Intent Data: What Each Category Can Actually See

Intent data is any observation that a business is moving towards a purchase, sold as a data product. That definition is broad on purpose, because the category is broad in practice: four quite different observation methods are sold under the one name, they watch four different surfaces, and they resolve to different things. A source is not better or worse than another source in the abstract. It can see what its surface exposes and nothing else, and that is the whole of the evaluation.

The disclosure belongs here rather than at the bottom, because it should change how you read the rest. We build Keaz Signals, which sells ecommerce leads with live buying signals attached. That puts us inside one of the four categories described below, and gives us an interest in how you weigh the other three. The section naming where the other categories beat us is not a courtesy; it lists cases where we are not a close second.

The B2B qualifier adds one thing and it is easy to miss. In consumer intent, the unit of observation and the unit of purchase are the same person. In B2B they usually are not: a source observes an individual, and then attributes that observation to the organisation the individual belongs to, because the organisation is what you can actually sell to. Almost every difficulty in this category descends from that one step. The attribution is an inference, its confidence varies enormously, and nothing in a delivered record tells you which kind you are holding.

So the useful question is never how much intent data a source has. It is three narrower ones: which surface can this source observe, at what resolution does it resolve an observation, and how long after the fact does the record reach you. This post is a map of the four categories against those three properties, written by a participant, with the parts we cannot support left out and said so.

Four observation methods are sold as intent data, and they differ by which surface they can watch: a surface you own, a surface someone else owns, a surface a company exposes technically, or a surface a company performs in public. Everything else about a source follows from which of those four it sits on.

1. First-party intent. Behaviour on surfaces you control: your website, your pricing page, your documentation, your product itself. You see the raw event and you see it immediately. This is the only category where you are not buying somebody's inference.
2. Third-party content intent. Reading and research behaviour observed across publisher networks and data co-ops, then aggregated by topic and attributed to an organisation. This is what most people mean when they say they bought intent data. It is the only category that can tell you about companies that have never touched anything of yours.
3. Technographic change. What a company runs, detected from the outside, and more usefully what it started or stopped running. A new tag on a site, a platform migration, a tool that appeared last month. The signal is not the state but the change in it.
4. Public activity. What a business does openly and datably: launching a product, starting to run ads, rebuilding a storefront, changing its posting cadence, sending newsletters again after a quiet quarter. No inference about who read what. An action, performed in public, with a date on it.

A fifth thing is often shelved next to these and is not intent data at all: firmographic and registry attributes, headcount bands, funding events, hiring activity. Those describe a company's situation rather than its behaviour. They are legitimate targeting inputs and they belong in the segment definition, not in the timing decision. Treating them as intent is how a list of plausible companies gets mistaken for a list of companies doing something this week.

Very few products sit purely in one category. Most combine two or three and present the result as a single score, which is convenient and also where the information about provenance gets lost. If a source will not tell you which surface a given record came from, you cannot evaluate it, because the four categories fail in completely different ways.

Each category has a blind spot that no amount of investment removes, because the blind spot is a property of the surface rather than of the vendor's engineering. Knowing the four blind spots is more useful than any comparison of coverage claims, and unlike a coverage claim you can check it yourself.

First-party intent cannot see anyone who has not arrived. It is the highest-confidence data in the whole category and it is structurally limited to people who already found you. For a company with traffic, that is a strong start-here. For a team whose problem is that nobody knows they exist, it is an empty room, and no configuration fixes that.

Third-party content intent cannot see the person, only a population. An observation is made somewhere in a network and then attributed upward to an organisation, so what arrives is a statement about a company: interest in this topic is elevated relative to a baseline. It is genuinely the only method that finds companies before they touch you. It is also the method where the delivered record is furthest from the underlying event, and where the word intent is doing the most work.

Technographic change cannot see motive. A company installed a tool. That is a fact, and it is compatible with a new initiative, a consultant tidying up, a trial nobody continued, or a tag left behind by an agency that left last year. The observation is solid and the story you attach to it is yours. That is a structural limit of observing from outside, not a claim that anyone's detection is unreliable.

Public activity cannot see anything a business does privately, which for most B2B companies is nearly everything that matters. It works to the extent that the business performs its operations in public, and most do not. This is our category, and the limitation is the reason the product is narrow rather than an accident of where we started.

There is a property that cuts across all four and decides more than the category does: whether the source stores a series or a snapshot. An observation is only an event if somebody was watching before it happened and kept the earlier reading. A source that overwrites its own records can describe a state accurately and cannot tell you when the state changed, which is the only thing a timing decision needs. We have written that argument out at length in build a lead scraper, or buy the panel, where it decides a build-versus-buy question rather than a purchase one.

Company-level intent was designed for a buying situation that an online store does not have. The method assumes a large organisation with many employees, a buying committee whose members research separately, and a purchase big enough that elevated topic interest across several people means something. Take those assumptions to a five-person store and each one fails individually.

Start with the aggregation itself. Attributing observations to an organisation only produces a signal when there are enough observations to lift above a baseline. A company of four thousand people generates that traffic. A store run by a founder and two part-timers does not, so the honest output for most of the market you sell into is silence — and silence in an intent product is indistinguishable from a company that is genuinely quiet.

Then the committee. The premise of account-based intent is that several people at one company are independently investigating, and that the pattern across them is more informative than any one of them. At a small store the company and the person are the same thing. There is no committee to detect, and the mechanism that adds value in enterprise adds nothing here.

And the resolution is wrong in a specific way. Company-level intent tells you an organisation is warm on a topic. Selling to stores, the question is rarely which store is warm on a topic in the abstract — it is which store did something last week that your service is the answer to. Those are different questions, and a source built for the first will not answer the second no matter how much of it you buy.

The compensation is that stores are unusually legible from outside. A store's advertising runs where anyone can look at it, its storefront changes visibly, its product pages appear and disappear, its social accounts and newsletters carry dates. What an enterprise does behind a login, a store does on a public web page. So the category that is weakest for this audience, company-level content intent, is weak precisely where the category that suits it, public activity, is strong. That is a difference in observation surface, not a claim that one method is more rigorous than the other.

The adjacent version of this argument, applied to bought contact databases rather than to intent products, is in what a sales intelligence platform knows, and what it cannot.

A deterministic signal is a record of something that happened, tied to a specific entity. A probabilistic signal is an estimate that something is happening, produced by a model. They are not two grades of the same measurement — they are different kinds of statement, and a percentage next to the second one does not turn it into the first.

The distinction matters operationally rather than philosophically, because the two fail differently and the failures need different handling. A deterministic signal that is wrong is usually wrong about timing or relevance: the event happened, and it did not mean what you hoped. A probabilistic signal that is wrong may have no event underneath it at all. The first produces a message that is slightly off. The second produces a message about something that never occurred, which is the version a recipient notices.

• Ask what the underlying event was, in one sentence, without a score in it. If the answer cannot be given that way, the record is a model output rather than an observation, and that is worth knowing before you write copy that references it.
• Ask what the entity is. A person, a company, a domain, a device, a household of employees inferred from a network address. Each of those is a different confidence level wearing the same label in the export.
• Ask for the date of the event, not the date of the record. Those are frequently different, and the gap between them is the part of the product that determines whether your message arrives while the thing is still true.

Probabilistic is not a criticism. Modelled intent is the only way to know anything about a company that has never interacted with you, and for a large-account motion that reach is worth a great deal of imprecision. The error is not using it. The error is writing an email whose first line asserts, as fact, something the source only estimated — because the confidence interval stays in the data warehouse and the assertion goes to a human being who knows whether it is true.

That is also the practical reason we only write from events we can point at. Our sales agent drafts per-store copy from the signal itself plus a knowledge base you write once, so the sentence in the email traces to a dated observation rather than to a score. What decides who enters a sequence at all is a separate question, covered in what a sales engagement platform does.

We sit entirely in the fourth category and nowhere else. We watch 4.1 million Shopify and WooCommerce stores across Europe and North America, re-scraped weekly, as of August 2026. That number moves, so treat the date as part of it. Because each weekly read is stored rather than overwritten, what we hold is a store plus what changed about it recently, which is the series-versus-snapshot property described above.

The events we currently detect are public actions with dates: new Meta ads running, active ad count rising, a product launch, an email marketing tool installed, social growth on Instagram and TikTok, newsletter activity, and storefront or site changes. The current list is the buying signal catalogue, and what sits underneath it is described on the ecommerce leads database page. Read both before a trial, because they are also the honest place to discover that no event we detect corresponds to what you sell.

Contacts arrive with the store, including founder addresses that are not publicly listed, so there is no separate enrichment step to buy. Segments are assembled with country, platform, follower and other conditions, with a live audience estimate as you narrow them. Sending runs through your own Instantly workspace, so domains and mailboxes stay yours and replies sync back. Access is capped per market, which is the mechanism that stops the segment you buy being sold to four of your competitors in the same week.

Several things are not there, and in an article about categories of intent data it matters which ones. Registry and firmographic triggers are marked planned on our own site, which means they do not exist today. Hiring signals are marked planned as well. There is no CSV export: leads move into campaigns and stay there. There are no LinkedIn signals and no LinkedIn sending, and no plans to add either. So three of the four categories in this post are things we do not sell, and one of them, first-party intent on your own site, is something we will never sell because it is not our surface to watch.

On data protection our position is deliberately narrow and this article does not widen it: we are GDPR-conscious by design, and responsibility for what you send stays with you. We will not claim more than that, and nothing here is legal advice about whether a given intent source may be used in a given jurisdiction. That is a statement about where responsibility sits, not a claim about anyone else's compliance.

Five situations make a different category the right purchase, and in each one we are not a close second. If any of them describes you, buy that instead and stop reading comparisons.

1. Your buyers are not online stores. Company-level content intent exists because most B2B buyers do their evaluating privately, and that is the correct tool for a private buying process. We watch a public surface, so for a manufacturer, an accountancy or a hospital group we see nothing at all. Not thin coverage — nothing.
2. You already have meaningful traffic and are not using it. First-party intent is the highest-confidence category in this article and you own the surface it runs on. If your pricing page gets real visits and nobody is acting on them, buying an external source first is spending money to reach strangers while ignoring people who arrived on their own.
3. You sell on a platform or in a region we do not cover. Ours is Shopify and WooCommerce in Europe and North America. Magento, BigCommerce, Wix, Squarespace, PrestaShop, Shopware, headless builds, marketplace sellers with no storefront: none of that is in our pool, and neither is Latin America, Asia, Africa or Oceania.
4. You need the data in your own systems. We do not export. If your CRM or your warehouse is the system of record, or you are building something on top of this data, that rules us out and it should. This is a straightforward loss for us and there is no version of the argument where we win the point.
5. Your motion is LinkedIn-first. We have no LinkedIn signals, no LinkedIn sending, and no plans to add them. If the channel that works for you is a connection request rather than an email, an intent source that only feeds email is the wrong shape regardless of how good the events are.

A sixth case is not about category at all. If your total addressable market is a few hundred stores, no intent product is warranted. You can watch a few hundred businesses by paying attention, and a subscription that tells you which of them moved this week is solving a problem you do not have.

There is also a limit that applies to every category equally. Better timing does not create capacity to answer replies. The clearest sign that any of this is worth money is paid sales capacity that is not fully booked. If your calendar is filled by referrals, you probably do not need an intent source yet, and that is true of ours as much as anyone's.

An article with this title is expected to end in a ranking of providers with accuracy percentages beside them. There is not going to be one, and the reasons belong in the text rather than in a note at the bottom, because the refusals are the most useful part of the evaluation we can offer.

We are not going to publish accuracy, coverage or freshness figures for any intent data vendor. Not because they would be unflattering, but because we have not measured them and no independent party has assembled the population you would need to. A number of that shape, produced by a competitor, is an invention with a decimal point on it, and once it is in print it gets quoted back for years.

We are also not publishing one for ourselves. We are not going to tell you what share of stores in your segment fire a signal in a given month, how often a detected event corresponds to a real budget, or what percentage of our own weekly reads complete on schedule. We could design a test for the last one; we have not run it, so the figure does not exist here. Stating it anyway would make this post feel more rigorous and be less true.

There will be no price-per-record table either. What each category meters is not the same thing: content intent is usually sold by topic and audience, technographic data by record, first-party tooling by traffic, and our own access by market. Lining those up in one column would be arithmetic dressed as a service to the reader.

And nothing here is a promise about outcomes. We will not quote a reply rate or a deliverability result for a signal-led send. A reply rate is a property of your offer, sent to your list, from your sending setup, and a vendor controls one of those three at most. A figure produced under someone else's conditions is not evidence about yours.

What survives once the unmeasured claims are stripped out is the structural map: four surfaces, four blind spots, one question about series versus snapshots, and one about resolution. You can check every part of that yourself, which is more than you can do with a percentage. Where we compare ourselves against named alternatives, including the cases where we recommend the other product, that work lives on the comparison pages so it can be kept current in one place, rather than fossilised in a blog post.

You can evaluate any intent data source in an afternoon without trusting a single published figure, including ours. Six questions do it, and all six are answerable from a trial export and a conversation with whoever is selling it.

1. Which of the four surfaces does this record come from? If the answer is a blended score with no provenance, you cannot evaluate what you are holding, and everything after this question is guesswork.
2. What is the underlying event, stated in one sentence with no score in it? Write the sentence down. If you would not put it in an email to the recipient, it is not going to survive being written into one.
3. What entity does it resolve to, and is that the entity you sell to? Company-level records against small businesses are the single most common mismatch in this category.
4. How old is the event when it reaches you, as distinct from how old the record is? Ask for both dates on the same row. Some sources cannot produce the first one at all, which is itself the answer.
5. Is this a series or a snapshot? Ask whether last month's reading still exists. If it does not, the source can tell you what is true and never when it became true.
6. Take twenty records and check them by hand against the open web. Twenty is enough to find out whether the events are real, and it is a measurement you took, which beats every number in every vendor's deck.

That sixth step is the one people skip and the only one that settles anything. It is also, deliberately, the test we would like you to run on us. If you sell to Shopify or WooCommerce stores in Europe or North America, there are 1,000 free leads on signup, which is enough to find out whether your segment exists in our pool and whether anything fires in it often enough to matter, before any money changes hands. What it costs after that is on the pricing page, and the events themselves are listed in the buying signal catalogue. We would rather you check twenty stores by hand than take our account of the category on trust.

• Keaz Signals product pages, retrieved 5 September 2026: /ecommerce-leads-database, /signals, /sales-agent, /compare and /pricing. Source for the detected event list, the segment builder, the per-market access cap, the Instantly sending model, the export limitation, the LinkedIn position, and the registry, firmographic and hiring triggers marked planned rather than shipped.
• Our own pool figures: 4.1 million Shopify and WooCommerce stores under watch across Europe and North America, re-scraped weekly, as of August 2026. These numbers move; the date is part of the number.
• 1,000 free leads on signup, as of August 2026.
• Related posts on this blog: what a sales intelligence platform knows, build a lead scraper, or buy the panel, what a sales engagement platform does and what sales prospecting tools actually do.
• The four-category map and the blind spot attributed to each are a structural argument from how each observation surface works, not a survey of vendors and not a measurement. No vendor is named, ranked or characterised anywhere in this article.
• No accuracy rate, coverage figure, freshness percentage, signal-firing rate, price comparison or reply rate is cited anywhere in this article, for any vendor including ourselves. That is deliberate, and the reason is in the section on what we are not going to tell you. We did not run those measurements, so we do not report them.