Skip to content
Antiqore

Why most intent data is noise, and what to keep

Topic surge scores rest on two assumptions that stopped holding: that an IP identifies an employer, and that reading about a subject predicts buying it. Here is what survives.

Antiqore5 min read

Third-party intent is the most confidently presented number in the go-to-market stack. A company appears with a topic, a surge score and a week, and the implication is that somebody there is shopping. It is worth being precise about what has actually been observed, because the gap between that and what the dashboard implies is where the money goes.

What is actually measured

A publisher network records that a browser at an IP address read pages tagged with a topic. A resolution service maps that IP to a company. An aggregator compares this week’s volume against a baseline and calls the difference a surge.

Two inferences are stacked in that sentence, and both are weaker than they look.

Inference one: this IP is that company

This was a reasonable assumption when most knowledge work happened on a corporate network. It is much less reasonable now. Home connections resolve to consumer ISPs. Mobile resolves to a carrier. VPN traffic resolves to wherever the exit node is. Shared and serviced offices resolve one range to a dozen unrelated tenants, so one person’s reading is attributed to all of them.

The resolution rate that gets quoted is a coverage number — the share of traffic that could be mapped to something. It is not an accuracy number, and the two are routinely conflated in sales conversations.

A coverage rate tells you how often the system produced an answer. It tells you nothing about how often the answer was right.

Inference two: reading predicts buying

Even where the mapping is correct, the reader is unidentified. A surge in “data warehouse” at a 900-person company is consistent with a buying committee doing research. It is equally consistent with one engineer having a curious week, a student on the guest network, a competitor doing exactly the research you are doing, or an analyst writing a report.

These are not edge cases. In a large organisation they are the base rate, and nothing in the score distinguishes them.

Where it still earns its place

This is not an argument that the category is worthless. It is an argument about where a weak signal is worth having.

SituationVerdictReasoning
10,000+ addressable accounts, small teamDefensibleYou cannot read them all. A noisy ordering beats alphabetical.
200–500 named accountsHard to justifyThe same budget spent reading those accounts properly returns dated, sourced reasons to call.
Deciding who to call this weekWeakAttribution error means a meaningful share of your week goes to the wrong company.
Deciding what to put in the messageUselessA topic is not a reason. It gives the rep nothing specific and true to open with.
Territory or campaign planningReasonableAggregate error averages out across hundreds of accounts in a way it never does for one.
The pattern: intent survives at aggregate scale and falls apart at the level of a single named account.

What to use instead at the account level

The substitute is not a better score. It is a different class of input: events that happened, with a date and a page you can open.

  • Roles posted that imply the problem you solve — with the posting linked and dated.
  • Leadership changes in the function that owns your budget line.
  • Public product or integration announcements that change what they need.
  • Funding, filings and reported expansion — sourced to the announcement, not to a database row.
  • Changes to their own site: new pricing, a new market page, a removed product.

There will be fewer of these, and that is the trade. Each one carries a date, an openable source and a specific thing to say on the call, which is the part a topic score can never supply. The tests a claim has to pass are set out in what actually counts as evidence for a buying signal.

First-party is a different argument

Behaviour on your own site is not subject to the first inference at all. The session is real, it happened on your property, and you can hold the consent record for it. It is still not identity — a session is a session until someone tells you who they are — but the chain is one link shorter and you own every link in it.

It needs the same discipline about age. Somebody who read your pricing page in March is not in market in July, and treating a four-month-old session as live is the first-party version of the same mistake. That argument is worked through in why a signal from March is not a signal.

Our own answer is to keep the two separate and to make the weaker one say so: Analytics records what happens on your site first-party, Leads researches companies from sources you can open, and neither is allowed to launder an inference into a fact on the way to the list.

Frequently asked questions

What is third-party intent data?
Aggregated browsing behaviour across a network of publisher sites, resolved to a company by IP address and sold as a topic score. You are buying the claim that somebody at a company read something about a subject, not the identity of who read it or why.
Why is IP-to-company resolution unreliable?
Remote work broke the assumption it rests on. A large share of traffic now comes from residential connections, VPNs and mobile networks that resolve to a carrier rather than an employer, and shared office ranges attribute one person's reading to every tenant in the building.
Is intent data ever worth buying?
It is most defensible in large, slow markets where you cannot cover the whole addressable list by hand and a weak prioritisation signal beats none. It is least defensible in a market of a few hundred accounts, where the same money spent on reading those accounts properly returns far more.
What is a better substitute for topic surges?
Observable events with a date and a source — a role posted, a product shipped, a filing made, a leadership change, a public integration announced. They are less numerous and far harder to argue with, and they tell you what to say rather than only who to call.
How should first-party intent be treated differently?
First-party behaviour on your own site is a real observation of a real session, so the identity problem is smaller and the consent position is cleaner. It still needs a decay rule: someone reading your pricing page in March is not in market in July.