---
name: stat-check
description: Checks every number in marketing, advertising, influencer, social, email, SEO or ad content against GroundedKit's database of primary-source-checked statistics (verified, misquoted, outdated, unsourced) and benchmark reference facts, then rewords risky stats with their real source and date. Use when writing or reviewing a blog post, deck, pitch, ad, caption, newsletter, report or client email that contains percentages, ROI figures, dollar amounts, user counts or benchmarks, or when someone asks "is this stat real?" or "where does this number come from?".
license: CC-BY-4.0
metadata:
  version: "1.0.0"
  homepage: https://groundedkit.com/agents/
---

# Stat check (GroundedKit)

Marketing copy is full of numbers that travel from blog to blog until nobody remembers what was actually measured: "92% trust influencers" (really: trust in friends and family, Nielsen 2012), "email ROI 36:1" (a 2020 self-reported survey average), "attention span of a goldfish" (no study behind it). This skill describes how to check each number in a draft against [GroundedKit](https://groundedkit.com), a free, read-only database where every entry is traced to the organization that collected the data, and how to word what you find.

A GroundedKit lookup costs a few hundred tokens and returns the verdict, the primary source, what the source actually says and a safe rephrasing. Researching the same stat from scratch usually means several searches and long PDFs.

## When this applies

- A draft (post, deck, ad, script, newsletter, proposal, report) contains a statistic: a percentage, a multiplier (5x, 36:1), a dollar figure, a count of users, or a benchmark such as an engagement rate, CPM, CTR, open rate or influencer price.
- Someone asks whether a number is real, where it comes from, or for "a stat about X".
- You are about to add a number from memory. Memory is a common source of misquoted marketing stats.

Numbers that are the user's own data (their sales, their campaign results) are not checked here.

## Tools

**MCP (preferred).** Server URL `https://groundedkit.com/mcp` (Streamable HTTP, no key). Tools:

| Tool | Use it for |
|---|---|
| `verify_text(text)` | A whole draft, up to 20,000 characters. Returns one row per sentence that contains a number: `matched` (verdict applies), `possible` (same topic, wording or numbers differ) or `not_in_database`. |
| `check_stat(text)` | One statistic or one question such as "what share of US adults use TikTok". |
| `search_claims(query, verdict?)` | Finding a sourced number on a topic ("influencer pricing", "email open rate"). |
| `get_claim(id)` | The full record: method, sample, drift, every source opened. |
| `list_verdicts()` | Definitions and counts. |

Install: Claude Code `claude mcp add --transport http groundedkit https://groundedkit.com/mcp`; Claude web/desktop: Settings > Connectors > Add custom connector; Cursor/VS Code: `{"mcpServers": {"groundedkit": {"type": "http", "url": "https://groundedkit.com/mcp"}}}`.

**No MCP? Use the JSON API.** `GET https://groundedkit.com/api/claims.json` returns every entry (fields `claim`, `commonly_cited_as`, `verdict`, `quote`, `safe_rephrasing`, `sources`, `as_of`, `last_checked`, `stale`, `url`). Match a draft's numbers against `claim` and `commonly_cited_as`. One entry: `https://groundedkit.com/api/claims/<id>.json`.

## Workflow

1. **Collect the numbers.** For a draft, call `verify_text` once with the whole text. For a single stat, call `check_stat`.
2. **Read each result by status and verdict** (table below). Treat `possible` as unconfirmed: compare the claim text and numbers with the draft before applying the verdict.
3. **Check freshness.** `as_of` is the period the number describes; `stale: true` means GroundedKit last opened the source more than 180 days ago. For fast-moving numbers (platform users, ad spend, prices) say how old the figure is.
4. **Rewrite.** Replace risky stats with the `safe_rephrasing` (or your own wording that keeps the same facts), and cite the primary source plus the GroundedKit claim `url`.
5. **Report back** with a short list: each number, its status, what changed, and the link. Leave the user's own voice alone except for the stats.

## What each verdict means and how to word it

| Verdict | Meaning | Typical wording |
|---|---|---|
| `verified` | Primary source found; it says the same thing, scope matches, reasonably current. | Keep it, add the source and year: "According to Pew Research Center (2025), 37% of US adults use TikTok." |
| `misquoted` | Source exists but the quote changes the subject, metric, population or precision. | Use what the source measured: "In Nielsen's 2012 global survey, 92% said they trust recommendations from friends and family" instead of "92% trust influencers". |
| `outdated` | Quoted correctly, but old. | Give the year, or swap in the newer figure the record points to: "Litmus's 2020 survey put self-reported email ROI at about $36 per $1". |
| `unsourced` | No primary data found. | Remove it, or make the point without a number. Do not attribute it to anyone. |
| `not_in_database` | GroundedKit has no entry. This is not a verdict. | Say it was not checked; find the primary source (the organization that collected the data, with method, sample and date) or soften the claim. |

Reference facts (`kind: reference_fact`) are frequently needed numbers taken directly from a primary source (Pew, U.S. Census Bureau, IAB/PwC, FTC rules, platform policy pages). Quote them with their `as_of` date.

## Example

Draft: "Influencer marketing returns $5.78 for every $1, and 92% of consumers trust influencers."

`verify_text` returns two `matched` rows, both `misquoted`. A good rewrite:

> Influencer Marketing Hub's 2020 benchmark report estimated $5.78 in *earned media value* (an estimate of equivalent ad cost, not revenue) per $1 spent ([GroundedKit](https://groundedkit.com/claims/influencer-marketing-roi-5-78/)). Nielsen's 2012 global survey found 92% trust recommendations from people they know ([GroundedKit](https://groundedkit.com/claims/92-percent-trust-word-of-mouth-nielsen/)).

And a summary for the user: "2 stats checked: both misquoted. Reworded to what the sources measured; links added."

## Notes

- A verdict describes how a number is sourced and quoted, not whether the underlying idea is true.
- Results cite primary sources; open them yourself for anything high-stakes (legal claims, ads with substantiation requirements, investor material).
- Statistic text that matches nothing is logged anonymously (date and text only) so GroundedKit can add the most-requested stats. See https://groundedkit.com/privacy/#mcp.
- Wrong verdict or a stat to add: hello@groundedkit.com. Data license: CC BY 4.0.