CiteSentry

A browser extension that sidecars every AI-search answer — Perplexity, ChatGPT search, Google AI Overviews, Gemini, Claude — and silently checks each footnote: does the source page exist, and does it actually say what the AI claimed it said?

Problem

AI search engines now answer the majority of long-tail queries with cited footnotes, but the citations are often wrong. Recent CJR/Tow Center testing pegs Perplexity's citation hallucination rate at roughly 37% — a real URL pointing at a real publication, with a paraphrased "fact" that the source never makes. Most users skim the answer, see five blue source chips, and trust it. There is no in-page signal that a citation is broken, paywalled, off-topic, or contradicted by the page it links to.

Target user

Knowledge workers who run >10 AI search queries a day — analysts, journalists, paralegals, consultants, founders doing competitive research. They already paste Perplexity answers into Notion and Slack; the job-to-be-done is trust before paste. Today's options are (a) open every citation by hand or (b) hope. CiteSentry sits in the AI answer panel and turns each citation chip green/yellow/red in real time.

MVP scope

• Manifest V3 extension that detects AI-answer DOM regions on Perplexity, ChatGPT search, Google AI Overviews, Gemini, and Claude.
• For each citation, fetch the cited URL through a CORS-safe proxy and extract main-content text.
• Run a local embedding similarity + sentence-pair entailment check between the AI's claim sentence and the page's text; score 0–100.
• Inline color badge on each citation chip: green (supports), yellow (loose match), red (unsupported), gray (404 / paywall / blocked).
• Hover-card showing the exact paragraph in the source page that best matches — or a "no supporting passage found" message.
• One-click "Copy with citations stripped of red flags" button for pasting into a doc.

Monetization

Freemium. 25 free citation audits per day on a single device. CiteSentry Pro at $6/month for unlimited audits, multi-device sync, and the "Copy clean" feature. A Teams tier at $9/seat/month adds an audit history dashboard for newsrooms, research firms, and compliance teams that need to show a paper trail of what was verified before publishing.

Why now

CJR/Tow Center and follow-on benchmarks published through early 2026 show citation hallucination rates between 11% and 37% across the major AI search products, even as those products eat into traditional search share. Google AI Overviews shipped sitewide by mid-2025, ChatGPT search left preview in late 2025, and Perplexity passed a billion queries per month. The behavior shift — "answer first, source maybe" — outran the trust infrastructure, and there is no in-browser primitive for verifying a citation without manually clicking it.

Risks & open questions

• Demand-side: do paid users exist outside power-user research roles, or does the casual majority simply accept hallucinations as cost-of-doing-business?
• Build-side: paywalled and JS-rendered sources (NYT, WSJ, JSTOR) are unreadable without a logged-in fetch; the extension must degrade gracefully rather than mark them red.
• Entailment is hard — false negatives ("page does support the claim, but in a table the parser dropped") will erode trust quickly; need a tight calibration on the entailment threshold.
• Platform risk: Perplexity and OpenAI can change their answer DOM weekly, breaking selectors. Need a resilient adapter pattern and a fast OTA selector update channel.
• Legal: scraping cited pages at scale on behalf of users may bump against robots.txt and individual sites' terms; structure as user-initiated, single-page fetches, not crawl.

Next step

Build a 24-hour proof-of-concept against Perplexity only: detect citation chips, fetch the source, run an off-the-shelf entailment model (e.g. a small fine-tuned NLI), and prove the green/yellow/red badge UX. If accuracy on a 100-citation gold set lands above 75%, promote to weekly prototype.

Sources

• https://suprmind.ai/hub/ai-hallucination-rates-and-benchmarks/ — 2026 hallucination-rate benchmarks across major AI products
• https://www.leapd.ai/blog/ai-visibility/how-chatgpt-google-ai-overviews-and-perplexity-source-information-in-2026 — how each AI search product currently surfaces and selects citations
• https://www.pravinkumar.co/blog/perplexity-chatgpt-google-ai-mode-citation-differences-2026 — comparative analysis of citation overlap and source choice across AI search engines in 2026