Third-Party Intent Data: What Is Worth Buying and What Is Not

A provider demo shows a list of companies "in-market for your category this week", and it is impressive: logos you recognise, a surge score, a topic cluster. The question nobody asks in that meeting is how the provider knows, and the answer changes what the list is worth by an order of magnitude.

Five quite different products are sold under the label intent data. They are collected differently and fail differently, and lumping them together is why so many teams write off the category after one disappointing year.

Five things are sold as intent, and they are not the same product

Research topic surges come from publisher co-operatives. A network of B2B publishers carries the vendor's tag, consumption is resolved to a company, and a surge is flagged when that company's reading of a topic deviates from its baseline. Note what this is: a model output, not an observation.

Review-site activity comes from G2, Capterra, TrustRadius and similar, the narrowest and strongest of the five. A company looked at your category grid, or a comparison page naming you and a competitor: unambiguous, close to a purchase decision, and correspondingly low in volume.

Technographic change is detected from public evidence: script tags, DNS and MX records, certificates, and what job adverts mention. Solid for anything customer-facing, unreliable for anything internal, and lagging, because a tag removed today persists in a vendor's index for weeks.

Hiring signals are scraped from applicant tracking pages and job boards. Reliable as facts, ambiguous as intent. A posting for a marketing operations manager proves a gap exists. It says nothing about budget, timing, or whether the gap gets filled by a person or a tool.

Funding and news events come from press releases and registry filings. Reliable, and public, so the informational advantage is zero: every competitor gets the same alert the same morning.

How it is collected decides how far you can trust it

The weak joint in the first two categories is IP-to-company resolution. It works reasonably for large organisations with owned address ranges and poorly for remote-first teams on residential connections, anyone behind a corporate VPN terminating in another country, and small companies sharing an ISP block. Error rates here are not marginal.

So ask two questions and write the answers down. What share of the signal comes from logged-in identity rather than IP inference? And what is resolution accuracy under two hundred employees, where most mid-market ICPs live? A provider who cannot answer has given you an answer.

Then understand the baseline. Surge measures deviation, and deviation is noisy at low volume: three people reading two articles can cross a threshold a company of nine thousand never will. Hence accounts appearing and disappearing month to month.

By the time it reaches you it is weeks old, and your competitors have it too

Two structural limits, neither of which any provider can fix. Latency compounds: aggregation windows of seven to fourteen days, then a weekly delivery, then the lag before anyone works the file. Three weeks between the behaviour and your first touch is ordinary, not bad. In a category where buyers shortlist inside a fortnight, you arrive after the shortlist was drawn.

Non-exclusivity is the other half. Co-operative surge data and review-site signals are sold to every vendor in the category, including the two you lose to most often. If a surge is real, it is on three desks. Your advantage cannot be the data, only what you do in the hour after receiving it, which is a statement about your routing.

Prioritisation input, never a trigger

The operating rule is one sentence: no bought signal, on its own, puts an account into a rep's queue. What bought intent may legitimately do is raise an account's position in a list you were working anyway, add it to an awareness audience, justify twenty minutes of research, or lower the threshold at which a first-party signal fires a play. All prioritisation. None of it action.

What it must never do is become the content of the outreach. The message beginning "I saw your team is researching customer data platforms" fails twice: the recipient has no idea what you mean, because they did not personally do it, and they now know you buy surveillance-shaped data about them. Keep bought signals inside your systems, out of your sentences.

Write that into the routing rules. The bought signal becomes an attribute on the account record feeding the priority model, while the trigger side reads only events you observed. The first-party intent layer has to exist first: buying signals before you can observe your own is paying for a rumour about a house you have never entered.

Test the provider against your own closed-won before you sign

This takes a week and almost nobody does it. Ask for retrospective data on a named account list, or run a trial covering one delivery cycle.

Take last quarter's closed-won accounts, at least thirty for the exercise to mean anything, and ask one question of each: did the provider flag this account before the first meeting was booked? That share is your recall. If fewer than half the deals you won were visible beforehand, the provider is not seeing your buyers, whatever its coverage claims say.

Then check the other direction, because recall alone can be bought with volume. Count how many accounts in your addressable market it flagged that quarter. If it surfaces thirty percent of them, it is not prioritising, it is describing your market back to you.

Two further checks. Insist on flag dates, because a signal arriving after the opportunity was created has no value regardless of accuracy, and retrospective datasets hide this beautifully. And measure overlap: if most flagged accounts sit in your signal ledger already, with a stronger first-party event attached, you are buying a duplicate of something better.

When the money is better spent elsewhere

Several cases, covering more teams than vendors would like.

If your ICP is under roughly five hundred accounts, buy nothing. A person can read five hundred companies' job boards, funding announcements and tech stacks over a few weeks, and will understand them better than any score. Spend the subscription on that time.

If your CRM is messy enough that account matching fails, the feed lands on duplicates and orphans and produces confusion rather than priority. Fix the record first. Equally, if nobody works the queue you have, more accounts only lengthen the backlog.

And if your category is new or narrow, topic taxonomies will not contain it. Surges get measured against adjacent topics half your market reads for unrelated reasons, and the output is noise with a confidence score attached. That is the honest limit of the category: bought intent describes plausible behaviour by somebody at a company, some weeks ago, and it was sold to your competitors too. Priced as a prioritisation hint it earns its keep. Priced as a pipeline source it disappoints.

Frequently asked questions

What types of third-party intent data are there?

Five distinct products: research topic surges from publisher co-operatives, model outputs measuring deviation from an account's baseline; review-site activity from G2 and similar, narrow but close to a purchase decision; technographic change detected from public web evidence; hiring signals scraped from applicant tracking systems; and funding or news events from public filings. They differ in collection method, reliability and latency, and treating them as one category is why teams write off the market.

How accurate is third-party intent data?

It depends how the signal was collected. Review-site activity and hiring signals are close to observations and hold up as facts, though hiring says nothing about budget or timing. Topic surge data depends on resolving anonymous traffic to a company by IP address, which works reasonably for large organisations with owned ranges and poorly for remote-first teams, VPN users and small companies on shared ISP blocks. Ask what share of a provider's signal comes from logged-in identity, and its accuracy under two hundred employees.

Why should bought intent data never trigger outbound on its own?

Because it is weeks old on arrival, it was sold to your competitors at the same time, and it describes a company rather than a person. Referencing it fails twice: the recipient did not personally do the thing you describe, and they learn that you buy behavioural data about them. Bought signals belong in the prioritisation model as an account attribute, while sequence triggers come from behaviour you observed.

How do you evaluate an intent data provider before buying?

Run it against your own history. Take at least thirty closed-won accounts from last quarter and check what share the provider flagged before the first meeting was booked, which gives you recall. Then count how much of your addressable market it flagged in the same period, because a feed surfacing thirty percent of your market describes it rather than prioritising it. Insist on flag dates.