Before using a clean-room report to measure buyer intent, reconcile the event totals at each boundary between collection and the final report. Start with eligible records and account separately for retries, duplicates, rejected records, unmatched records, restricted outputs, and late arrivals. The business question is simple. When a count gets smaller, did something change in buyer behavior or did an event stop being visible in the analysis? Without a reconciliation, those explanations can look identical.
On November 3, 2026, IAB Tech Lab announced final guidance for providing ECAPI events to data clean rooms and separately opened a version 1.1 specification update for public comment. Event transport and clean-room guidance give implementers a starting point for agreeing on how data arrives. The next step for a buyer is deciding how to account for the records that do and do not appear in a report. The reconciliation below is a proposed operating method for that problem. It reports no Cortex campaign results.
Choose an outcome whose business meaning survives the trip
A team that sells software for multi-site maintenance can observe a request, a held discovery meeting, an accepted opportunity, or a signed agreement. Each can be useful, depending on the question. Suppose the team wants to know whether external buyer research combined with owned-site engagement is associated with accepted opportunities. Define an accepted opportunity in sales terms before creating the event. A sample rule might require a serviceable customer, a documented maintenance problem, and a mutually agreed evaluation step. Preserve the date when the criteria are first satisfied.
A raw form event might be technically easy to transport and still be the wrong business outcome. It can include support questions, duplicate requests, or accounts in markets the business does not serve. Likewise, a deal stage recorded at the end of the month may reflect data entry timing rather than the actual sequence of events. Give the event owner a description and an example of a row that fails the definition. That makes it easier for the receiving team to notice when a source starts producing something different under the same name.
Use the right grain for each check. A receipt log counts transmissions, including retries. A deduplicated event table counts distinct approved occurrences. An account-level outcome table counts accounts meeting a qualification rule during a defined window. One account can send several events and still contribute only one outcome. If the presentation shows an account qualification rate, use a denominator of eligible accounts rather than a count of incoming payloads. Store the campaign cohort and observation window with the account outcome so later teams can reconstruct the rate.
Make a ledger for the places a record can go
Write the path in a way that can be verified. Begin with outcomes recorded in the source system for the period. Count records excluded because they are outside the approved purpose or data-use scope. Count each approved event once at the business level and keep a separate total for delivery attempts. Then account for accepted events, rejected events, unresolved transmissions, distinct match candidates, usable matches, and records withheld from output. A clean-room partner might only permit approved aggregate information about some stages. Request the smallest aggregates that can still distinguish meaningful failure modes.
Assign an owner to each transition. Sales operations may own the opportunity definition. Engineering may own delivery, retry behavior, and source-side totals. A privacy or data governance owner can approve the purpose, fields, access, retention, and outputs. The analytics owner should decide which populations can appear in a numerator or denominator and explain why. When all four responsibilities are implicitly assigned to the partner receiving events, the organization can know that a payload was accepted while being unable to say whether it was appropriate or analytically useful.
A good ledger also records when counts were produced. Compare intake for events occurring in a defined interval against a settled receipt period. If delivery may lag, report a provisional cut and a closed cut. An event can move from unresolved to accepted after the first report. A dashboard that overwrites the earlier total without identifying the cut makes it look as if a previously published result was calculated incorrectly. A clear report can say which data arrived late and when the final count was locked.
Walk through a fictional reconciliation before seeing production data
Imagine a fictional maintenance vendor whose CRM records 100 accepted-opportunity events across a set of approved dates. Ten events are outside the permitted data-sharing scope. That leaves 90 eligible business events. The sending system records 96 delivery attempts because six attempts are retries. The receiver eventually confirms 86 distinct accepted events, identifies two distinct rejections, and still has two unconfirmed distinct events. In this example the useful arithmetic is 86 plus two plus two equals 90. The 96 attempts tell the engineering team about delivery activity. They do not belong in the acceptance denominator.
Suppose only 65 of the 86 accepted events have a usable match for the account-level question. Then 21 accepted events remain outside the matched view. The analyst should ask whether those missing matches cluster by source, region, or customer type, using only information the teams are authorized to inspect. If regional operators match less often than national companies, the final report may overrepresent the latter. A large matched total can still describe a biased subset. If one outcome group is also held back by a minimum-cell rule, the missing output conveys something about the reporting control. It should not be presented as evidence of zero opportunities.
These fictional numbers illustrate distinct dispositions. In practice, agree on a precedence order or a separate table for overlapping reasons. A record can be outside the time window and ineligible under a sharing rule. Counting it in two mutually exclusive categories makes the sums fail. An alternative is to keep a single final disposition for reconciliation while also recording multiple diagnostic flags inside an approved environment. That provides a defensible total and enough detail to investigate the problems.
Keep identity confidence and data permission on separate lines
A data clean room can provide a controlled environment for approved matching and analysis. Its value depends on the setup, the participating parties, and the permitted inputs and outputs. The IAB Tech Lab data clean room guidance describes use cases and safeguards for collaboration. The safeguards may include access restrictions, computation controls, and limits on outputs. Ask which rules actually apply to your setup and who can change them.
An identity match can suggest a relationship between an event and an account, depending on the identifiers, source quality, and method. It cannot establish that the same employee performed every action tied to that account. A network used by multiple businesses, a central purchasing group, or a name shared by subsidiaries can complicate interpretation. Where person-level behavior is relevant, establish the evidence necessary to support that claim and record ambiguous matches separately. A match also carries no automatic permission to activate an audience or contact a person. Establish the authorized purpose and channel eligibility independently.
For the maintenance vendor, one parent company may have several facilities in different regions. If the sale is approved by a local site, matching activity only to a parent organization may be too broad to support a claim about the facility's intent. The report can still show a parent-level association if the grouping and outcome are defined that way. Keep the site-level question open until an approved matching method can answer it. It is easier to publish a meaningful narrow result than to defend a broad claim whose unit changes from row to row.
Give the two exposure groups the same chance to be observed
For a combined-signal analysis, define external-intent-only accounts and external-intent-plus-site accounts using a common entry time and the same lookback. Treat unknown site engagement separately when coverage is incomplete. An unmatched visit is not enough to conclude that the account stayed away from the site. Decide whether your study is limited to accounts with comparable measurement coverage or whether you will report an explicit unknown group. Keep the opportunity rule and follow-up window the same across groups.
Avoid defining the combined group using visits that occurred after an opportunity was already accepted. That allows downstream engagement to masquerade as a predictor. A prospective setup can take a snapshot when accounts enter the study and observe what happens afterward. In a retrospective setup, reconstruct the snapshot from the facts available at that time. Check timestamp precision and time zones, particularly when a website event and CRM change occur on the same date. If ordering cannot be established reliably, set aside the affected cases or report a sensitivity analysis that shows whether they matter.
When presenting the rates, use the eligible account counts for each group and show accepted opportunities observed by the cutoff. Report raw counts alongside rates when permitted. If a rate rises from one comparison group to another, distinguish the difference in percentage points from the relative difference. State the observation dates, definition of acceptance, match coverage, and important changes in campaign activity. An observed gap can support a hypothesis that the combination is useful for prioritization. A controlled test is needed to estimate the effect of actually acting on it.
Plan the failure review before campaign reporting begins
Ask to see the settlement report when no conversions occur, when events arrive late, and when the reporting policy withholds a small group. These outcomes should be distinguishable to the reviewer where the governance setup permits it. A true zero can be a business result. A withheld cell is a reporting decision. An unconfirmed delivery is an operational issue. Treating all three as blank spreadsheet cells can send marketing looking for a demand problem that belongs to engineering or governance.
Review a dry run with harmless test records before relying on real campaign inputs. Include a duplicated event, an invalid event, a record that matches only at a broader level, and an event dated outside the lookback. Decide who clears an unresolved count and what evidence allows a previously rejected record to be reprocessed. Keep the original event identifier stable when correcting a delivery problem. If a true business correction changes whether an opportunity qualifies, record why and publish a revised report cut only under the agreed policy.
For a published Cortex benchmark in the future, a useful release package would include anonymized counts at each relevant stage, exposure and outcome dates, cohort definitions, the share of each group with observable outcomes, and the limitations of the permitted analysis. If aggregate outputs conceal a small group, protect the group and explain the reporting constraint at a safe level. No individual client records are needed in the public article. A method someone can understand is more valuable than an impressive-looking number whose data path cannot be explained.
Use feedback to improve the next decision
Cortex’s two-way intent approach lets a client choose relevant topics, websites, sources, and buying signals. Website behavior and reviewed outcomes from leads, customers, and won or lost deals can help refine the client’s prioritization. For this kind of measurement, document which feedback entered the model and which evidence was reserved for evaluation. Customers that the business supplies from outside the original list can inform learning, while acquisition credit still follows the campaign evidence. A reconciliation shows where the analysis has information strong enough to support a conclusion and where it has room to improve.
For a companion question, read Cortex’s guide to reporting website identity coverage. If you are evaluating a data collaboration, call Cortex Marketing AI to request a consultation and bring a small, anonymized map of where an event is recorded, sent, accepted, matched, and reported. A good first conversation can establish which count you trust today, where the losses could occur, and which next comparison would be worth running.
One useful procurement question is whether the parties can explain a mismatch without exposing individual records. For instance, ask whether an approved report can distinguish missing identifiers from a deliberate policy exclusion. If the report can show only a combined number, decide in advance which decisions that number can support. The same question belongs in the analytic handoff. A future reader should know what was measurable with the approved data and what would require a different study design.
If the team cannot reconcile totals across all stages, publish the section it can explain and label the unobserved transition. Assign a separate task to improve that boundary before the next report.
