Use Synthetic Buyers to Catch Broken Intent Data Before Launch

November 16, 2026•10 min read

Run a small suite of synthetic buyer journeys before relying on production intent data. Give each fictional record a known starting state, perform an approved action, and write down which event should exist, where it should appear, and what must stay absent. This is especially useful when a perfectly healthy dashboard can hide a broken handoff that affects only one buyer path.

Start with the buyer decision

Imagine an industrial coatings supplier rolling out new forms for sample requests. A test engineer opens the guide in one browser, another test user submits an inquiry in a separate context, and a known account uses an existing contact record. A sloppy identity join could attach all three events to the known contact. The test has succeeded only when the system records the right event at the right grain without inventing a person-level history.

Keep these records out of live segmentation, model learning, and reported results. Use a nonproduction setup when feasible and coordinate any production test with the system owners. Carry a marked test ID from the event through processing, then inspect exact fields and timestamps. A live-looking name is a liability when a test accidentally reaches a seller or recipient.

The outside evidence and its limits

IAB Tech Lab's finalized ECAPI clean room guidance addresses shared data exchange formats. Apply the same discipline to your own test contracts while verifying the workflow's actual semantics separately. See the IAB Tech Lab ECAPI clean room guidance.

A fictional buyer walkthrough

Send an original event labeled test-104 and resend the identical event after an artificial timeout. Then issue a new event labeled test-105. Expect two distinct event records and only the approved number of sales actions. If all three produce tasks, the team needs to inspect the retry boundary.

A fully suppressed fictional contact completes a test journey. Review whether a permitted event is recorded internally while a message is prevented. Deleting every trace of the test makes it harder to prove that the suppression worked.

Positive path

Submit a permitted test inquiry using a declared account. Expect one accepted form event and one routed test record with the agreed timestamp and channel label. Apply the rule while the source evidence is still available. Check processing time separately so a delayed arrival does not rewrite the time the action happened.

Treat the positive path check as a release check when the team can identify an expected state in advance. A tester should document differences between a warning, an expected missing result and a real violation. Those classes may look identical if they are collapsed into a null or zero. Do not fill missing evidence with a guess that fits the dashboard. They often show a gap in timing, coverage, or ownership that a filter would hide.

Choose fixtures that represent real implementation failures

A fixture is a known test input and expected outcome. Make one for a timestamp near a day boundary, one for a missing identifier and one for the same event arriving more than once. A buyer submitting a request after midnight can fall on a different reporting date from an analyst in another time zone. Store occurrence time and preserve the reporting conversion separately. If the data moves between systems, verify that the output respects the same meaning and does not infer a stronger identity relationship.

Negative path

Run an adverse example through the negative path check. Load a page without submitting the form. Expect permitted aggregate viewing activity only when that capture is authorized. A sales stage should never advance on the strength of the test view.

Put the next step in the operating notes. Document missing events that are intentional and make them part of automated assertions if supported. Keep the evidence from the negative path check with the system owner. If two sources appear to disagree, compare the units and processing times before changing any rule. A known control record can make the issue easier to reproduce and protect the production data from exploratory edits. If the evidence is too weak for a person-level action, retain it only for the approved account, diagnostic, or aggregate purpose.

Put a failure example beside the standard example for the negative path check. For the first example, require the reviewer to explain why the rule should stop. For the second, ask why it should advance. This exposes definitions that sound precise in a meeting but do not lead to consistent decisions under ordinary conditions.

Duplicate path

Repeat the same event with its original ID, then make a distinct submission. These two operations should have different effects in a deduplication design.

Give the owner of the duplicate path check a way to compare versions. A failed the duplicate path check calls for a record of the first stage that diverged from expectation. Inspect the upstream event and the immediate downstream effect. If the upstream state cannot be verified, the failure may be a traceability problem as well as a processing problem. The review record for the duplicate path check should retain this disposition. Verify a retry causes no extra CRM action and a genuinely new event stays visible.

To verify the handoff, ask the receiving owner to explain the decision in their own words using only the approved record.

The best evidence for the duplicate path check is the one an authorized second person can check. Use an approved source rather than a copied conclusion, record the relevant version, and retain the date. This is especially important if a later report uses journey assertions to support a change in time or spending.

Suppression path

Mark the fictional contact as suppressed and send it through the decision step without a live send. Observe the suppression result in the processing log. For this case, the expected next step is clear. Fail the release if a suppressed record becomes eligible simply because a fresh score arrived.

Treat the suppression path check as a release check when the team can identify an expected state in advance. A tester should document differences between a warning, an expected missing result and a real violation. Those classes may look identical if they are collapsed into a null or zero. Keep a clean test case that proves the correction under controlled conditions before extending the repair to every record. Record whether the expected action changes and why.

Test the boundary of the suppression path check using a case that almost meets the documented condition. Record the exact reason it falls outside. If two reviewers give different answers, inspect the underlying definition before you widen the rule. The disagreement may show an ambiguous requirement or a field that should be confirmed at intake.

Separate checks on data from checks on buyer-facing side effects

A test can verify that an event exists and independently verify that no real message was sent. Build assertions for both. When using a production path under approved controls, keep fictional contacts out of model learning and campaign audiences. Check other systems that might consume the same event, including task automation and reporting. Do a final run after changing mappings, exclusion rules or tenant settings. A passed initial suite establishes what happened for those test records under that configuration, and it remains valuable as a regression reference.

Out-of-order path

Write the expected result for the out-of-order path check before inspecting any live rows. Queue a form and a later outcome, then simulate the outcome arriving before the form. Establish whether the integration can safely wait, reconcile, or route to review.

To close the review, write down what the decision owner is expected to do. Retain the original event chronology and note how the system handles late data. Keep the evidence from the out-of-order path check with the system owner. If two sources appear to disagree, compare the units and processing times before changing any rule. A known control record can make the issue easier to reproduce and protect the production data from exploratory edits. An exception that vanishes into a generic queue cannot teach the team much.

Turn the out-of-order path check into a short review record with the observed evidence, the proposed action and its owner. The next operator should be able to continue the work without being asked to infer missing details. If a check is unresolved, note which subsequent action must wait and what evidence can release it.

Build the review worksheet

Use a worksheet with one row per expected event and action. Include test ID, fixture version, intended tenant, occurrence timestamp, expected event count, expected action count, and observed result. If the test fails, link the owner to a permitted trace instead of copying sample payloads into general reporting.

Script distinct cases for missing event time, duplicate delivery, invalid geography, a valid request with unavailable match, and suppression applied just before activation. Expected behavior will differ by configuration. Require specific proof that data and action stay in the intended client context.

How to calculate journey assertions

In the fictional worksheet below, the denominator consists of scripted fictional test cases per build. The numerator counts cases that met all expected positive and negative conditions. Both figures belong in the report. A reader who has only a percentage cannot tell whether a handful of outcomes drove the apparent difference.

Illustrative groupCount meeting ruleEligible countRate
Build with explicit edge-case assertions232688.46%
Previous build tested on the same synthetic journeys192673.08%

The first result is 23 out of 26, or 88.46 percent. The second is 19 out of 26, or 73.08 percent. On unrounded rates the first is 15.38 percentage points above the second. If you present a relative change, keep the comparison denominator explicit. Using 73.08 percent as the reference gives a difference of about 21.1 percent in relative magnitude.

A failed suppression case can block release even if a build passes most tests. Classify failures by consequence before comparing build-level pass totals.

Record the snapshot date beside the journey assertions result. If a reviewer changes the comparison rule, keep the original journey assertions calculation and clearly identify the revised version.

Reconcile the count used for journey assertions to a frozen sample or a permitted source extract. Keep intentional exclusions and unavailable observations in separate rows of the audit. If the reader cannot inspect the reason the denominator changed, hold the comparison until the data owner can explain the movement.

Give each assertion a check for downstream effects

A test may pass in the collection log while the same event updates the wrong CRM client. The test fixture should therefore include expected destinations and any forbidden side effects. Use a marked event as a trace through each approved component. When a record is intentionally withheld, verify that result at the handoff rather than inferring it from missing sales work. Write down how the next operator can reproduce the result after a future update.

The journey assertions counts leave 3 cases that did not meet the rule in the first fictional group and 7 in the second. In actual data, inspect those remaining cases before calling them all failures. For example, some may be awaiting a decision, while others may be complete and outside the rule. Keep a reconciliation showing the final and unresolved dispositions. That distinction matters more than rounding the percentages to another decimal place.

Give the person investigating journey assertions a test identifier, the expected result, and a permitted way to see the difference between the two states. A root cause might sit in the source, the integration, the comparison rule, or the report. Record the earliest confirmed divergence and retest from there after correction.

Protect feedback from test data

Customer feedback can make two-way intent data more useful, so protect it from test noise. An artificial lead should not acquire a won, lost or qualified label in the model just because a test successfully reached a CRM stage. Verify that excluded fixtures remain excluded after the system refreshes a segment.

For more on the signals behind this decision, read the Cortex intent data insights.

Make the diagnostic repeatable

Block activation when an unsafe path fails and rerun the complete test suite after repair. Walk an approved synthetic record through the riskiest handoff. If an unintended record can create a task or send, fix that path before asking what the score could do in production. A small acceptance test can reveal exactly where responsibility changes hands.

Visit Cortex Marketing AI for an overview of how website activity and sales feedback inform buyer prioritization.

Bring a sample reconciliation, an approved test case, and the expected output. To put the decision into practice, call Cortex Marketing AI to arrange a consultation.

Back to Blog