# Partner network marketing audit protocol | SourceX

Canonical: https://refer.sourcex.si/guides/how-to-audit-the-claims-used-across-a-large-partner-network
Last updated: 2026-11-05

Audit claims the way an internal auditor tests expenses: list everything partners publish, draw a random sample sized to the hours your reviewers actually have, check every sentence about SourceX against a dated register of approved facts, and correct by severity. Then re-sample the partners and cohorts that failed. A clean sample does not prove a clean network: 30 clean items only tell you that problems are probably rarer than about one in ten.

This page is for whoever runs a network of introducers, such as an association, an advisory firm's partner bench or a software ecosystem, whose members describe SourceX in their own posts, emails and decks. It is not about auditing how licensed data is used after a deal; SourceX's main site covers that in [can I audit how my data is used](https://sourcex.si/questions/can-i-audit-how-my-data-is-used).

## What the audit counts, and why claims decide introduction quality

Agree four definitions before anyone reads a post.

| Term | Definition used in the audit |
|---|---|
| Active partner | Published material mentioning SourceX, or submitted at least one introduction, during the audit period |
| Item | One piece of material a partner uses to reach companies or other partners: a post, email, landing page, deck, webinar slide or call script |
| Claim | Any sentence about the partner reward, a company's payout, eligibility, timing, buyers or the partner's own status |
| Useful introduction | Made with the CEO's knowledge, not a duplicate, and passing the published checks on paper |

Claims and introduction quality are the same problem seen from two ends. A company introduced on the line "your business could get USD 50,000 plus 5% of the deal" arrives expecting money the program never offered it, and the CEO who discovers that the figure is the partner's reward, not the company's payout, will discount everything else the partner said. Count unique companies rather than mentions, too: only one partner is credited per company, so a network's totals will always be lower than the sum of its partners' own claims.

## The claim sampling and correction protocol

Regulators already expect this kind of routine in consumer marketing. In its endorsement FAQ, the FTC says advertisers promoted through networks of endorsers need reasonable programs to train and monitor network members, scaled to the risk of harm: tell members what they can and cannot say, periodically search what they are saying and act on questionable practices. It adds that nobody can know every statement, but a company must make a reasonable effort. That guidance is written for consumer advertising, and whether it reaches a business-to-business introduction network is a question for counsel. Its design logic transfers either way.

| Step | What you do | What you keep | Owner |
|---|---|---|---|
| 1. Inventory | List every item that mentions SourceX by partner, channel and date; ask partners to submit links and files each quarter | The inventory, one ID per item | Channel operations |
| 2. Stratify | Split items into partners in their first 90 days, the busiest tenth by introductions, partners reactivated by a campaign, and everyone else | A stratum label on each item | Channel lead |
| 3. Size to hours | Divide reviewer hours by minutes per item; give each high-risk stratum at least 30 items when hours allow | The sample list and the random seed used to draw it | Reviewer |
| 4. Extract | Copy every sentence that falls into one of the register's seven classes | A claim log | Reviewer |
| 5. Code | Mark each claim Approved, Stale, Unsupported or Prohibited against the register version in force on the item's date | Code plus register version | Reviewer, with a second reader for Prohibited |
| 6. Correct | Prohibited: down within two business days. Unsupported: rewritten within ten. Stale: fixed at the next edit, within thirty days | A ticket with dates | The partner, checked by the reviewer |
| 7. Re-sample | Read all new material from any partner with a Prohibited finding next period; re-test a stratum whose rate passed your threshold | Follow-up results | Channel lead |

The deadlines in step 6 are recommendations, not program rules; choose your own and publish them to partners before the first audit. Draw the sample with a random number generator, not by scrolling. The posts you happen to see are the popular ones, not the typical ones.

Sample size is where most audits fool themselves. If a share of all items carries a problem and you draw items at random from a large inventory, the chance that every sampled item comes back clean is one minus that share, raised to the power of the sample size.

| Clean items in a row | Chance of a clean sample if 5% of items have a problem | Chance if 10% do | What a clean sample supports, at 95% confidence |
|---|---|---|---|
| 20 | 36% | 12% | Problems probably affect under about 14% of items |
| 30 | 21% | 4% | Under about 10% |
| 60 | 5% | 0.2% | Under about 5% |
| 100 | 0.6% | Under 0.1% | Under about 3% |

The last column solves for the share at which a clean sample would happen only 5% of the time, which comes out close to three divided by the sample size. Twenty clean items are weak evidence; a hundred are decent evidence; none of them is proof.

## The claim register: seven classes and what each must match

The register holds facts, not sentences. Partners keep their own voice; every fact in their wording has to match the register version in force on the day they published.

| Class | A line that fails | What the register holds | Severity if wrong |
|---|---|---|---|
| 1. Partner reward | "Earn USD 50,000 on every referral" | the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap, on a company's first qualifying transaction; repeats within 12 months pay 15% with no second floor; never calculated on SourceX's fees or revenue (terms, section 4) | Prohibited |
| 2. Company payout | "Your company could get 15% on top from SourceX" | The company's payout is fixed per dataset in its signed Dataset Deal Confirmation; earlier estimates are indicative | Prohibited |
| 3. Payment timing | "Paid as soon as the company signs" | Payable only after the Acquisition Partner's final written acceptance, completed handoff, cleared payment to SourceX, payment of the company, finance approval and the partner's tax details; deadline 60 calendar days after the later of acceptance and handoff (section 5) | Prohibited |
| 4. Eligibility | "Any business can apply and get paid" | 50 or more full-time employees at peak, contractors excluded; documented history; rights to license; an authorized contact; geography as published on the who-qualifies page | Unsupported |
| 5. Buyers | Naming an AI developer as a buyer | Never name buyers; partners receive no confidential buyer information (section 7) | Prohibited |
| 6. Partner status | "Certified SourceX partner" | A signed Channel Partner Agreement lets a partner submit introductions; the terms describe no certified or official status, and the signed agreement governs confidentiality (section 8) | Unsupported; Prohibited for "certified" or "official" |
| 7. Version drift | A recruiting deck that calls the agreement new partners sign by its earlier title, or ties payment to "the AI lab's acceptance" | Section 11 lists dated term changes; the current trigger is worded around the Acquisition Partner, and only agreements already signed under the earlier title keep it | Stale |

Class 7 deserves its own calendar entry. Section 11 of the program terms lists five dated versions between 12 and 22 October 2026. Each one turned some accurate partner sentences into stale ones overnight, without any partner doing anything wrong. Update the register within days of a change and re-check that class. The sentence-by-sentence editing method in [interesting angle or unsupported claim](/guides/the-difference-between-an-interesting-angle-and-an-unsupported-claim) is what reviewers use inside step 5.

## One quarter's audit across 140 partners (Illustrative)

Illustrative: an invented network with invented counts. The network has 140 partners and 612 items in its inventory. Two reviewers have 24 hours for the quarter at about 15 minutes per item, which buys 96 items.

| Stratum | Partners | Items in inventory | Sampled | Sampled items with a finding | Prohibited findings |
|---|---|---|---|---|---|
| First 90 days | 38 | 142 | 30 | 9 | 3 |
| Busiest tenth | 14 | 188 | 30 | 4 | 0 |
| Reactivated by the spring webinar | 22 | 96 | 21 | 11 | 2 |
| Everyone else | 66 | 186 | 15 | 2 | 0 |
| Total | 140 | 612 | 96 | 26 | 5 |

**The underperforming cohort.** Eleven of the 21 sampled items from the reactivated partners had findings, most of them Stale. The cause was one shared deck: the spring webinar's slides predated the October term changes, named the agreement by its earlier title and tied payment to the AI lab's acceptance. The cohort's problem was its source material, not its people. Replace the deck at the source, tell the 22 partners what changed and why, and re-sample the stratum next quarter.

**The duplicate-risk case.** Two partners in the first-90-days stratum each posted that a named 400-person regional distributor was "my introduction to SourceX this month." Neither post can settle anything: SourceX decides attribution from dated records, credits one partner per company and credits nobody for a company it already knew. The audit's job is narrower. Under this network's register, naming a company publicly without its approval is Prohibited, so both posts came down within two days, and both partners were pointed to the [checklist for preventing duplicate claims](/guides/how-to-prevent-duplicate-claims-before-they-reach-sourcex). The network did not rule on credit.

**The capacity limit.** In week nine a recruiting push added 210 new items, taking the inventory to 822. Reviewer hours did not move, so the share of material read fell from about 16% to about 12%. The protocol's answer: keep 30 items in each high-risk stratum, cut "everyone else" to the floor and, if even that will not fit, pause recruiting until it does.

## How many items your reviewers can actually read

| Pressure | Signal to watch | What to do |
|---|---|---|
| Reviewer hours | Items in inventory per reviewer hour rising quarter on quarter | Hold the high-risk strata at 30 and shrink the low-risk stratum first |
| Support load | Corrections that turn into long threads with partners | Budget time per correction, and fix shared materials once instead of item by item |
| Concentration | A few partners producing a large share of items; here the busiest tenth made 188 of 612 | Sample them as their own stratum, never in proportion to volume |
| Overlap | Two partners' material describing the same company | Log it separately and send partners to SourceX's duplicate review |

The same arithmetic governs introductions themselves when review is scarce; the guide to [prioritizing introductions when review capacity is limited](/guides/how-to-prioritize-introductions-when-review-capacity-is-limited) applies it on SourceX's side of the line.

## Gates before the next cohort, and the partner's right to reply

- **Recruit only through a clean gate.** Open the next recruiting wave when no Prohibited finding is still open and the new-partner stratum's finding rate sits at or below a threshold you set in advance, such as one in ten.
- **Keep the register current.** After any change to the program terms, update the register and re-check the affected class within a couple of weeks.
- **Let partners contest a finding.** A partner who shows the source for a claim gets the finding withdrawn, and the reviewer records the decision with the register version. Corrections go to the partner privately, never into a league table.
- **Correct the register too.** When many partners trip over the same class, the register's wording is probably unclear. Rewrite it, then re-test.

Quality measures for introductions, as opposed to claims, are set out in [measuring referral quality instead of volume](/guides/how-to-measure-referral-quality-instead-of-referral-volume).

## Where sampling misses

- **Private messages and calls.** One-to-one conversations rarely reach any inventory. Asking introduced CEOs what they were told about payment covers some of the gap.
- **Deleted material.** Partners can remove a post before review; ask for quarterly submissions and date-stamp what you receive.
- **Small strata.** Fifteen clean items say little, as the sample-size table shows.
- **No benchmarks.** SourceX has published no data on claim error rates across partner networks, and the counts above are invented. Your own first two quarters are the only baseline you will have.

## 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.

Before the next recruiting push, build your register from the current [program terms](/terms) and [who qualifies](/who-qualifies), run one sampled audit, and have each partner who passes [register for a partner link](/register) so the introductions they make can be tracked.

## Sources

- https://refer.sourcex.si/terms
- https://www.ftc.gov/business-guidance/resources/ftcs-endorsement-guides-what-people-are-asking
- https://refer.sourcex.si/who-qualifies
- https://sourcex.si/how-it-works
