Intent Data Providers: Questions to Ask Before a Pilot

Before you pilot any of the intent data providers on your shortlist, get answers to twelve questions grouped under method, identity, freshness and activation, and design a blind holdout so the pilot measures the data rather than your team's enthusiasm for it. An easy evaluation mistake is asking how high the scores go before asking what behavior the scores summarize. A composite number from 0 to 100 tells you nothing you can audit. An observed signal definition, resolved to an account you recognize, is something you can check against reality.

That distinction matters because intent products describe research activity at the account level, and that activity is evidence to interpret, not proof that any individual is making a purchase decision (Bombora). Behavior that may indicate interest still has to be evaluated separately from fit, source methodology and timing (Demandbase). Your pilot should test each of those separately too.

Signal definition versus vendor score

Every intent product ships two layers. The bottom layer is observed behavior: pages consumed, topics researched, searches run, content downloaded, resolved to a company. The top layer is a model that compresses that behavior into a score or a "surge" flag.

Vendors naturally lead with the top layer because it is simple to demo. But the score is only as good as the observed layer and the resolution logic underneath it. Two providers can show the same account at 85 for entirely different reasons: one because three people read comparison content this week, another because a single bidstream event matched a broad topic. If you cannot see which behaviors triggered a score, you cannot debug a bad quarter, and you cannot tell whether a score change reflects your market or a model update on the vendor's side. Insist on access to the underlying intent data fields, not just the composite.

The twelve questions

Print this as a worksheet, one column per vendor, and require written answers. Verbal answers in a demo have a way of softening by the time the contract arrives.

A 50-account blind holdout, worked through

Here is an illustrative pilot design. All counts are hypothetical and exist only to show the arithmetic.

Suppose your ICP universe contains 300 accounts. The vendor scores all 300 and flags 100 as showing elevated intent. Split them into three cells before anyone acts:

1. Flagged and activated: 50 accounts. These get your standard signal-triggered play.
2. Flagged and held out: 50 accounts. Randomly selected from the flagged 100. Keep the signal flag hidden from the sales operators working these accounts; an experiment owner retains the assignment list. Continue the same baseline activity used for the unflagged group, with no added signal-triggered play. This is the blinded holdout.
3. Unflagged baseline: 50 accounts. Randomly sampled from the 200 the vendor did not flag. No treatment changes here either.

Then measure two things separately over the pilot window:

• Exposure: did the account actually exhibit behavior you can independently corroborate, such as inbound visits, replies to existing sequences, or opportunity activity that started organically?
• Response: did the account progress on your chosen outcome metric, defined before the pilot, such as a qualified meeting or a new opportunity?

The comparison that tests the data is cell 2 versus cell 3: flagged-but-untouched accounts against unflagged-but-untouched accounts. If flagged holdout accounts progress at a materially higher rate with zero treatment difference, that is observational evidence the signal carries information — descriptive counts to weigh against base rates and small-sample uncertainty, not proof of predictive power, because the vendor's flagging is not random assignment. The comparison that tests your play is cell 1 versus cell 2.

Keep the arithmetic honest. If 7 of 50 holdout accounts progress versus 4 of 50 baseline accounts, that is 14% versus 8%, a 6-point gap on small samples. Whether that gap justifies the contract is a decision you should make in advance: write down the minimum gap you would need to see, and the sample size caveat, before the pilot begins. Do not let a vendor compute lift for you after the fact, and do not report a lift number internally that the sample cannot support.

Tradeoffs and failure handling

Identity resolution failures look like signal failures. If the vendor maps behavior to the wrong domain in a corporate hierarchy, a real signal lands on the wrong account and your holdout comparison degrades. Test question 6 against your five messiest multi-brand accounts before the pilot, and document data provenance for each delivered field so you can trace a score back to a source.

Broad topics inflate flags. A taxonomy topic like "cloud software" can flag a large share of your universe. Narrower topics reduce volume but risk missing early research. Pick topics at the level your buyers actually research, and accept that this is a tuning decision you will revisit, not a one-time setting.

Mid-pilot model changes contaminate results. If the vendor updates its scoring model during your window, your before-and-after comparisons break. Question 9 exists for this reason; get change notification in writing and log the date of any change against your metrics.

Activation dead-ends waste good data. A signal that arrives as a weekly CSV nobody opens is functionally no signal. Decide before the pilot where scores land, who acts on them, and within what window. Feeding intent fields into your own lead scoring model, alongside fit and engagement data, gives you more auditable control than acting on the vendor score alone. For a broader view of how intent fits among other business signals, see our guide to buying intent and business signals.

Where AstroFabric fits

AstroFabric treats intent as one signal type inside a larger data layer. You describe an objective, and autonomous agents discover accounts, verify and enrich records, monitor buying-intent and marketplace signals, score relevance against your criteria, and stream structured intelligence into your CRM, warehouse or Slack. That means an intent flag from a provider can be corroborated against hiring, funding, technographic and news signals rather than acted on in isolation, and the resulting scored targets land in the systems your team already uses. The pilot discipline above still belongs to you: define the signal, hold out a cell, and measure before you scale.

FAQs

What is the difference between an intent signal and an intent score? A signal is an observed behavior resolved to an account, such as topic-level content consumption. A score is the vendor's model output summarizing those behaviors. You can audit a signal definition; a score you can only test. Always ask what raw behavior sits underneath the number.

How many accounts do I need for a meaningful intent pilot? Enough to fill three cells: flagged and activated, flagged but held out, and an unflagged baseline. Fifty per cell is a suggested practical floor for spotting large differences. Decide your minimum decision gap before the pilot starts instead of rationalizing small differences afterward.

Should intent data replace fit-based targeting? No. Intent describes behavior that may indicate interest; fit describes whether the account belongs in your ICP at all. A surging account outside your ICP is usually noise. Evaluate fit, methodology and timing separately and combine them in your own scoring.

Ready to put intent signals next to fit, hiring and funding data in one scored dataset? Start with AstroFabric and run your next pilot with a holdout built in.

Sources

• Bombora: intent overview
• Demandbase: what is intent data