Step 1
It finds every sentence.
When an AI Overview finishes loading, the extension splits it into individual claims and notes which links Google attached to each one.
AI Overview
- Up to 400 milligrams of caffeine a day is considered safe for most adults.
- That's about the amount in eight cups of brewed coffee.
- Children and teens should keep caffeine under 100 milligrams a day.
- Energy drinks often pack more caffeine than their labels suggest.
4 sentences found, each with its own sources
Step 2
It opens the page Google cited.
The extension opens each cited page itself, without your cookies, and reads all of its text, skipping menus, buttons, and scripts.
AI Overview
- Up to 400 milligrams of caffeine a day is considered safe for most adults.
- That's about the amount in eight cups of brewed coffee.
- Children and teens should keep caffeine under 100 milligrams a day.
- Energy drinks often pack more caffeine than their labels suggest.
4 sentences found, each with its own sources
Caffeine: How much is too much?
Step 3
It finds the passages that matter.
Each page is broken into short passages and ranked against the claim, with extra weight on matching numbers, where errors tend to hide.
AI Overview
- Up to 400 milligrams of caffeine a day is considered safe for most adults.
- That's about the amount in eight cups of brewed coffee.
- Children and teens should keep caffeine under 100 milligrams a day.
- Energy drinks often pack more caffeine than their labels suggest.
4 sentences found, each with its own sources
Caffeine: How much is too much?
That's about the amount of caffeine in four cups of brewed coffee, 10 cans of cola or two "energy shot" drinks.
Step 4
It shows you the receipt.
Gemini Nano, the AI model built into Chrome, weighs the claim against the best passages on your own computer. The quote you see is checked word for word against the page, and links straight to it.
AI Overview
- Up to 400 milligrams of caffeine a day is considered safe for most adults.
- That's about the amount in eight cups of brewed coffee.
- Children and teens should keep caffeine under 100 milligrams a day.
- Energy drinks often pack more caffeine than their labels suggest.
4 sentences found, each with its own sources
Caffeine: How much is too much?
That's about the amount of caffeine in four cups of brewed coffee, 10 cans of cola or two "energy shot" drinks.
The page says 400 mg is about four cups of brewed coffee, not eight.
That's about the amount of caffeine in four cups of brewed coffee, 10 cans of cola or two "energy shot" drinks.
Open the source at this passage
Four verdicts. No guessing.
Groundmark judges one thing: whether the page Google cited backs the sentence. It doesn’t decide what’s true in the world, and it says so when it can’t tell.
- Supported
The cited page says it, or it follows directly. Every number, name, and qualifier matches.
For example: “Up to 400 milligrams of caffeine a day is considered safe for most adults.” The FDA page states the 400 mg daily level is safe for most adults.
- Not found in its source
The cited page doesn't back the specifics. It may be true, but Google's own citation doesn't show it.
For example: “Children and teens should keep caffeine under 100 milligrams a day.” The page discusses caffeine and children but never gives a 100 mg limit.
- Contradicted by its source
The cited page says something that conflicts with the sentence.
For example: “That's about the amount in eight cups of brewed coffee.” The page says 400 mg is about four cups of brewed coffee, not eight.
- Couldn't check
The page was a PDF, blocked, paywalled, or missing, so no verdict is given.
For example: “Energy drinks often pack more caffeine than their labels suggest.” The cited source is a PDF, which Groundmark can't read yet.
Check a claim right here.
The same checker the extension uses, running in your browser. Paste a sentence and the source text that’s supposed to back it up.
Example sources are written for this demo.
A fact-checker should be checkable too.
A tool that grades AI has to hold itself to the same standard. These are the design choices that make Groundmark trustworthy, and the places where it isn’t. Read the full methodology, including the exact prompt.
The only thing that goes out
- A normal request to open each page Google cited, the same as clicking the link
- Sent without your cookies, so sites never see your logged-in accounts
Stays on your computer
- Your search query and the AI Overview's text
- Every verdict, quote, and explanation
- Your Google account, cookies, and history
Receipts are real text.
The model must quote the source word for word. Every quote is checked against the page, and if it doesn't match, the closest real sentence is shown instead and the confidence is lowered.
Pages can't give orders.
Web pages can hide instructions meant for AI. Source text is passed to the model as quoted data with an explicit rule to ignore anything inside it that reads like a command.
No server, no account, no bill.
Checking runs on Gemini Nano, the model built into Chrome, on your own computer. There's nothing to sign up for, and nobody, including us, sees what you search or what gets checked.
Access only when you allow it.
Reading cited pages needs permission to open websites, so the extension asks once, with an explanation, and still runs only on Google search pages. It never opens local-network addresses.
Honest about uncertainty.
"Not found" is not "false." PDFs, paywalls, and blocked pages get "Couldn't check," never a guess. Confidence comes from checks on the evidence, not from the model grading itself.
Works on more computers, honestly labeled.
Where Chrome's on-device model isn't available, a lite mode matches words and figures instead. It never claims a contradiction, and every lite result says so.
Read past the summary.
Free and open source, with nothing to sign up for. Requires Chrome 138 or later.
Step 1
Add it to Chrome
Install Groundmark from the Chrome Web Store. A setup page opens automatically.
Step 2
Finish setup
Allow it to read the pages Google cites, and download Chrome's on-device AI if your computer supports it.
Step 3
Search as usual
When an AI Overview appears, each sentence gets a verdict. Hover one to see the evidence.
Install without the Chrome Web Store
- Download the zip below and unzip it somewhere you'll keep it.
- Go to chrome://extensions and turn on Developer mode in the top right.
- Click Load unpacked and choose the unzipped folder.
The full source code is on GitHub.
What it can’t do
- It checks whether the cited page supports a sentence, not whether the sentence is true. A sentence can be correct and still not be backed by its citation.
- Verdicts come from a small on-device model reading short passages. It is less accurate than large cloud models, can miss context elsewhere on a page, and can be wrong. Treat a verdict as a reason to look, not a final answer.
- The on-device model needs a recent computer: about 22 GB of free disk space and either a GPU with more than 4 GB of memory or 16 GB of RAM. Other computers get lite mode, which is much less capable.
- It reads English Overviews on Google search only, and can't yet read PDFs, videos, or pages behind logins or paywalls.
- Google changes its page layout often. When the layout shifts, the extension may miss sentences or attach them to the wrong links until it's updated.