The study is built so that a disappointing result is publishable. If owners rate every variant the same, or say they would ignore all of them, that is the finding, and it changes the advice partners get.
Facts the study starts from, and the question it asks
| Kind | Statement | Source |
|---|---|---|
| Program fact | Introductions made without the decision maker's knowledge are usually not a fit | SourceX who-qualifies page, checked 2026-10-26 |
| Program fact | The partner reward is paid separately by SourceX and never deducted from the company's agreed compensation | Program terms, Version 2026-10-22.2, since replaced by version 2026-11-04 |
| Regulator guidance | A payment is a material connection to disclose when a significant minority of the audience would not expect it; vague labels do not explain it | FTC Endorsement Guides FAQ |
| Conditional | Whether an introduction leads to an application, review or payment | The owner's choice, SourceX review and the payment conditions in the terms |
| Proposed thesis | A permission ask and a way to verify matter more to owners than the sender's seniority | A question for the study, not a finding |
| Term | Definition in this study |
|---|---|
| Factor | One feature of the message the study changes on purpose |
| Level | One version of a factor; each factor here has two |
| Vignette | A hypothetical message an owner reads and rates; it is never sent to a real company |
| Stated outcome | What the owner says they would do, on a fixed scale |
| Negative case | An owner who would not reply to any variant, or chooses to report the message |
The five factors and their two levels
This is the core of the trust-factor message study. Each factor is a candidate drawn from SourceX's published etiquette and screening guidance or the FTC's disclosure guidance, and each has two levels an owner could plausibly receive.
| Factor | Level A | Level B | Why it is a candidate |
|---|---|---|---|
| 1. Sender | An advisor the owner has worked with for years | Someone the owner met once at an event | Relationship history may carry more trust than any wording |
| 2. Permission | Asks whether the owner wants an introduction | Says an introduction has already been sent | SourceX treats introductions without the decision maker's knowledge as usually not a fit |
| 3. Disclosure form | States who pays the reward, how it is calculated and that it never reduces the company's payout | Says only "I may benefit from this" | FTC staff warn that vague labels do not explain a relationship |
| 4. Ways to verify | Links the published terms and fit criteria and suggests checking SourceX through its own site | No links | Owners told to verify independently may trust the message more, or read it as a warning |
| 5. Promise | Says nothing is binding unless the owner agrees terms and signs | Says the records "could be worth a lot" | Outcome hints may raise interest and lower trust at the same time |
Both disclosure levels disclose. No variant hides the reward: SourceX's warm introduction etiquette asks partners to say plainly in the first message that they may earn one, and testing a concealed version would mean writing a message the program tells partners never to send. The study asks how the disclosure is phrased, not whether it appears.
Level B of the permission factor breaks the program's own rule on purpose. Because vignettes are hypothetical and never sent, the study can measure how much a skipped permission ask costs in trust without any real owner being ambushed.
The message anatomy: one full variant
Here is the all-Level-A variant, annotated. The other 31 combinations swap one or more marked parts.
[Sender, A] "We've worked together on your audits for eight years, so I'll keep this short. [Permission, A] Would you like an introduction to SourceX, which licenses operating records from established companies for AI training and evaluation? [Disclosure, A] If your company licenses data through them, SourceX would pay me a referral reward of the greater of $50,000 plus 5% and 15% of what it actually pays your company, with no cap, funded separately so it never reduces your payout. [Verify, A] Their program terms and fit criteria are public at refer.sourcex.si; check them through their own site rather than this email. [Promise, A] Nothing is binding unless you agree terms and sign, and you can say no in one line."
For partners who want a ready message rather than an experiment, the warm introduction brief and the guide to discussing referral rewards transparently cover the wording.
The study design, written before any owner reads a message
Design. Five factors at two levels give 32 variants. Each owner reads four, drawn at random, in random order, and rates each one separately. Random assignment means any difference in ratings between levels is caused by the message feature, within this hypothetical setting.
Outcomes, fixed in advance.
| Outcome | Scale | Role |
|---|---|---|
| Likelihood of replying | 1 to 7 | Primary |
| Belief that the message is honest | 1 to 7 | Secondary |
| Chosen next action | Reply yes; ask a question; check SourceX independently; ignore; report as spam | Secondary, includes negative options |
| Free-text reason | Optional, one line | Coded after the main analysis |
Analysis, fixed in advance. For each factor, the difference in mean primary rating between Level A and Level B, with a 95% interval. One interaction is pre-registered: whether a long relationship (Factor 1, Level A) offsets a skipped permission ask (Factor 2, Level B). Any other comparison is labeled exploratory.
Sample size rule. Set it from the smallest difference worth changing advice for, half a point on the seven-point scale, and from the spread of ratings in a pilot of about 30 owners. Publish both numbers with the pre-registration. This page sets no target count because no pilot has run.
Who takes part, and how consent works
Sampling frame. CEOs and owners of companies that would plausibly meet the baseline: 50 or more full-time employees at peak, contractors excluded, with operating history. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.
Two recruitment routes, reported separately. Owners invited by partners who already know them, and owners reached through a named research panel if one is used. Partner-invited owners already trust the person asking, so pooling the routes would inflate every rating.
Consent. A plain-language notice before the first question; no company name, records or financials collected, only size band, industry, role and country; withdrawal allowed until the data freeze; any incentive stated up front and never tied to answers. Partners who invite owners see no individual responses.
Counting everyone, including the owners who say no
The flow reports invited, started, completed and excluded, with a reason for every exclusion (for example, a failed attention check or a respondent outside the frame). Three groups stay in the analysis even though they are inconvenient:
- Universal decliners. Owners who rate every variant at the bottom or choose "ignore" for all four. They are evidence that, for some owners, no wording helps.
- Reporters. Owners who choose "report as spam" for any variant. Their share by factor level is a result in its own right.
- Partial completers. Shown with how far they got. Their completed ratings count; their missing ones are never filled in.
Descriptive versus causal. The experiment supports causal claims about stated responses to hypothetical messages, nothing more. A second phase, coding real introductions with partner and owner consent for the same five features and their outcomes, can describe what happens in practice but cannot separate cause from selection: partners choose different wording for owners they know differently. The report keeps the two phases in separate sections.
Materials released with the results
- All 32 variants, word for word, and the rating instrument.
- The pre-registration with its date and every deviation.
- The participant flow with counts that reconcile.
- Factor-level results with intervals, including factors that made no difference.
- The coding frame for free-text reasons, applied by two coders, with their agreement rate.
- A one-page worksheet of the five factors for partners, with the study date, a method note and a link to this page.
Where the conclusion could fail
Counterexample one: distrust that wording cannot fix. Illustrative: the owner of a 180-person staffing firm distrusts AI companies on principle. Every variant gets the lowest rating and "ignore". The five factors explain nothing about her answer; the objection is to the industry, not the message. The page on owners who distrust AI companies covers that conversation, and why trust a new company lists the facts an owner can check.
Counterexample two: trust without fit. Illustrative: a 30-person design studio receives a model introduction from its accountant of ten years and replies within the hour. Review declines it, because the company has fewer than 50 full-time employees at peak. The message earned trust and still cost the owner, the partner and the reviewer time. Trust is a reason to send a well-made message only after the company passes who qualifies; the cost of an unqualified introduction describes how to measure that waste, and poor-fit patterns shows the usual causes.
Conditions that would change the recommendation. If the plain disclosure lowers ratings, the advice is to keep it and improve its wording, not to drop it. If the permission factor shows no effect, the program rule still stands, because introductions without the decision maker's knowledge are usually not a fit regardless of how owners rate them. If sender relationship swamps every other factor, partners should introduce only owners they know well and treat second-degree introductions as a separate, slower route, compared in warm introduction versus cold outreach.
Other limits. Owners rating hypothetical messages may answer more generously than they act. Panel owners may differ from owners partners know. The study tests English-language messages only.
Start with a pilot. Run it: 30 owners, four variants each, primary outcome only. Publish the rating spread and the share of universal decliners, then set the full sample. A partner can run a smaller version on their own practice: log every permission ask for a quarter, note which of the five features each message had, and record replies within 14 days. That log is descriptive, but it shows which features you already use. The annual report's research blueprint explains how such logs feed a larger study. Introductions are submitted only after the partner has signed the SourceX Channel Partner Agreement.
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.
When a company passes the fit screen and you know the owner well enough to ask permission, register for a partner link, check the company against who qualifies, read the program terms, and send a message that asks first and discloses plainly. Related study designs are listed in the guides hub.