About this plugin
Two problems, one plugin.
1. Pasted chat sessions. When someone copies an answer out of ChatGPT or Claude and drops it into the editor, the reply usually comes with its packaging: a “Certainly!
” at the bottom. The pattern scanner looks for that packaging. It runs locally, costs nothing, and needs no API key.
Hits near the end of a long post are weighted more heavily, because that is where the tail of a chat reply lands.
2. Facts that have moved on. A model writes confidently from training data with a cutoff date.
Version numbers, prices, “the current largest”, “as of 2023” — all of it can be quietly wrong by the time it publishes. The plugin pulls the most checkable sentences out of a post and runs them past a search-grounded model. Anything contested or low-confidence gets a second look from Google.
Verdicts are one of: accurate, outdated, inaccurate, or unverifiable. “Outdated” is the one to care about — it means the claim used to be true, which is exactly the failure mode a knowledge cutoff produces.
Choosing a fact-check provider
You can use Perplexity, Gemini, or both.
Both are grounded: Perplexity searches at request time by default, and the Gemini requests always attach the Google Search tool. That grounding matters more than which model you pick. An ungrounded model answers a fact-check from its own training data, which means it carries the same cutoff blind spot the plugin exists to find — it would be checking stale facts against stale facts.
Running both costs twice as much but is meaningfully stronger: claims the two agree on are reported with higher confidence, and claims they split on are marked disputed so a human decides rather than the plugin guessing.
What it looks at
- Dangling offers to keep working (“Would you like me to…”, “Let me know if you’d like…”, “3 more prompts”)
- Assistant self-reference (“As an AI language model”, “I don’t have access to real-time data”)
- Prompt echoes and delivery framing (“Certainly! Here’s a draft”, “As requested”)
- Unfilled scaffolding ( [Insert company name], “Word count: 1,450”, stray SEO briefing fields)
- Raw markdown that never converted ( **bold**, ## headings )
- Knowledge-cutoff language (“as of my last update”, “the latest version is”)
- Stock generated phrasing (“delve into”, “in today’s fast-paced world”, “a testament to”)
The last category is deliberately low-weight. Plenty of human writers say “it’s important to note”. On its own it proves nothing, and the scoring reflects that.
External services
Pattern scanning is entirely local. The plugin makes no network requests at all until you add an API key and turn on AI mode.
When you do enable fact-checking, the plugin sends data to whichever services you have configured. In every case what leaves your site is the same: individual sentences extracted from the post being checked, plus the post title and its publication date for context. Never the full post, never user accounts, never visitor data.
Perplexity AI — used to fact-check claims when Perplexity is the selected provider.
Data sent: extracted claim sentences, post title, post publication date.
ai/chat/completions
ai/hub/legal/terms-of-service
ai/hub/legal/privacy-notice
Google Gemini API — used to fact-check claims when Gemini is the selected provider. Google Search grounding is enabled on these requests, so your claim text is used as a search input.
Data sent: extracted claim sentences, post title, post publication date.
com/
dev/gemini-api/terms
com/privacy
Note that Google treats its free and paid Gemini tiers differently. On the free tier, Google’s terms state that submitted content may be used to develop Google products and that human reviewers may read it. On a paid tier it is not used to improve their products.
If you are checking unpublished drafts, read those terms before using a free key.
Google Fact Check Tools API — optional. Looks for published fact-checks matching a claim.
Data sent: keywords derived from a single claim sentence.
com/
com/terms
com/privacy
Google Programmable Search (Custom Search JSON API) — optional, and only active if you supply a Search Engine ID. Used to corroborate contested claims and judge how recent the supporting sources are.
Data sent: keywords derived from a single claim sentence.
com/customsearch/v1
com/terms
com/privacy
Responses are cached for 24 hours so repeated checks of unchanged content do not re-send anything.