# Referral conversion trust experiment: a dashboard | SourceX

Canonical: https://refer.sourcex.si/guides/a-referral-pilot-that-tracks-complaints-alongside-applications
Last updated: 2026-11-04

Put complaints on the same stage-by-stage dashboard as applications, and give every complaint two stages: where it surfaced and where the expectation behind it was set. Keep the complaint window open past the application read, because a misleading line often draws its complaint only after the company hears SourceX's decision. Read the pilot at day 90 and again on a later date fixed in advance. A variant wins only if it leads on applications and does not trail on complaints at both readings.

How to build two truthful versions of a message is covered in [testing eligibility messaging without hiding restrictions](/guides/how-to-test-eligibility-messaging-without-hiding-restrictions). This page is about the complaint column such a test needs: where each complaint came from, and when it arrives.

## Why one application count rewards the wrong version

A line that hints the company qualifies, or that the sender speaks for SourceX, can lift applications in the first weeks. Its cost lands later and somewhere else: with a CEO whose company is declined after being told it looked like a fit, with the partner's own client relationship, and in SourceX's review. SourceX's screening guidance already warns that a line reading as "our companies qualify" implies an acceptance nobody has given.

Experimentation teams guard against this with guardrail metrics: measures of things they do not want to make worse, even when they do not expect to improve them. Microsoft Research's Experimentation Platform team states the rule: "Key success metrics and guardrails for the product should be examined for large degradations while the experiment is in execution. If degradations are discovered, the experiment should be stopped before the predetermined end-time." In a referral pilot, complaints are the guardrail, and the hard part is that some of them arrive after the experiment looks finished.

## What counts as a complaint

ISO's complaints-handling guideline (2004 edition) defines a complaint, in paraphrase, as an expression of dissatisfaction made to an organization about its products or its complaints handling, where a response or resolution is expected explicitly or implicitly. So a complaint need not ask for anything: "who gave you my name?" counts. A complaint about how you handled an earlier one counts again.

A plain decline, silence and a neutral question are not complaints, and each has a row of its own. Complaints reach a partner from the CEO, others at the company, colleagues who hear one secondhand, and SourceX when it passes on a concern about outreach. Log the channel and date, never the person's name.

## Two stages for every complaint

Most complaints surface at one step and were set up at another. A CEO who complains after a decline is often reacting to a sentence on a page they read two months earlier. Record both.

| Step | What happens | Who controls it | What can set up a false expectation |
|---|---|---|---|
| S1 First message | The partner writes to a CEO it knows | Partner | A sentence about fit, speed, money or SourceX's view of the sender |
| S2 Follow-up | Any reminder or second note | Partner | Frequency, tone, implied deadlines |
| S3 Partner page | The page or resource the message links to | Partner | Examples, headlines and calculators that read as a forecast |
| S4 Application | The company applies itself at SourceX's own form | SourceX and the company | SourceX's own wording |
| S5 Review | SourceX reviews history, documentation, rights and relevance | SourceX | SourceX's own process |
| S6 After the outcome | The company hears where it stands and looks back at what it was told | Nobody; this is where late complaints surface | Nothing new; complaints here point back to S1 to S5 |

The partner can change S1 to S3 only. A complaint whose expectation was set at S4 or S5 goes to SourceX and is never charged to a variant. Partners see each referral's stage on their dashboard; they hear about a decision only if the dashboard shows it or the company says so, which is one reason late complaints are under-counted.

## Complaint types

| Type | The person objected that... | Charged to a variant when its origin is S1 to S3 |
|---|---|---|
| Expectation | What they were told or shown was contradicted by what happened: fit, timing, money or endorsement | Yes |
| Contact | Too many messages, the wrong channel, the wrong moment | Yes |
| Information | They were asked for files, exports, samples or logins, or were uneasy about what the sender knew | Yes |
| Authority | Someone other than the CEO was pressed to decide, or the CEO heard last | Yes |
| Disclosure | They learned afterward that the sender stood to receive a reward and objected that it went unsaid | Yes |
| Process | Their concern is about SourceX's own steps | No; forwarded to SourceX |

Two people assign the type and the origin step independently, without seeing which version the company received, then settle disagreements by rereading the message and page the company saw.

## The trust-inclusive experiment dashboard

| Row | Current version | New version | Rule |
|---|---|---|---|
| Companies contacted | Fixed list | Fixed list | The denominator for every row below |
| Reached each step, S1 to S6 | Count per step | Count per step | Drawn as a funnel, same scale in both columns |
| Verified applications at day 90 | Count | Count | Read from the partner dashboard stage |
| Complaints by step surfaced | Count per step | Count per step | Used to decide what to repair |
| Complaints by step of origin | Count per step | Count per step | Used to decide which version wins |
| Process complaints forwarded | Count | Count | Never charged to either version |
| Open companies at the late reading | Count | Count | Reported as open, not as complaint-free |

The two complaint rows answer different questions. Surfaced tells you where people got upset; origin tells you which sentence caused it. Cells show a dash, not a zero, before their reading date, and no company is named.

## Hypothesis, unit, versions, allocation and readings

- **Hypothesis (example).** "A partner page that opens with SourceX's published fit criteria produces as many verified applications per company contacted as our current page, and no more complaints originating at S1 to S3."
- **Unit.** A company the partner already knows, reached through its CEO. No bought or cold lists.
- **Versions.** Both truthful and both carrying every published condition; one step differs.
- **Allocation.** A coin toss per company on the fixed list, recorded before the first message; the same person sends both versions.
- **Early reading, day 90.** Verified applications are counted 90 days after each first message, the length of the attribution window in terms section 2.
- **Late reading.** A second date fixed before launch, such as day 180, when complaints are counted again. Companies still in review are reported as open.
- **Data.** The pilot never asks for or stores company records, and the partner signs the SourceX Channel Partner Agreement before submitting any introduction it produces.

## When the pilot pauses, and when a version wins

Written and dated before the first message:

1. **Information, authority or disclosure complaint with origin S1 to S3.** Stop both versions that day, fix the step in both, and resume only with companies not yet contacted.
2. **Two expectation complaints traced to the same step of one version.** Retire that step's wording in that version and log the change.
3. **Three contact complaints in one version.** Remove its follow-up.
4. **A concern from SourceX about the outreach.** Pause both versions until it is answered.

The new version replaces the current one only if, at both readings, its verified applications per company contacted are at least equal and its complaints originating at S1 to S3 are no higher, and no stop rule fired against it. Below 30 companies per version, report the counts and make no call; that floor is this page's choice, not a statistical standard.

## A pilot read twice (Illustrative)

Illustrative. An invented accounting firm contacts 40 client CEOs, 20 per version. The new version's page adds: "Built for established firms like yours."

| Step | Current version | New version |
|---|---|---|
| S1 First message | 20 | 20 |
| S2 Replied after follow-up | 12 | 12 |
| S3 Visited the page | 8 | 10 |
| S4 Verified application at day 90 | 3 | 5 |
| Complaints by day 90 (origin) | 1 contact, S2 | 1 contact, S2 |
| Complaints added by day 180 (origin) | None | 2 expectation, S3 |

At day 90 the new version looks better: more applications, equal complaints. By day 180 two CEOs whose companies were declined have written that the page told them the company was the kind SourceX wants. Both trace to the same S3 sentence, which fires stop rule 2 and fails the decision rule at the late reading. The sentence goes; the current page stays.

## Missing, declined and unsuccessful cases

Every company contacted stays in the denominator: no reply, declines, companies that turn out not to fit, and removal requests. Complaints from companies that later applied stay on the dashboard too. A sentence claiming one version caused fewer complaints needs the comparison fixed in advance and a gap that clears a threshold written before the count; everything else is description.

## Repair the step that caused it

With a few dozen companies, complaint counts are small, so treat each one as a repair order for the step that caused it. Repair by origin, not by where the complaint surfaced: rewriting a decline message will not fix a page that set up the wrong expectation. Report both readings side by side in any write-up, including the reversal when there is one. When a pilot ends with two pages answering one question, [when to consolidate landing pages after a referral test](/guides/when-to-consolidate-landing-pages-after-a-referral-test) applies; a network that needs one person answering complaints can use [a referral network needs an escalation owner](/guides/a-referral-network-needs-an-escalation-owner); and [the experiment behind a better partnership landing page](/guides/the-experiment-behind-a-better-partnership-landing-page) tests a frame change on identical sentences.

## Blind spots of a stage-aligned dashboard

- **Late complaints are under-counted.** Many companies never tell the partner how review ended.
- **Origin is a judgment.** Two coders reduce the guesswork; they do not remove it.
- **The late reading still ends somewhere.** Complaints after it are logged but cannot change the decision already made.
- **It measures the pilot, not the outcome.** Acceptance, licensing and any reward are decided in SourceX's review, on its timetable.

## 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 both readings come back clean and a CEO asks for the conversation, [sign up as a partner](/register), compare the company with [the published fit criteria](/who-qualifies) and go through the [program terms](/terms) first.

## Sources

- https://refer.sourcex.si/terms
- https://refer.sourcex.si/how-it-works
- https://refer.sourcex.si/who-qualifies
- https://www.iso.org/obp/ui/#iso:std:iso:10002:ed-1:v1:en
- https://www.microsoft.com/en-us/research/group/experimentation-platform-exp/articles/patterns-of-trustworthy-experimentation-during-experiment-stage/
