# Data referral language research: a study design | SourceX

Canonical: https://refer.sourcex.si/guides/the-data-referral-vocabulary-owners-actually-understand
Last updated: 2026-11-04

Which words in a data referral owners understand is an open question, and this page publishes the study that would answer it, not an answer. Owners whom a partner already knows would paraphrase ten program terms, such as Actual Supplier Payout and verified application, each shown either as published or as a plain rewrite, and coders who cannot see which version was shown would score every paraphrase. The sample, instruments, codes and reporting rules are fixed here first. No results are published, and none are implied.

This design is for partner research and content leads who write introductions, one-pagers or FAQs for networks of owners and want evidence about wording before they rewrite anything. It tests words one at a time. Whole messages are tested in the [executive understanding interview](/guides/a-referral-content-test-focused-on-executive-understanding), partners' own knowledge in the [payment-trigger survey](/guides/how-to-measure-whether-partners-misunderstand-the-payment-trigger), and disclosure wording in the [professional network trust survey](/guides/the-professional-network-trust-barometer-for-ai-data-referrals).

## A word can be read and still misread

A referral reaches an owner carrying defined terms from SourceX's pages: Actual Supplier Payout, Dataset Deal Confirmation, Closing, exclusive for AI training. Each has a published meaning, and each has a plausible wrong one. An owner who reads "referral reward" as money taken from the company's payout, or "verified application" as an approval, may decide on a false picture, whatever the rest of the message says.

The U.S. government's plain-language guidance on digital.gov has a case that shows why words need their own test (our paraphrase of its paraphrase-testing page): readers of a Veterans Benefits Administration letter said it was clear, yet their paraphrases showed several held a different definition of "service-connected", one that would have led them to act wrongly. The FTC makes a related point about disclosures. Its endorsement guides FAQ ties a disclosure's adequacy to what readers would actually understand and asks for the essential information in words that are easy to understand (our paraphrase). A partner's note to an owner carries both kinds of words: program terms and the disclosure that the partner could be paid.

The study asks three questions, worded so that no answer is assumed:

1. For each term, how often does an owner's paraphrase match SourceX's published meaning, and how often does it contradict it?
2. For each term, does the plain rewrite change that pattern?
3. Which misreadings, if acted on, would change an owner's decision?

## Ten terms, each paired with a rewrite from SourceX's own wording

Every rewrite quotes or closely follows a SourceX page, and every item carries the terms version it was built from (2026-10-22.2). Before fieldwork, the full set goes to SourceX for a meaning check.

| Term as published | Where it appears | Plain rewrite tested | Misreading that would matter |
|---|---|---|---|
| 1. Actual Supplier Payout | Terms, section 4 | The compensation SourceX actually pays your company | An estimate, or what an AI developer pays |
| 2. Verified application | Terms, section 2 | None yet: SourceX's pages do not define it | That the company has been accepted |
| 3. Acquisition Partner | Terms, section 5 | None yet: SourceX's pages do not define it | That the company signs with it |
| 4. Exclusive for AI training | FAQ | You agree not to license the selected data to anyone else for AI training | That it covers every record the company holds |
| 5. Master Supplier Licensing Agreement | How it works | One agreement, signed once after qualification | That signing it licenses data |
| 6. Dataset Deal Confirmation | How it works | A short agreement fixing the exact dataset, your payout, rights, the processing route and the deadlines | A summary with no force |
| 7. Closing | How it works | When SourceX has accepted the data in writing and every condition is met | The day the agreement is signed |
| 8. Deidentified dataset | FAQ | A copy SourceX de-identifies before use; the agreement records that re-identification may still be possible | That nobody could ever be identified |
| 9. Referral reward | Terms, section 4 | A payment SourceX may make to whoever introduced you, separate from your payout and never deducted from it | That it comes out of the company's payout |
| 10. 50 or more full-time employees at peak | Terms, section 1; who qualifies | Your highest full-time headcount, contractors not counted | Today's headcount, or a count with contractors |

Terms 2 and 3 are tested for what owners take them to mean, and no paraphrase of them is coded right or wrong until SourceX supplies its meaning in writing. If SourceX's pages change during fieldwork, the study pauses, the affected items are rebuilt and the interviews before and after the change are reported separately.

## Who is asked, how they agree and who is left out

**Frame.** Owners and CEOs of companies with 50 or more full-time employees at peak whom a participating partner knows personally. The partner sends a research-only invitation; the interviewer is someone else.

**Strata.** Owner-run companies, and companies whose CEO was hired by a board, because the two may meet contract language differently. Within each, owners are asked whether they had heard of licensing company records for AI training before; the answer is recorded, not screened on.

**Two rounds.** Round 1 runs six to nine interviews per stratum, the range digital.gov suggests for paraphrase testing; its purpose is to find misreadings and repair rewrites, and rewrites may change only between rounds. Round 2 is fixed before it starts at 40 completed interviews, 20 per form. That number is a design choice that supports description, not a significance test.

**Never asked.** Companies whose CEO has declined SourceX or a partner's suggestion, so research never becomes a way back in; companies with an open introduction or application; anyone the partner knows only through board materials; and anyone the partner cannot name as a personal contact.

**Consent, read aloud before any card.** The interview takes about 20 minutes. You will read short sentences about a data licensing program and say what they mean to you. Nothing about your company's data is asked. Your words are recorded and reported without your name or your company's. You can stop at any point without giving a reason. Taking part has no bearing on any SourceX decision, now or later. The study is run with a SourceX referral partner, who can earn referral rewards from SourceX for introductions; this interview is not one.

## The paraphrase interview, card by card

1. **Warm-up.** One practice card about an unrelated topic, so the owner hears that there are no wrong answers.
2. **Ten cards.** Each holds one short sentence containing one term. On Form A, odd-numbered terms appear as published and even-numbered terms as rewrites; Form B swaps them. Card order is shuffled for every owner.
3. **Two prompts per card.** "In your own words, what does this sentence tell you?" and "What would you expect to happen next?" Never a yes/no question, which digital.gov's guidance says yields little.
4. **No correction during the cards.** The interviewer writes the owner's words as spoken.
5. **The other-readers prompt.** "Which of these words might other owners find confusing?" The guidance notes people find it easier to point out confusion on others' behalf.
6. **Side-by-side preference.** Each term's two versions, shown together, with "Which would you rather receive, and why?" This is secondary and never replaces the paraphrase.
7. **Close.** The interviewer hands over links to the SourceX pages each card came from, so no misreading leaves the interview uncorrected, but only after every answer is recorded.

An example card pair: Form A reads "Any reward is based on the Actual Supplier Payout." Form B reads "Any reward is based on the compensation SourceX actually pays your company."

Digital.gov's testing guide places paraphrase and usability testing before a controlled comparison. Round 1 is the paraphrase stage; Round 2's form swap is the comparison, and it runs only after the rewrites have survived round 1.

## Codes, coders and every case in the count

| Code | Meaning |
|---|---|
| M, match | The paraphrase carries the published meaning |
| P, partial | The core is right, a stated detail is missing |
| X, misread | The paraphrase contradicts the published meaning |
| D, decision-changing misread | An X that would change what the owner does, such as expecting the reward to reduce the payout |
| N, don't know | The owner says they cannot tell |
| H, held | A paraphrase of term 2 or 3, waiting for SourceX's meaning |

Two coders score every paraphrase against the published meaning without seeing which form the owner received; a third settles disagreements. Agreement between the first two, as percent agreement and Cohen's kappa, is reported before any count. Every invitation is counted: invited, agreed, completed, stopped partway and declined. An owner who stops at card four contributes four coded answers and appears in the stopped count.

## Reporting rules, with the results section left empty

- For each term and each version: completed interviews, then the count in every code, never a code without its denominator.
- A cell with fewer than 10 answers is shown as a count, never a percentage.
- Every decision-changing misreading is quoted in full, without names, where the owner consented to quotation.
- No league table of "most understood" words, and no claim that a rewrite works beyond what its counts show.
- Round 1 is reported as a list of misreadings found and rewrites changed, without counts.
- The report states its frame: owners reachable through partners who took part, interviewed in English, on the terms version named.
- Until fieldwork ends, the results section of this page stays empty. As of 2026-10-27 no interview has been run.

## What could make a word-level study wrong

- **Context.** A term on a card reads differently from the same term inside a full email; a match on a card does not prove a match in the wild.
- **The frame.** Owners who know a partner well may be more familiar with the program than owners reached cold.
- **Interviewer effects.** Tone can signal a right answer; a script and a separate interviewer limit, but do not remove, that.
- **Moving terms.** SourceX publishes new versions of its terms; results tie only to the version tested.
- **Understanding is not action.** A clearer word may not change who applies. Testing that needs an outreach experiment, such as the [one-variable comparison](/guides/how-to-compare-two-outreach-angles-without-changing-everything).

## Writing referral copy while the study is only a design

Use SourceX's own wording for defined terms and link the page each one comes from. Where a term has no published definition, say so rather than supplying one. Put the disclosure in plain words, as the FTC's guidance asks, and never in shorthand. For the evidence partners can cite about program results today, see [partner results research](/guides/partner-results-research); more research designs sit in the [guides hub](/guides).

## Next step

> **Refer a company. No cap on rewards.** Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.

To run this design with owners you know, [register as a partner](/register) first, keep every introduction separate from the study, check companies against the [published screening profile](/who-qualifies) and quote terms only from the current [program terms](/terms).

## Sources

- https://refer.sourcex.si/terms
- https://refer.sourcex.si/who-qualifies
- https://sourcex.si/how-it-works
- https://sourcex.si/faq
- https://digital.gov/guides/plain-language/test/paraphrase-testing
- https://digital.gov/guides/plain-language/test
- https://www.ftc.gov/business-guidance/resources/ftcs-endorsement-guides-what-people-are-asking
