Predictive intent data is a ranking to test, not

Abmatic AI's 2026 buyer's guide, written by Jimit Mehta and published 27 April 2026, gives the cleanest definition. Predictive intent data is the output of a model that estimates which accounts are likely in market for a category before they have declared explicit interest. The word to hold onto is "estimates".

Abmatic splits intent into three kinds. Third-party intent observes topic surges across a publisher network. First-party intent captures behavior on the owned properties of the company itself. Predictive intent infers. It combines historical patterns, observed signals, firmographic features and machine-learning models to surface accounts that look statistically likely to buy.

That makes it a different kind of object from the other two. A first-party record says a company did something on your site. A predictive score says a model thinks a company resembles others that bought. One is a record. The other is a guess presented as a ranking.

Abmatic calls predictive intent "one of the most powerful and most-misused layers in the modern intent stack". It also discloses its own lean: its architecture is first-party-primary, with predictive intent as a layer rather than the foundation. That is worth remembering whenever a vendor writes about its own category. Every source here sells something.

The inputs also differ by vendor. LeadGenius announced LG Intent on 8 August 2026. It combines research signals, hiring, funding and news, then applies AI-powered suppression to filter out illogical or low-confidence activity. Two feeds that both say "predictive" can be built from quite different ingredients.

When does this not matter? If all you need is to know which companies already reached your site, you don't need a model at all. Predictive intent earns a place only when you want accounts beyond the ones that found you.

Your next step is small. Before any demo, ask the vendor to list its inputs in writing, and mark which ones are observed behaviour and which are inferred from company traits.

Demandbase's Jay Tuel, writing on 2 March 2026, opens with a scene most revenue teams know. One account downloads an eBook. Another spikes in website visits. A third opens every nurture email. All three look like engagement.

Then the cost lands. Marketing pushes these "engaged" leads into the pipeline, and sales complains they're unqualified. Campaigns bloat with the wrong accounts. Sales cycles stall out. Demandbase's own question is the useful one: which of these actions indicate real purchase intent, and which are just noise?

Predictive models are one attempt to answer it. Demandbase says intent data promises to reveal what your buyers are actually researching and when they're in market. That's a promise, not a result. Nothing in the four sources reviewed here shows that any single signal settles the question, and none publishes an accuracy figure you could hold a vendor to.

The sales-side pressure is real, though. A MarketsandMarkets guide published 15 January 2026 says 67% of the buyer's journey happens digitally before a prospect engages with sales. It also says the average B2B salesperson spends only 28% of the week actually selling. Treat both as that guide's claims, since it is a vendor-adjacent source with no method given. They do explain why a ranking is tempting: a rep with 28% of the week for selling wants someone else to sort the list.

Here's the trade-off. A ranking that cuts rep time spent on dead accounts is worth having, even if it's only right some of the time. A ranking that fills the pipeline with accounts sales rejects costs you more than the list you had before.

This doesn't apply if your sales cycle is short and your market is small. When a team can name every likely buyer by hand, a model adds little.

Pick one decision this week. Count how many accounts sales rejected last month, and write down which signal had sent each one over.

The ranking guides all describe what a model can infer. They skip what you already hold: your own site's record of which companies came back, which pages they read and in what order. That is first-party intent, in Abmatic's terms, and you can read it before paying anyone.

Here's a method to run this week.

1. Fix the rules first. Write down what counts as a signal for your team before you open the data. For example: a company returning on a separate visit, or reading pricing after a comparison page. If you write the rules after, the data will happily fit whatever you already believed.
2. Read at company level only. Count which companies returned across separate visits. Never attach the read to a named individual.
3. List the pages each company read, and the sequence. Mark pricing, comparison and docs pages.
4. Put the list beside the accounts a predictive feed ranks highest.

Order is the part the top guides leave out. Demandbase's scenario treats a spike in website visits as one flat signal. But a company that reads a pricing page after a comparison page is in a different place from one that opened a blog post twice. Same visit count, different meaning.

Take a worked case from the Demandbase scenario. One account spikes in website visits. Read at the level of pages, that spike could be three views of a blog post, or it could be a comparison page followed by pricing. Only the second is a decision-stage path. A count can't tell them apart. The page sequence can.

Now the comparison. Where your returning companies and the feed's top accounts match, the feed is adding confirmation. Where they don't, one of the two is wrong. Sales' rejection notes can say which: if reps reject the feed's picks and accept the companies your site shows returning, you have your answer.

This method fails with little traffic. If few companies return, there is no sequence to read, and noise looks like order. In that case a predictive feed, or a direct conversation with a target account, may be the only evidence you can get. Say so plainly rather than reading three visits as a pattern.

One more limit. A site record only covers companies that reached you. It tells you nothing about the accounts that never came. That is the one gap a predictive feed claims to fill, and the one place you should ask it to prove itself.

The honest counter-case is this: a model is only as good as what feeds it, and the buyer can rarely see the inputs. Abmatic names historical patterns, observed signals and firmographic features. LeadGenius names research signals, hiring, funding and news. Those lists are not the same, so the outputs can't be expected to agree.

There is a quieter problem. Firmographic features describe what a company is, not what it's doing. A model that leans on them may rank a company high because it looks like past buyers. That's useful for finding lookalikes. It's weak evidence that the company is buying now.

LeadGenius gives a hint of how it handles this. It says LG Intent applies "logical suppression" to filter out illogical or low-confidence activity. That's a fair admission that raw signals contain activity that makes no sense for a buyer. It also means someone has decided, in code, what counts as illogical. You'd want to know the rule.

The MarketsandMarkets guide says intent data lets organizations identify in-market buyers "with surgical accuracy". No figure comes with that phrase in the excerpt reviewed. Compare it with Abmatic, which calls the same layer one of the most misused. When one source offers precision with no number and another warns about misuse, believe the one that tells you where it breaks.

To be fair to predictive models, they do something your site record cannot. They can flag a company before it has ever visited you. If your market is large and your traffic is thin, that may be the only early view you get. The cost is that you're trusting an estimate you can't inspect.

When does the counter-case not apply? When a vendor will show you which inputs fed each account's rank and let you check a sample against your own records. Then the model isn't a black box, and you can judge it like any other source.

Your decision here: ask for the inputs behind ten ranked accounts. If the vendor won't give them, treat the ranking as unverified.

Put the pieces together. A predictive feed is worth buying if it ranks accounts in a way your own results can confirm. That is a testable claim, so test it.

Take the feed's top accounts and the companies your site shows returning. Hand both lists to sales without saying which is which. Then record what reps accept and what they reject. This is the same loop Demandbase describes in reverse: instead of marketing pushing "engaged" leads that sales calls unqualified, sales judges the lists blind and marketing learns from the rejections.

Set the pass mark before you start. A mark written after the results tends to match the results. If you'd run the test again next month, keep the rules unchanged so the two rounds are comparable.

Market matters, and the sources are thin here. None of the four says how coverage differs between US and Indian companies, or how either market's privacy rules change what a feed may collect. So don't assume. Ask each vendor how many of the accounts in your target list sit in its data for the US and for India, and get the price in USD with the rupee figure beside it. At ₹96.88 to the dollar on 9 October 2026, a USD price shifts in rupees whenever the rate moves, so ask which currency the contract is fixed in. A US team and an Indian team may also differ in how many people can read a feed, so size the process to the team you have.

For privacy, consent requirements vary by jurisdiction and your consent setup. Use your own legal review before you connect any data source, and keep your reads at company level.

When doesn't this apply? If you have no sales team to judge the lists, the blind test can't run. Then a founder talking to a handful of target companies will teach you more than a feed will.

Your next step is concrete: send both lists to sales, blind, and write the pass mark down before they reply.

PageFox does not bring new visitors to your site. Its Visitor Agent works on the traffic you already have, and the Research Agent, which finds the right people when traffic is low, is coming soon. If you need more visitors first, fix that before you add any tool.

Decide now that a predictive feed is a hypothesis, and that your own returning companies are the first thing it has to match. This week, write down what counts as a signal, then list which companies came back to your site and which pages they read, in order.