Receipts

Checks each claim in a Google AI Overview against the pages it cites, and shows you the sentence it found (or tells you it found nothing).

Date: 2026-10-05 Form factor: Browser extension (Manifest V3) Browser surface: content script on Google search result pages, toolbar popup, service worker (fetches cited pages), chrome.storage Status: Prototype

What it is

On a Google search that shows an AI Overview, you click the Receipts icon and then Check the overview's sources. Receipts splits the overview into individual claims, reads each page cited next to a claim, and finds the closest passage on that page. Each claim gets a badge in place: found, partial, differs (the passage covers the same thing with different numbers), or not found. The popup lists the flagged claims first, each with the quoted passage, so you can judge it yourself in a few seconds.

Receipts doesn't decide whether a claim is true. It answers a narrower question: does the page Google pointed to actually say this? Then it shows you the passage, so you can see the evidence it's working from.

Why it has to be an extension

The AI Overview only exists in your own search results page. It's generated for each query, can differ from one user or session to the next, and Google offers no API for it. A website can't see your search results. A server could scrape Google, but that breaks Google's terms and gets back a different overview from the one you're reading. A content script can read the overview as it rendered in your tab, including the citation links next to each paragraph. A service worker can then fetch those cited pages from your browser and compare them, which a web page can't do because cross-origin fetches are blocked without the extension's host permission.

Who it serves

People who use AI Overviews but can't afford to take one at its word: students and teachers, reporters and researchers, and anyone looking up a medical, financial, or home-repair figure before acting on it. Their problem is specific. The overview looks sourced, with a link chip right next to the sentence, and clicking through to check is slow enough that almost nobody does it. Receipts turns that click-through into a quoted passage you can read in about two seconds.

Why it could be profitable

Model. Freemium. The free tier gets 10 overview checks a day. Pro, at around $4/month, removes the cap, keeps a history of past checks, and would extend coverage to other AI answer surfaces with inline citations (Bing's Copilot answers, Perplexity, ChatGPT search). The stronger revenue line is probably site licenses for schools, libraries, and newsrooms. Those buyers already pay for research and verification tools. Whether they'd budget for one aimed at AI answers is an open question I can't answer from desk research.

Why now. On 2026-09-29, Washington University in St. Louis published results from an audit of 55,393 Google searches run between March 13 and April 21, 2026. Of 98,020 verifiable claims in AI Overviews, about 7% weren't supported by the pages Google cited and about 4% contradicted them. Only 41.9% of overviews were fully grounded. The researchers also found that source quality and claim fidelity were largely independent, so a trustworthy-looking citation doesn't mean the sentence next to it is faithful. That finding is the case for this extension. Separately, Google keeps adding more citation surfaces to AI Overviews (a "Further Exploration" section, subscription labels, inline link previews, announced in May 2026), so there are more links to check, not fewer.

Where the case is thin. Most searchers won't pay to double-check a search result, and the 11% problem rate means a casual user sees a flag on only one or two claims in a typical overview. Competition already exists. "AI Fact Checker" on the Chrome Web Store sends a highlighted selection to Gemini to verify it against the sources. Receipts takes a different approach: it checks the whole overview at once, runs locally without sending your searches to any model, and shows you a quote rather than a verdict. Whether that difference is enough to win paying users is untested. The biggest operational risk is that Google changes the overview's markup. The content script anchors on the visible "AI Overview" heading rather than class names, but it will still break sometimes and need fixes quickly.

How to load it in Chrome

1. Open chrome://extensions and turn on Developer mode.
2. Click Load unpacked and select this folder's extension/ directory.
3. Search Google for a question that usually gets an AI Overview (question-form searches trigger one far more often than single words). Click the Receipts icon, then Check the overview's sources. The first check asks for permission to read pages on other sites. That's how it fetches the cited pages. (The prototype ships without an icon, so Chrome shows its default badge in the toolbar.)

How to try the demo

1. Open index.html in any modern browser. The sample data is also inlined in the page, so it works from file:// with no server.
2. Click the receipt icon at the right of the mock toolbar.
3. Click Check the overview's sources. After a short pause, badges appear on the overview and the popup lists the results. Expand any claim to see the passage. Hover a badge on the page to see the same quote.

The search page and the five cited sites are invented, and their text comes from sample-data.json instead of being fetched over the network. Everything else runs the extension's own files: content.js finds the overview and extracts claims from the mock page, matcher.js does the matching, and popup.js and popup.css render the popup.

Permissions, and why each one

What's in this prototype

• A content script that finds the AI Overview by its visible heading, extracts sentence-level claims with the citation links next to each, and adds status badges in place (with the quoted passage as a tooltip). It guards against double injection, and all its class names start with rcpt-.
• A service worker that fetches each cited page once, strips it to text, and treats paywalled or script-rendered pages that come back nearly empty as "unread" rather than "not found".
• matcher.js, a lexical matcher shared by the service worker and the demo. It scores word overlap and checks whether every figure in the claim appears in the passage. When the passage has different figures, it reports differs.
• A popup with a summary bar, flagged claims listed first, an expandable quote for each claim, a daily free-tier counter, and Copy report (plain text for pasting into notes or an assignment).
• Stubbed: payment and the Pro plan, check history, and every search engine other than google.com.

The matcher is deliberately simple. It catches missing support and changed numbers well. It can't catch a claim that rewords its source while flipping the meaning (a negation, or "used to" vs. "still does"), so a found badge means "the words and numbers are there," not "the claim is right." The quote is there so you can check that yourself.

Roadmap

• Use Chrome's built-in on-device model (the Prompt API) to judge the passages the lexical matcher flags, so negations and reworded contradictions get caught without sending anything off the device.
• Support Google's regional domains and AI Mode, then Bing Copilot answers and Perplexity's inline citations.
• Check history with a per-domain view: which sites keep getting cited for things they don't say.
• A publisher-side version: alert a site owner when an AI Overview cites their page for a claim the page doesn't make. This would need a backend, and it's a separate paid product.
• A classroom mode that exports a check as a worksheet: the claim, the quote, and a blank for the student's verdict.

Sources

• https://source.washu.edu/2026/09/washu-research-finds-gaps-in-google-ai-overview-citations-claims/ — WashU audit (published 2026-09-29): 55,393 searches, 98,020 verifiable claims, 7% unsupported and 4% contradicted by cited sources, 41.9% of overviews fully grounded; to be presented at ACM IMC in October 2026.
• https://arxiv.org/html/2605.14021v1 — "Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact," the preprint behind the audit.
• https://thenextweb.com/news/google-ai-overviews-publisher-links-search-traffic — Google's May 6, 2026 AI Overview changes (Further Exploration links, subscription labels, inline link context) and the Ahrefs February 2026 finding of a 58% click-through drop for top-ranking pages.
• https://chromewebstore.google.com/detail/ai-fact-checker/ajlnhclklmceofjieakhamopoieiagmj — an existing extension that verifies a highlighted AI Overview claim with Gemini; the closest competitor.