The Intent Data Contract: What Every Signal Should Explain
Buyer intent data becomes more useful when every signal arrives with an explanation of where it came from, what it represents, and what your business may do with it. A high score alone cannot answer those questions. Before connecting a new source to a sales workflow, create a short data contract that travels with the signal. This article proposes a practical contract for evaluating custom buyer intent. It is an operating framework, not a report of measured Cortex customer results.
The immediate buyer problem is simple. Marketing wants more relevant opportunities, sales wants a dependable queue, and operations wants fewer exceptions. Those goals can conflict when the same record means different things to different teams. A researcher can become a supposed decision maker, an account match can become a named person, and a delivery date can be mistaken for the date of the original behavior. A written contract makes those transformations visible before money or attention is committed.
An October development worth understanding precisely
Google’s privacy policy, effective October 16, 2026, explains that users may allow certain Google services to access information from other Google services for personalization, including ads. It also distinguishes that processing from sharing personally identifying information with advertisers. Read the primary policy and its personalization section. This development does not establish that Cortex receives Gmail, Drive, or Photos data. It illustrates why a buyer should distinguish what a platform processes internally from what an advertiser or external provider can actually obtain.
Our practical interpretation is that broad phrases such as “AI-powered signals” are insufficient procurement descriptions. Ask which data reaches your business, at what level, under which permission, and with which restrictions. An announcement about one company’s products does not expand another company’s access. A provider should explain its own source relationships without borrowing the capabilities of a large platform. That distinction helps your team evaluate an offer without either dismissing useful information or assuming access that has never been demonstrated.
Start with the unit the signal describes
The first field in the contract is the unit of observation. Is the record about an individual, a browser, a household, an account, a location, or an aggregate audience? Write the answer in ordinary language. If it is an account, say that the evidence concerns the account. If it is a browser, do not silently promote the record into a named individual. Even a technically successful match may still leave uncertainty about who performed the behavior or why they performed it.
Consider a hypothetical industrial supplier evaluating research around packaging equipment. Activity associated with a company could reflect an engineer investigating an upgrade, a student visiting from a shared network, or a purchasing team comparing alternatives. The supplier may reasonably use the evidence to prioritize account research. It should not tell a particular employee that the employee personally searched for a machine. The operating action should fit the strength and resolution of the evidence. More precise language makes the resulting outreach more credible.
Separate four dates that often become one
A useful record distinguishes the behavior date, the provider’s processing date, the delivery date, and the date your team acted. These dates support different decisions. A signal delivered today may describe activity from an earlier period. A salesperson may wait several more days before using it. If those delays are combined into one “fresh” label, a manager cannot tell whether performance reflects the source, the transfer process, or the team’s response time.
The contract should state whether timestamps are exact, rounded, inferred, or unavailable. It should also identify the time zone and define a rule for late arrivals. For a service with immediate demand, old evidence may need to leave the priority queue quickly. For a long equipment purchase, older research can still matter when interpreted alongside newer events. Neither rule is universal. The correct time horizon depends on the decision you are trying to support and the sales cycle you actually serve.
Describe the source without promising unlimited coverage
Cortex’s proposed starting point is customer choice: relevant websites, topics, sources, and buying signals should reflect the market the customer wants to reach. That choice is valuable only when source availability and meaning are explained. A buyer selecting a competitor website should ask what “selecting” means operationally. Does the source contribute observed research, a topic association, an aggregate indicator, or another licensed signal? Does coverage vary by geography or device? The contract should document the answer actually supplied for that implementation.
Avoid interpreting a source preference as permission to observe every visitor to any website. The preferred source map and the obtainable dataset are separate objects. Keep both. The first expresses your commercial hypothesis; the second describes the evidence available to test it. When they do not match, you can revise the hypothesis, choose an available alternative, or stop the use case. This is more useful than accepting a long list of domains without knowing how any of them contributes information.
Specify what the score means before specifying a threshold
A score might order records by relative relevance, estimate the probability of a defined outcome, or summarize several recent behaviors. Those meanings are not interchangeable. If the score is a ranking, a score of eighty does not automatically mean an eighty percent chance of purchase. If it is a probability, ask which outcome and observation window it predicts. If it is a composite, request the inputs and the treatment of missing information at a level the provider can responsibly disclose.
Your contract does not need the provider’s complete intellectual property. It does need enough information to use the output sensibly. Record the intended action, relevant limitations, version identifier, and review owner. A sales queue based on a ranking should be evaluated against the same queue capacity over time. A probability forecast needs calibration checks. A behavior summary may require human interpretation. Without this distinction, teams can debate whether a number looks impressive while never establishing what would count as success.
Make permissions part of the workflow
Source availability, identity matching, and permission to activate data are separate checks. A record may be useful for approved aggregate research without being eligible for a specific advertising upload. Google’s Customer Match policy, reviewed October 26, 2026, limits uploads to information collected in a first-party context and imposes additional requirements. A purchased intent list should not be presumed eligible. Confirm the actual collection context, account eligibility, intended use, and current platform rules before planning that activation.
For internal operations, maintain a destination register. For each intended destination, record the permitted purpose, authorized fields, responsible owner, and removal process. An analysis workspace may need only pseudonymous account keys and aggregated outcomes. A salesperson may need a narrower set of approved business contact details. An advertising destination may reject the dataset entirely. A clear register lets a team benefit from useful research without assuming that every field should follow every record into every connected system.
Write the return-signal contract too
Two-way intent data depends on feedback, so your outgoing information deserves the same attention as incoming signals. Define “qualified,” “customer,” “lost,” “pending,” and “outside our service area” before sending outcomes. Include the event date, the date it became known, and whether the label is final. A deal that is still open should not become a failed opportunity merely because a reporting month ended. A rejected record should carry the reason that made it unsuitable, not just a generic negative label.
Customers who were never in the original list can still provide learning information when the implementation permits it. Preserve their origin. Their behavior may inform which patterns fit your offer, but their presence does not prove that the intent program acquired them. Keep learning eligibility and acquisition attribution in different fields. This allows your customer-specific model to benefit from broader experience while maintaining an honest account of where revenue came from. The distinction is especially useful when several marketing channels operate simultaneously.
Use a concrete acceptance test
Before a full rollout, ask the provider to walk through a small authorized sample from source to decision. Choose examples that include a missing timestamp, an unresolved identity, a duplicate account, a late outcome, and a record that must not be activated. Check whether the explanation survives every handoff. The goal is not to create a polished demonstration with only perfect records. It is to learn whether ordinary uncertainty remains visible when the information reaches the person expected to act.
In a hypothetical supplier pilot, one record might correctly reach an account-research queue, another might remain in an aggregate analysis, and a third might be excluded because the service region is unsupported. Those different outcomes are evidence that the workflow respects its rules. They are not failures merely because they produce fewer sales tasks. A useful acceptance test demonstrates correct routing as well as correct matching. Record the expected action before testing so that a surprising result cannot be redefined as success afterward.
Give changes an owner and a stopping rule
A data contract is a living operating document. Sources can change, definitions can be revised, and your own product can move into a new market. Assign one owner to approve material changes and another to confirm downstream consequences. Keep a dated change log. If a source disappears, do not quietly replace missing observations with zero interest. Mark the coverage change and decide whether affected records should pause, retain their last known status, or enter a different review process.
Define stopping conditions in advance. Examples include an unexplained change in the unit of identity, an unresolved permission question, or a missing timestamp that makes the intended action unreliable. A stopping rule does not mean abandoning buyer intent data. It means pausing a specific use until its assumptions are restored. This keeps a local problem from becoming a company-wide reporting error. It also gives both the customer and provider a concrete question to resolve instead of an unstructured argument about quality.
Make corrections reversible and traceable
Plan how a correction moves through the same workflow as the original record. If an account match is later withdrawn, the sales queue, analytical dataset, and permitted downstream destinations should receive an understandable update. Assign a stable record identifier so that the correction updates the intended observation instead of creating another apparent prospect. Keep the date and reason for the correction where retention is permitted. This also helps customer support: a provider can investigate a specific disputed record without requiring the customer to send an entire database. Test one correction before launch and confirm that the person using the queue can recognize the revised status.
Turn the contract into a buying decision
A provider comparison can use the same contract for every candidate. Ask each provider to explain the observation unit, source coverage, dates, score meaning, feedback requirements, and allowed outputs for your actual use case. Then compare the clarity of the answers and the operational fit. A provider with broader coverage may still be less useful if your team cannot interpret its records. Another may support a narrower but better-defined workflow. The best choice depends on the decision you need to improve.
For Cortex, the relevant question is how custom buyer intent and customer-specific feedback would work for your niche, geography, offer, and sales cycle. Scoring can begin about twenty-four hours after list subscription according to the supplied Cortex brief, but starting a process is different from demonstrating reliable learning. Ask what must be configured first and what evidence will establish progress later. Proprietary learning should mean that your outcomes shape your model. It should not be interpreted as a promise that every underlying contact is exclusively yours.
Questions to take into the meeting
Ask what happens when a record cannot be matched confidently. Ask how withdrawn or corrected data reaches downstream systems. Ask whether historical scores remain available after a model update. Ask which outcome fields are essential and which are optional. Finally, ask the provider to show one example of an action the system should refuse to recommend. These questions reveal whether the implementation has a usable boundary around its capabilities, which is often more valuable than a broad feature list.
No anonymized Cortex event-and-outcome dataset was available for the analysis in this article, so no conversion lift or accuracy benchmark is asserted. A documented contract is the preparation that makes later measurement trustworthy: consistent units, dates, definitions, and permitted uses allow meaningful comparisons. To map those requirements to your business, book a Cortex consultation. Bring one real decision your team struggles with and a sanitized example of the records available today. Start with a workflow you can explain, then evaluate whether the evidence improves it.
