Intent Without the Black Box: We Backtested a Signal-Native Buying Predictor

Every intent-data vendor promises the same thing: tell me who is about to buy. In practice, most of them tell you a topic spiked - some anonymous account read a few articles about your category this week - and leave you to guess what it means. Your reps have learned to ignore it, and they are right to.

So we tried something different. We took the signal data we already produce - more than a billion buying signals across 700+ types - and asked a sharper question: can these signals actually predict who buys, before they buy? Then we backtested it. Here is what we found.

Intent data has a black-box problem

Today’s intent products share three flaws:

• They are a black box. A score goes up. You never learn why, or whether it means anything.
• They lean on one signal. Usually web visits or bid-stream - a single, eroding proxy stripped of everything else a buyer is doing.
• They never learned from outcomes. They were never trained on which patterns actually ended in a purchase.

That last one is the real gap. If you want to predict buying, you need to know what buying looks like in the data - and that means learning from deals that actually closed.

The missing ingredient: the answer key

Here is the thing almost no one has. Everyone in this market sits on signals. Almost no one has labels - a clean record of which companies actually bought, and when. Signals are the exam. Outcomes are the answer key. Without the answer key, you are guessing which patterns matter. With it, you can measure exactly which signals precede a purchase and weight them accordingly.

Everyone has signals. Almost no one has the answer key.

So we tested it

We trained gradient-boosted models on 14 months of our signal history - 82.8 million signals across 31 types - and pointed them at real, observed technology-adoption outcomes. The question: given only the signals a company was exhibiting, could the model predict which companies would adopt a given category of software within the next 90 days?

We built five category-specific models. Every one of them worked - and worked well.

The results

Measured by top-decile lift - how many more real buyers show up in the model’s top 10% of predictions versus random targeting:

Generic baseline, with no category model: 78.8% AUC and 2.28× lift. The category-specific models beat it by two to four times over.

In plain English: if your reps only worked the accounts these models flagged, they would hit real buyers four to seven times more often than working a random list. That is the difference between a full pipeline and a wasted quarter.

Why it works: signals no one else fuses

The model did not lean on any single magic signal. It keyed on a combination:

• Tech-stack freshness - how recently a company last adopted anything new.
• Adoption velocity - how many new tools they added in the last 90 to 180 days.
• Total tech footprint - how much they already run.
• Hiring signals - roles opening in the relevant function.
• Signal recency - how fresh the underlying evidence is.

Any one of these is noise. The combination is a buying signature. And because it runs on gradient-boosted trees rather than an opaque neural net, every prediction is auditable - you can trace a score straight back to the signals behind it. No black box.

From tech adoption to what actually closed

Technology adoption was just one kind of outcome we could observe at scale. It proved the method: signals plus real outcomes equals prediction. But the most valuable label of all is not tech adoption - it is what actually closed-won, across every category, for every kind of product.

No single company has enough of that data alone. That is why we are building a co-op: a privacy-safe way for teams to contribute anonymized closed-won and closed-lost outcomes, so the model learns the buying signature for their market - and hands back a list of the companies exhibiting it right now.

The takeaway

Intent does not have to be a black box. When you layer real outcomes on top of a billion signals, you can predict who is in market, at what stage, and show the exact signals that prove it. We proved the thesis. Now we are opening it up.

If you have closed-won data and want to help build the next intent engine - and get a map of who is about to buy in your market - join the Intent Initiative.