TL;DR
There are four types of B2B intent data: topic-level, behavioral, predictive, and owned. Three of them can be bought, which means a competitor can buy the same feed and see the same account surge on the same day. Compare them on source, resolution, latency, exclusivity, consent basis, and best use before you sign anything.
Buyers cannot evaluate intent vendors sensibly because the category names are not defined consistently across the market. One provider’s “behavioral” is another’s “topic,” and “predictive” is used to describe everything from a regression model to a keyword surge. This article sets out the four types, compares them on six dimensions, and adds the dimension almost every page leaves out: whether anyone else can buy the same thing.
What is intent data?
Intent data is any evidence that an account or a person is researching a problem you solve. It is not a purchase commitment, and it is not a lead. It is an observation, or in one case an inference, about research behavior, and its usefulness depends entirely on how it was captured.
Four types make up the category:
The first three can be purchased. The fourth cannot.
The four types compared
The table below uses identical dimensions in identical order for all four rows, which is the only way a comparison of this kind means anything. Read the exclusivity column last, because it changes how you read the other five.
| Dimension Topic-level Behavioral Predictive Owned (first-party) | ||||
| Source | Content consumption across publisher networks and data cooperatives | Tracked digital actions on web properties, ads and marketing automation | Model output built on firmographic, technographic and behavioral inputs | Calls, meetings, email threads, support tickets, product usage, calendar, internal channels |
| Resolution | Usually account or domain level, rarely contact level | Contact level where the visitor is identified, otherwise anonymous | Account level, expressed as a score or ranking | Contact level, with a named person and a named account |
| Latency | Reported on a lag; describes a window rather than a moment | Immediate for observed actions, decays quickly | As fresh as the last training and feature update | Available as soon as the conversation is processed |
| Exclusivity | None. The same feed is sold to your competitors | Partial. Your own property data is yours, syndicated and partner data is not | None. The same model, or an equivalent one, is sold widely | Complete. Nobody else can buy a conversation that happened inside your company |
| Consent basis | Depends on storage and access on a third party’s site, which under PECR requires consent | Same requirement where storage or access technologies are used, including pixels and web storage | Inherits the consent basis of every input feeding the model | Engages UK GDPR, employment, and confidentiality considerations rather than PECR storage and access rules |
| Best use | Early TAM narrowing and ABM list shaping | Retargeting, session-level prioritization, form-fill follow-up | Territory prioritization when observed data is thin | Contact-level routing, expansion, and renewal defense |
Topic-level intent data
Topic-level intent tells you that people at a given company have been consuming content on a subject across a network of publisher sites. It is the broadest and cheapest form of intent, it is genuinely useful for narrowing a large addressable market, and it is the type most often oversold.
Where it comes from. Publisher networks and data cooperatives aggregate content consumption and match it back to a company, usually by IP resolution or by an identity graph.
What resolution it gives. Account or domain level. That is a real limitation rather than a temporary one. Gartner’s B2B buying research finds that buyers loop through six buying jobs, revisiting each at least once, and that 99% of B2B purchases are driven by organizational changes. If the purchase is decided by a group looping through jobs, an account-level surge tells you a building is interested. It does not tell you who inside that building is doing consensus creation.
Where it breaks. Large organizations generate topic signal constantly for reasons that have nothing to do with buying. A researcher, a student intern, and a competitor doing win-loss analysis all read the same articles. Aggregation also means the signal describes a window rather than a moment, so by the time a surge is reported, the conversation it reflects may have moved on.
Behavioral intent data
Behavioral intent is the record of actions actually taken: pages viewed, assets downloaded, ads clicked, pricing pages revisited. It is the most directly observed of the four types and, where the visitor is identified, the only purchasable type that reaches contact level.
Where it comes from. Web analytics, marketing automation, advertising platforms, and, where syndication is involved, partner properties.
What resolution it gives. Contact level for known visitors; anonymous or account level for the rest. That split matters more than most teams admit, because the portion of traffic that is anonymous is usually the larger portion.
Where it breaks. Behavior on your own property is a strong signal but arrives late in the process. Toman, Adamson, and Gomez found in Harvard Business Review that 65% of customers spent as much time preparing to speak to a rep as they expected the whole purchase to take. Most of that preparation happens somewhere you cannot see. A first visit to your pricing page is not the start of the buying process; it is a late waypoint in one that started elsewhere.
On the consent question. Where behavioral capture relies on storing or accessing information on a user’s device, the Information Commissioner’s Office is explicit that this applies regardless of the technology used, so pixels, fingerprinting, web storage, and tag-based scripts are all in scope, not just cookies.
Predictive intent data
Predictive intent is different in kind from the other three. It is not an observation. It is a model output that infers an account is likely to be in market from patterns in other data, which means it carries a confidence level and a false-positive rate that a directly observed signal does not.
Where it comes from. Firmographic, technographic, hiring, funding, and behavioral inputs, combined by a model trained on historical outcomes.
What resolution it gives. Account level, usually as a score or a ranked list.
Where it breaks, honestly. Confidence is a property of the model, not of the account. A score that is right most of the time is still wrong for a knowable share of your list, and you will not know which share until you have worked it. The cost lands as wasted outbound: sequences sent, connect attempts burned, and rep confidence in the whole intent program eroded when the third scored account in a row has no project.
None of that makes predictive intent a bad category. When observed data is thin, which is the normal state early in a territory, a model output is better than a guess. It should simply be treated as a prioritization aid rather than as evidence.
Owned (first-party) intent signals
Owned intent is the evidence generated inside your own organization: what customers and prospects said on calls, in meetings, in email threads, in support tickets, in product usage, and in internal channels. It is the fourth type, and structurally it is the only one nobody can sell you.
Where it comes from. Calls, meetings, email threads, support tickets, product usage, calendars, and internal channels.
What resolution it gives. Contact level with a named person and a named account. Not a domain that surged, but a person who said something, on a date, in a system you already own.
Why it matters for resolution specifically. The buying group has grown from an average of 5.4 people to 6.8. Gartner names consensus creation as a distinct buying job that groups return to repeatedly. An account-level score cannot tell you which of those roughly seven people is doing the consensus work. A recorded objection from a named finance stakeholder can.
The exclusivity fact. Topic, behavioral, and predictive intent are purchasable. That is a structural fact about the supply side, not a criticism. It means a competitor with a budget can buy the same feed and see the same account surge on the same day you do. Owned signals cannot be bought by anyone, because the conversation happened inside your company. That is a taxonomy fact.
A caveat that belongs here. Owned intent is not automatically compliant. It sits outside the PECR storage-and-access questions that dominate the other three, but it engages UK GDPR, employment, and confidentiality considerations of its own. The honest position is that the two categories raise different questions, and that procurement teams in the UK and EU increasingly ask the third-party question first.
fifth’s Revenue AI is the signals layer for this category, and Buyer Signals is the product term for what it produces from those sources.
Which type should you use for which motion?
No type wins outright. The right question is which motion you are running, because each motion has a different tolerance for false positives and a different need for contact-level precision.
| Motion Primary type Supporting type Why | |||
| Cold outbound | Topic-level | Predictive | You need volume and TAM narrowing before precision |
| ABM | Behavioral | Owned | Small named list, so contact-level precision pays for itself |
| Expansion | Owned | Behavioral (product usage) | The evidence already exists inside the account relationship |
| Renewal defense | Owned | Behavioral | Risk shows up in tickets and sentiment before it shows up in a surge |
How to combine types without double counting
The failure mode is straightforward. One account triggers a topic surge, a predictive score uplift and a website visit in the same week, three records land in the queue, and a rep works the same account three times.
Set precedence rules before you set thresholds:
Refresh frequency and signal decay
How long a signal stays actionable varies by type, and this is the dimension the search results almost never cover. Vendors set refresh intervals individually and no neutral benchmark exists, so the useful comparison is relative decay rather than a number of days.
Directly observed actions are actionable immediately and decay fast. The pricing-page visit is worth most within hours.
Aggregated topic surges are reported on a lag and describe a window rather than a moment. They stay directionally useful longer but were never precise about timing.
Model outputs are as fresh as their last training and feature update, which means a predictive score can be stale in a way that looks identical to a current one.
Owned signals carry their own timestamp, because a conversation happened on a date. Decay depends on the substance of what was said rather than on a refresh schedule.
This is why decay matters more than volume. A large feed of signals you cannot date is worth less than a small set you can.
Worked example one: outbound
A mid-market software company narrows a 4,000-account territory using topic-level intent, which reduces the working list to accounts showing research activity in the relevant category. Predictive scoring ranks what remains. Both are purchasable, so the assumption is that at least one competitor is working a similar list.
The differentiator is what happens next. The company checks its own systems and finds that eleven of those accounts already appear in owned signals: a support conversation, a partner introduction mentioned on a call, a demo request from a colleague eighteen months ago. Those eleven go to named contacts with a named reference point. The rest go to a sequence.
Worked example two: expansion
An industrial manufacturer runs expansion motions across a large installed base. Purchased intent is close to useless here, because the accounts are already customers and the surge signal describes a building that is already talking to the company daily.
The evidence that matters is owned: a services conversation where a second business unit was mentioned, a support thread naming a system the account is not licensed for, a meeting where a new site came up in passing. Approved reference for this framing is Atlas Copco (industrial manufacturing; RevOps and voice AI). None of that is purchasable, and none of it reaches the account owner unless something reads the conversations and routes it.
Conclusion
Four types, six dimensions. Topic, behavioral, and predictive intent all have legitimate uses and all share one property: they are for sale. Owned intent is the only category on the taxonomy that a competitor cannot acquire, because it is generated by conversations inside your own organization. That is a fact about the structure of the supply, not a claim about accuracy.
FAQs
Q1. What are the different types of intent data?
There are four. Topic-level intent infers interest from content consumption across publisher networks. Behavioral intent records tracked digital actions. Predictive intent is a model output that infers readiness. Owned or first-party intent comes from your own calls, tickets, and product systems. The four types compared sit in the table above.
Q2. What is topic-level intent data?
Topic-level intent data infers that an organization is researching a subject from content consumption across publisher networks and data cooperatives. Its resolution limit is important: it usually resolves to an account or a domain rather than to a contact, so it tells you a company is interested without telling you which person is driving the evaluation.
Q3. What is behavioral intent data?
Behavioral intent data records actions actually taken, such as pages viewed, assets downloaded, and ads clicked. The distinction most pages blur is that topic-level intent is inferred from consumption across third-party networks, while behavioral intent is a direct observation of a tracked action, usually on properties you or a partner control.
Q4. What is predictive intent data?
Predictive intent data is a model output rather than an observation. It infers that an account is likely to be in market from patterns in other data, which means it carries a confidence level and a false-positive rate that a directly observed signal does not. Confidence is a property of the model, not of the account.
Q5. Which type of intent data is best for outbound?
It depends on the motion rather than on a single winner. Cold outbound usually starts with topic-level intent for TAM narrowing, supported by predictive scoring. ABM, expansion, and renewal defense lean on owned and behavioral signals because they need contact-level resolution. The motion table above sets out the pairings.
Q6. How do the types differ in refresh frequency?
Vendors set refresh intervals individually and no neutral benchmark exists, so compare relative decay. Observed actions are actionable immediately and decay fast. Topic surges are reported on a lag and describe a window. Model outputs are as fresh as their last update. Decay matters more than volume, because an undatable signal cannot be prioritized.
Q7. Which intent data types can be sourced internally?
Owned or first-party intent only. The internal sources are calls, meetings, email threads, support tickets, product usage, calendar, and internal channels. The other three types depend on data captured outside your organization, which is why they can be purchased by anyone, including your competitors. See first-party intent signals in depth.
