The announcement is a starting point, not a result

Attain's November 17 announcement names it a data partner for The Trade Desk's Open Measurement, planned for later this quarter. Permissioned purchase observations are an input, not proof of incremental sales. The following is original practical analysis, not a product test.

For a marketing leader, the useful question is what kind of decision a new measurement input can support. A transaction can strengthen the evidence that a purchase occurred within an observed population. It does not, by itself, reveal whether the same person would have purchased without advertising. Those are different questions requiring different evidence. Confusing them can turn a more detailed report into an unjustifiably confident budget recommendation. Before examining a dashboard, write down the claim you intend to make and the alternative explanations that could invalidate it.

This article proposes a working method for a fictional retail campaign. It does not assert that the announced framework follows these procedures or has passed these tests. The central discipline is to keep observation, comparison and causal interpretation separate. Teams can prepare definitions and questions while a service is still planned, without describing it as available, assuming access in a particular country or assigning it performance that has not been demonstrated. Preparation should make a later evaluation sharper, not become an advance endorsement of an untested implementation.

Define the purchase before measuring its relationship to media

Begin with an outcome contract: what counts as a purchase, which products qualify, when the event is recorded and how cancellations or returns change the total. A paid order, a shipped order and a retained purchase are not interchangeable. Decide whether the unit is a person, household, transaction or account, and keep that unit consistent throughout the analysis. If one customer places several orders, an order count and a buyer count answer different questions. Neither should quietly replace the other because it produces the more attractive result.

Then define the population you can actually observe. Record who is eligible for inclusion, what proportion can be linked to the selected measurement process and which outcomes remain outside coverage. A missing observation is not automatically a nonpurchase. Equally, a successful link does not establish that the linked subset represents everyone reached by the campaign. Ask how observable and unobservable groups differ in ways relevant to the decision. If you cannot answer, preserve that uncertainty rather than filling the gap with a convenient assumption of completeness.

Finally, fix the outcome window before reading results. A short window may miss delayed decisions, while a long window may include more unrelated influences. Choose a window that fits the task and explain its trade-off rather than searching several windows and reporting only the strongest. Keep a dated definition and a record of changes. If a correction becomes necessary, show what changed and rerun comparable periods where possible. Measurement credibility depends as much on stable definitions as on the apparent precision of the transaction record.

Permission and analytical usefulness require separate checks

Permissioned data deserves careful handling, not a shortcut around purpose review. A team evaluating a measurement proposal should understand the permitted purpose, recipients, retention arrangements and available withdrawal processes at the level appropriate to its responsibility. This is a practical review agenda, not legal advice or a statement that a particular configuration meets local requirements. Ask the relevant specialists to assess the actual arrangement before transferring or using data. The existence of permission for one use does not justify inventing broader authorization for a different workflow.

Keep the analysis proportionate. An aggregate campaign question usually does not require a marketer to inspect identifiable purchase histories. Specify the minimum fields and level of aggregation needed to answer the question, and avoid collecting extra detail simply because it might be interesting later. Give each output an intended audience and an owner. A useful report should make uncertainty and coverage understandable without exposing people unnecessarily. More granular access can increase operational risk while adding little value to the budget decision the team actually needs to make.

Separately examine whether the data can answer the chosen business question. A permission record does not establish representativeness, stable matching or balanced comparison groups. Likewise, technically clean observations do not establish that a proposed use is permitted. Put these checks in separate review fields so that success in one cannot conceal a gap in the other. If either remains unresolved, reduce the scope or postpone the proposed analysis. Do not treat a polished procurement presentation as a substitute for a documented decision about both suitability and permitted use.

Separate observed association from an incremental effect

Suppose a report connects campaign exposure with later purchases. That association may be informative, but people who saw the campaign could differ from those who did not before the campaign began. Existing customers may be easier to reach, shoppers may already be searching, or promotional timing may coincide with stronger demand. These are possible explanations to investigate, not claims about the announced service. A measurement plan should state how it addresses such differences rather than assuming that a transaction occurring after exposure was caused by that exposure.

A suitable randomized comparison can strengthen a causal interpretation when assignment, delivery, outcome collection and analysis are appropriately designed. It still needs scrutiny: did the groups remain distinct, did other channels contaminate the comparison, and was observation comparable? Where randomization is not feasible, an observational comparison requires explicit assumptions and a clear account of what remains uncertain. Label it accordingly. Do not imply that adding a new outcome source automatically solves selection bias, interference between groups or changes in the wider market during the measurement period.

Original analytical diagram separating observed transactions from incremental impact, with matched records and attribution windows on one side and comparison design and uncertainty review on the other

Original analytical illustration: Observed transactions ≠ Incremental impact. Matched records and attribution windows address observation; comparison design and uncertainty review address an incremental interpretation. Keep the two questions separate. This is a proposed framework, not a product screenshot or a research finding. Context: the primary announcement.

The illustration separates two questions deliberately. A team should establish what was observed through the matched record and attribution window, then examine the comparison design and uncertainty needed for an incremental interpretation. Only afterward should it consider changing a budget. These are practical review distinctions, not a universal statistical method. Their value is preventing a visually persuasive chart from skipping the questions that determine whether a recommendation is warranted. If a check fails, the appropriate next action may be a better test rather than a larger campaign.

A hypothetical calculation that keeps denominators visible

Consider an invented experiment with 10000 eligible units assigned equally: 5000 to a treatment group and 5000 to a comparison group. Assume the same purchase definition and observation window, with outcomes observable for every unit. Suppose 300 treatment units and 250 comparison units purchase. The observed rates are 300 / 5000 = 6% and 250 / 5000 = 5%. The difference is one percentage point, and the relative difference is (6% - 5%) / 5% = 20%. These expressions describe different scales of the same numerical contrast.

Applying the comparison rate to the treatment group gives 5000 × 5% = 250 expected purchasers under that simple benchmark. The observed excess is 300 - 250 = 50 purchasers. It is not 300 incremental purchasers. Even the estimate of 50 needs the experiment's design and execution to support a causal reading, together with an uncertainty assessment. We are showing arithmetic, not reporting a statistically established effect. A final report should include an appropriate interval and explain its method; it should not manufacture an interval from missing assumptions merely to appear complete.

Now suppose the campaign cost is 6000 currency units. Dividing by all treatment purchasers gives 6000 / 300 = 20 currency units per observed purchaser. Dividing by the estimated excess gives 6000 / 50 = 120 currency units per estimated additional purchaser. The two figures are not alternative labels for the same metric. If the excess estimate is unstable or compatible with no effect, the second ratio is also unstable and may not be useful for a decision. Do not call either calculation profit without contribution margins and the relevant additional costs.

Stress-test coverage, timing and the story told by the report

Revisit the hypothetical example with a practical question: what if some outcomes are missing? Do not simply retain the original denominator while silently treating the missing purchases as zero. Nor should you drop unobserved units without considering whether missingness differs between groups. Show the eligible count, observable count and exclusions separately. If observation changes after assignment, examine how that affects the interpretation. The answer may require redesign or a qualified estimate; the existence of a larger dataset is not permission to ignore a smaller but systematically different missing group.

Test sensitivity to the choices that genuinely matter. Returns may arrive after the initial reporting date, product categories may be defined inconsistently, or several records may describe the same transaction. Create reconciliation rules before comparing channels. A channel that appears efficient because its outcomes are counted sooner should not automatically receive more investment than one measured on a different clock. Keep an initial result and a later settled result distinct. Explain when each can be used rather than continuously replacing history without a visible record of revision.

Also inspect the narrative around the numbers. A report might correctly calculate observed purchases while its headline claims caused sales. A presentation might show a relative increase without the underlying rates, or omit the population whose behavior was never observed. Review charts, captions and executive summaries together with the calculations. Ask a colleague unfamiliar with the project to state what they think the result proves. If their interpretation exceeds the evidence, repair the communication before it reaches a budget meeting. Accurate arithmetic cannot rescue an inaccurate claim.

Build a bounded evaluation before changing investment

Select one decision, one campaign context and a manageable evaluation period. Write the business question, eligible population, outcome definition, comparison strategy and stopping conditions before results arrive. Assign separate owners for data quality, methodological review and the investment decision, even if a small team has people holding several roles. Record disagreements instead of averaging them into a vague confidence score. The goal is a decision record that another reviewer can understand, including why the team chose a particular design and what it deliberately left outside scope.

Ask potential providers for the information needed to evaluate the proposal without presuming the answer. What will actually be available at launch? Which populations and outcomes can be assessed? What documentation supports the method? How are revisions and exclusions explained? These are due-diligence questions, not features we claim the announcement promises. Avoid a pilot whose success depends on unconfirmed access or functionality. While availability is planned, preparation can include a synthetic-data exercise and a draft reporting template without representing either as a live product evaluation.

Before increasing spend, agree what strength of evidence would justify the change and what result would lead to no change. Consider operational capacity, margin and customer experience alongside the measured outcome. A positive signal in one narrow setting does not guarantee that a larger campaign will behave the same way. Expand in controlled steps and retain the ability to stop when assumptions no longer hold. Measurement should improve decisions under uncertainty, not eliminate uncertainty from the language of the recommendation while leaving it unresolved in the underlying analysis.

The strategic takeaway: stronger observations need disciplined claims

My practical recommendation is to prepare an evidence checklist before buying a more impressive measurement narrative. Separate four statements: a purchase was observed, it was linked to a campaign context, a comparison suggests a difference, and the design supports an incremental interpretation. Each statement requires additional support. Keeping them distinct allows a team to use valuable transaction information without overstating what it establishes. It also makes a disagreement productive: reviewers can identify the exact unsupported step instead of arguing broadly about whether a platform should be trusted.

The announcement creates a reason to ask sharper questions, not to rewrite a budget on the strength of a future offering. Keep the release date, planned availability and actual evaluation status separate. Do not assume that a method, data population or permission arrangement transfers unchanged to another market. The responsible next step is a bounded assessment once the required information and access exist. Until then, a well-specified test and an honest list of unknowns are more useful than an early claim that measurement has finally become objective or complete.