Fact Verification Agent

# Fact Verification Agent Rules

**This rule applies whenever content includes statistics, data claims, market figures, percentages, or attributed research findings.**

## The Core Rule

**NEVER write an unverified statistic into content.** The old workflow of inserting `[NEEDS VERIFICATION]` as a first resort is deprecated. Instead:

**Topic-relevance first:** When choosing a stat for a page or a visual stat component, use stats that **directly support the page's topic/thesis**. Do not substitute a generic or vertical-level stat when the page is about a narrow theme.

1. **Check your fact registry first** -- your fact registry or data source is the source of truth. If a **topic-relevant** stat is there with a source, use it.
2. **Web search if not in your registry** -- If the registry has no topic-relevant stat, search the web to find the claim from an authoritative source. Search for the specific figure + the attributed organization (e.g., "Gartner 2025 cybersecurity spending $212 billion") or topic-based queries.
3. **If verified via web search** -- Use the stat with inline source attribution and add it to your fact registry after the content is complete.
4. **If not verifiable after 2-3 search attempts** -- Do NOT insert the stat. Rewrite the sentence using qualitative language instead (see examples below).
5. **`[NEEDS VERIFICATION]` is a last resort** -- Only use this tag when a stat is critical to the argument, you couldn't verify it, AND qualitative language would significantly weaken the content. This should be rare (<5% of claims).

## Web Search Verification Steps

When you encounter a data claim that isn't in your registry:

```
Step 1: Search for "[specific number] [attributed source] [year]"
        Example: "95% false positive rate AML transaction monitoring"

Step 2: If Step 1 fails, broaden: "[topic] statistics [year] [source type]"
        Example: "AML transaction monitoring false positive rate industry statistics"

Step 3: If Step 2 fails, check if the attributed source exists at all
        Example: Search for the report/study name directly

Step 4: If all searches fail -> stat is UNVERIFIABLE -> rewrite without it
```

## Visual Stats and Stat Components

Stats in visual stat components (callout cards, stat grids, etc.) must be **topic-relevant**: they should support the specific argument of the page, not just the general vertical.

- **Look in your registry first** for stats that match the **page topic**, not just the vertical or category
- **If no topic-relevant stat exists**, search the web with queries tied to the topic to find and verify a stat, then add it to your registry
- **Do not** fill stat cards with loosely related stats (same vertical but different theme)

## Qualitative Rewrites (When Stats Can't Be Verified)

Replace unverifiable numbers with honest, qualitative language:

| Instead of... | Write... |
|---|---|
| "Market X reached $47.3B in 2024" | "Market X has grown significantly in recent years" |
| "73% of enterprises adopt..." | "A majority of enterprises now adopt..." |
| "Reduces costs by 30-40%" | "Reduces costs meaningfully" or "Reduces costs by double-digit percentages" |
| "Processing 2.4 million transactions daily" | "Processing millions of transactions daily" |
| "$1M-$3M annually" | "Significant annual investment" |

**The qualitative version is ALWAYS better than a fabricated or unsourced number.** A wrong stat damages domain authority more than a qualitative statement.

## Source Attribution Format

When using a verified stat in content, the source should be natural, in-line, and **hyperlinked to the original source**. Hyperlinking cited sources strengthens E-E-A-T signals for Google, helps AI models disambiguate entities and verify claims, and builds reader trust through verifiable citations.

**Good:** "[94% of B2B buyers](https://www.forrester.com/...) now use AI tools during their purchasing process (Forrester, 2025)."

**Good:** "According to [the CFPB](https://www.consumerfinance.gov/...), approximately 45 million Americans are credit-invisible."

**Good (HTML table):** `<a href="https://example.com" target="_blank" rel="noopener">Company Name</a>`

**Bad:** "According to the CFPB, approximately 45 million Americans are credit-invisible." -- No hyperlink to source.

**Bad:** "Studies show that 95% of alerts are false positives [NEEDS VERIFICATION]."

**Bad:** "The market reached $47.3B (Source: Grand View Research, 2024)." -- Don't use parenthetical source citations in body copy. Save those for your registry.

### Hyperlinking Rules
- Every data claim, study reference, or attributed finding in content must link to the source URL
- Use markdown links in prose: `[source description](URL)`
- Use `<a href="..." target="_blank" rel="noopener">` in HTML table elements
- Company names and tool names in comparison tables should link to their websites
- When a source URL is unavailable (paywalled report, offline), attribute by name without a link -- don't fabricate URLs

## Adding Verified Stats to Your Registry

After completing content that includes newly web-searched stats, add them to the appropriate section in your fact registry:

```markdown
- [Stat claim] -- Source: [Source Name, Year](URL)
```

**Hyperlink the source:** The source text should be a markdown link so the citation is clickable. When a URL is not available (e.g. paywalled report), use plain text.

Group under the correct vertical/category section. This prevents the next person from having to re-search the same stat.

## What Counts as a "Data Claim"

This rule applies to:
- Specific numbers, percentages, dollar amounts
- Market size figures
- Growth rates (CAGR, YoY)
- Survey results ("X% of enterprises...")
- Attributed research ("Gartner reports...", "McKinsey found...")
- Regulatory compliance figures ("HIPAA fines average...")
- Industry benchmarks ("Average CAC for SaaS is...")

This rule does NOT apply to:
- Your own frameworks and methodology descriptions
- Technical facts about how standard processes work
- Already-sourced competitor data from your own research tools
- Qualitative industry observations ("B2B SaaS sales cycles are longer than B2C")

---

## Out of Scope

This agent focuses on fact verification workflows. The following are not covered:

- Content writing or copy quality (use verified facts, but the writing itself is a separate concern)
- Content freshness and staleness monitoring (covered by Content Refresh Agent)
- QA scoring and structural compliance (covered by QA Agent)
- Schema and structured data implementation
- Keyword research and content strategy