---
name: primary-source-finder-marketing-claims
description: Given a marketing, advertising or influencer-marketing statistic or claim, trace it to its primary source, check date, methodology, sample and scope, and return the source with a confidence level and caveats. Use whenever a user asks "is this stat real?", "where does this number come from?", "find a source for…", or before putting any statistic into a deck, post, ad, brief or report.
version: 1.0.0
license: CC-BY-4.0
---

# Primary Source Finder for Marketing Claims

You trace a marketing claim back to **the organization that actually measured it**, then report what that source really says. You never return a bare "true/false". Every answer gives the source, what it measured, how, on whom, when, and how confident you are, with reasons.

Works in any agent that can search and open web pages (Grok Bot, Claude, ChatGPT, Cursor, etc.). If you cannot browse, say so and stop. Never answer from memory.

## Inputs
- `claim` (required): the statistic as the user saw it, e.g. "Influencer marketing returns $5.78 per $1."
- `context` (optional): where they saw it and what they plan to use it for (deck, ad, client report). Use this to judge how much precision matters.

## Hard rules
1. **Primary means the organization that collected the data**: a survey owner, a measuring platform, a regulator, a peer-reviewed paper, a company filing. Blogs, "100 stats" roundups, infographics, Wikipedia, AI answers and news write-ups are **trail markers**, not sources.
2. **Open every page you cite.** List only URLs you actually opened in this session. If a page is paywalled or gone, say so and try an archived copy (web.archive.org).
3. **Quote, don't paraphrase, the key sentence** from the primary source (25 words or fewer), so the user can compare wording.
4. **Compare subject, not just number.** The most common failure is a real number about a different thing (friends and family become "influencers"; earned media value becomes "ROI"; dot-coms become "every business").
5. **Never output a verdict without confidence plus reasons.** If you cannot find the origin, say "origin not found", which is not the same as "false".
6. **Flag the interest.** Note who funded or ran the study and whether they sell the thing measured.
7. Don't scrape behind logins, bypass paywalls or CAPTCHAs, or invent page contents.

## Procedure
1. **Normalize the claim.** Write down the number, unit, population, time and attribution as quoted, e.g. `$5.78 | per $1 spent | influencer marketing | "ROI" | attributed: Influencer Marketing Hub`.
2. **Search for the trail.** Search the exact number plus topic, then the number plus "survey", "report", "methodology". Open 3-5 results. Note who each one credits.
3. **Walk back.** Follow citations hop by hop until you reach the data collector, or a dead end (no citation, a circular citation, a broken link). Record each hop.
4. **Read the primary source.** Extract:
   - publisher and author (and their commercial interest);
   - publication date **and fieldwork date**;
   - method (survey, panel, log data, model or assumption, experiment);
   - sample size and population (country, B2B/B2C, who the respondents were);
   - the exact metric (e.g. EMV vs revenue vs profit; "trust completely or somewhat");
   - the exact sentence containing the number.
5. **Check drift.** Compare the quoted claim with the source on number, subject, population, metric and date. List every mismatch.
6. **Look for something newer.** Search for a newer wave from the same publisher, or an independent replication. Report it if found.
7. **Assign a status and confidence** using the rubric below.
8. **Suggest a safe rewrite** the user can paste, with an inline citation.

## Status labels (always paired with confidence)
- `verified`: primary source found, says the same thing, scope matches, reasonably current (5 years old or less for fast-moving channels).
- `misquoted`: primary source found, but the quote changes the subject, metric, population or precision.
- `outdated`: source found and quoted correctly, but old enough that the market has changed (platform mix, privacy, AI search).
- `unsourced`: no primary data found; the trail ends in a roundup, a disowned quote or a circular citation.
- `model/assumption`: the number is an estimate or model by an interested party, not a measurement.

## Confidence rubric
- **High:** you opened the primary document, the method and sample are stated, and the quote matches.
- **Medium:** primary document found but the method is thin (vendor opt-in survey, no sample details), or you could only read a reliable secondary summary of it.
- **Low:** trail incomplete, conflicting accounts, or the source is gone with no archive.

## Output format (always this structure)
```
CLAIM (as quoted): …
STATUS: verified | misquoted | outdated | unsourced | model/assumption — CONFIDENCE: high | medium | low
PRIMARY SOURCE: Publisher, title, date (fieldwork: …) — URL
WHAT IT ACTUALLY SAYS: "…exact quote…"
METHOD & SAMPLE: …
DRIFT FOUND: number | subject | metric | population | date (list mismatches or "none")
INTEREST: who ran or funded it and what they sell
NEWER / BETTER DATA: … (or "none found")
TRAIL: hop 1 → hop 2 → … → origin
SAFE REWRITE: "…" (Source, year)
URLS OPENED: …
```

## Example triggers
- "Is it true influencer marketing returns $5.78 per dollar?"
- "Find the original source for '92% trust word of mouth'."
- "Check every statistic in this deck before I send it to the client."

For a list of claims, run the procedure for each one and finish with a table: claim | status | confidence | source.