Why intent data fails to improve pipeline: four common reasons

Most of the common reasons why intent data fails to improve a pipeline lie in how the feed is used, and each can be tested on your own CRM. This post takes four in turn, each with the test that shows whether it applies and the fix: no rule for what a rep does differently, a score with no evidence to act on, a window that does not match your sales cycle, and no holdout to measure against. A worked holdout then separates what a feed predicts from what acting on it adds.

Four common reasons why intent data fails to improve your pipeline

Intent data fails to improve pipeline for four common reasons. No rule says what a rep does differently when an account is flagged. The score arrives with no evidence a rep can act on. The window the score reads does not match the length of the sales cycle. And no holdout group exists to measure the effect against.

Run the tests in this order: a later one means little while an earlier one fails.

Why intent data doesn't work for a B2B sales pipeline on its own

Intent data doesn't work for a B2B sales pipeline on its own because a score is information, and pipeline comes from action. A flag adds opportunities only when it changes which account a rep works, when, or with what opening. Where the queue, the timing and the message stay as they were, the pipeline stays as it was.

No rule for what a rep does differently

The usual failure is a score that lands in a CRM field and stops there, while reps keep their own order of work.

Test it on logged activity for flagged and unflagged accounts in the same fit tier of your account scoring. If first touch comes no sooner and touches are no more frequent, the data was never used, and the quarter says nothing about whether it works.

The fix is a play, a rule written as a sentence with an owner and a deadline: when a strong-fit account enters the top band, its owner makes first contact within five working days and opens with the newest signal. Say what the rep stops doing to make room. Bands deserve different rules: 6sense's documentation of its predictive buying stages names five, from Target to Purchase, and describes a treatment for each, such as broad-based engagement for the Awareness stage and outbound prospecting by sales at the Decision stage.

A score with no evidence to act on

A number tells a rep that something happened and leaves out what. Without that detail the rep has no opening line and no way to judge the flag, and after a few empty calls stops trusting the list.

Feeds differ in what travels with a score, so ask for the evidence fields before you design the play. G2's Buyer Intent documentation says each signal represents a buyer interaction on G2 and lists the types, among them a view of a product's profile, its pricing page or a comparison page. A Fokals score comes with the dated evidence behind it, so the reason reads as something the company did on a stated day. Whatever the feed, let an account into the top band only when its evidence is in the queue beside the score.

A window that does not match the sales cycle

Three clocks run at once: how far back the score reads, how far ahead it speaks, and how long you wait before judging it. Vendors document the first two.

• Bombora's Company Surge compares the most recent three weeks of activity with a 12-week historical baseline, for every pairing of account and topic.
• 6sense's intent score gives the likelihood that an account opens an opportunity in the next 90 days, and is updated daily.
• A Fokals score is written weekly for each company and topic, once the week closes, and an intent surge flag marks a company that has become active on a topic against its own record.

The third is yours: measure the lag from first flag to opportunity, and read the median and the slowest quarter. Use each score as it stood on its date, which takes point-in-time data: a feed that overwrites its scores answers last spring's question with today's number.

If the median lag is 120 days and you judge the feed on opportunities opened within 30, you will miss at least half of the opportunities that follow. If the lag is 15 days and a flagged account reaches a rep three weeks after the flag, the signal was spent before anyone called. Set the outcome window from the measured lag, and keep the score version and your cut-offs fixed while you measure.

No holdout to measure against

Without a baseline, two wrong conclusions are open. One credits every opportunity at a flagged account to the data. The other sees flat pipeline and blames the data, when the play was never run. Both overlook that a flag selects accounts that are already active, which can open opportunities at an above-average rate with nobody acting on it.

The UK government's Magenta Book on evaluation calls that missing comparison the counterfactual: what would have happened in the absence of the intervention. It describes a randomised controlled trial as random allocation to intervention or control groups, so that measured differences can be put down to the intervention alone. For a revenue team the intervention is the play, and the control group is a random share of flagged accounts whose flag the reps never see.

Why intent signals fail in B2B sales: false positives along the buyer journey

Intent signals fail in B2B sales when they are false positives: an account is flagged and no opportunity follows. Along the buyer journey that happens in four ways. The activity belongs to another entity, the actor has no part in a purchase, the research runs quarters ahead of any budget, or the topic is wider than your product.

Vendors document controls for the last two. 6sense's documentation describes accounts in its Awareness stage as likely to be just starting their buyer's journey, with a lower likelihood of converting to an opportunity in the next 90 days than accounts in later stages. Bombora's guidance on thresholds treats a score of 60 or more on a topic as a spike, and recommends a topic threshold as well: a minimum number of topics an account must spike on, set to at least 25% of the topics in a report or cluster.

Entity and actor are settled by evidence. Resolve each account to a company by domain matching and its corporate hierarchy. Where the signal is a public act of the company, read it for what it records: a posting records an intention to hire, and a detection shows presence on the website. Each signal in Company Signals names the company and carries its date and a line of detail, so a rep can open it and settle both questions.

Sort 50 flagged accounts that opened no opportunity into the four kinds, or into a fifth: a sound signal on a deal that was lost or is still open. The largest pile names the first fix.

Predictive accuracy of account-level B2B intent data: false positives and the research

We found no published standard for the predictive accuracy of account-level B2B intent data in a search of the IAB and IAB Tech Lab sites: no accuracy figure, no test method. Vendors publish descriptions of their own scoring. So measure accuracy on your own accounts: precision, recall and lift, against a holdout, with the false positives counted.

The nearest standard we found there concerns disclosure. The Data Transparency Label that the IAB Tech Lab and three other trade bodies unveiled in beta in 2018 set out what the buyer of an audience segment was to be told: who provided the data, how the segment was constructed, including any modelling applied and the event lookback window, and where its components were sourced. It makes a sound list of questions for an intent feed.

The US Food and Drug Administration's statistical guidance on diagnostic tests has the vocabulary for the measurement: a result is a false positive when the test is positive and the condition is absent, and the predictive value of a positive result is the share of positive tests in which the condition is present. For a revenue team the test is the flag, and the condition is an opportunity opened within the outcome window.

• Precision is the predictive value of a flag: the share of flagged accounts that open an opportunity.
• Recall is the share of all opportunities that came from accounts flagged beforehand.
• Lift is precision divided by the base rate, which is the share of all target accounts that open an opportunity in the same window.

Precision depends on the base rate as much as on the feed. Where 2% of target accounts open an opportunity in a quarter, a flag that triples that rate is still wrong 94 times in 100, and a rep who expected a list of buyers concludes that the data is broken. Lift is the fair measure of the feed, and the cost of the false positives is the measure of the play.

A worked example: Acme Robotics measures a feed against a holdout

The example is illustrative: Acme Robotics is an invented company, and every figure is an assumption.

Assumptions:

• Acme has 4,000 target accounts of strong fit.
• With no intent play, the base rate is 2%: 80 of them open an opportunity in a quarter of 90 days.
• An intent feed flags 600 of the 4,000 at the start of the quarter.
• The 600 are split at random. On 300, reps see the flag and run the play: first contact within five working days, opening with the evidence. The other 300 are the holdout, worked in the usual order with the flag hidden.
• An account under the play takes 1.5 hours of rep time.
• After 90 days, opportunities have opened at 24 of the 300 play accounts, 15 of the 300 holdout accounts and 50 of the 3,400 unflagged accounts.
• Acme decided beforehand to act once the difference between the two groups is at least twice its standard error.

1. The feed predicts, and most flags are still false positives. Untouched, flagged accounts open opportunities at 2.5 times the base rate. Even so, 285 of the 300 holdout accounts open none, and 50 of the 80 baseline opportunities come from accounts the feed never flagged. The flag orders the list, and reps should be told so.
2. The play adds the pipeline. A report with no holdout would credit all 24 opportunities to intent, 2.7 times the 9 that the play added, because 15 would have opened anyway.
3. One quarter is not yet proof. Acme uses the test for equal proportions in the NIST/SEMATECH e-Handbook of Statistical Methods. The pooled proportion is 39 / 600, or 6.5%. The standard error of the difference is the square root of 0.065 × 0.935 × (1/300 + 1/300), or 2.0 points, and z is 3 / 2.0, or 1.5. Under its own rule Acme runs the same split for a second quarter. At the same rates, 48 of 600 against 30 of 600, the standard error is 1.4 points and z is 2.1.

Set the 50 hours beside what an opportunity costs from the rest of your outbound.

Running the four tests on Fokals intent data

Each test asks something of the data, and the Fokals intent dataset is built to supply it.

• Evidence, for the rule and the rep. Intent Scores holds a weekly score from 0 to 100 for each company and topic, and every score comes with the dated evidence behind it. Company Signals delivers dated company signals daily.
• The record as it stood, for the lag. Each weekly row is written once, after the week closes, and is never revised.
• A fixed definition, for the holdout. A score keeps the definition it was written under, and a breaking change ships with at least 90 days' notice, so a test run under one definition measures one thing.

The guide to prioritising accounts with intent scores turns the scores into a weekly queue with a holdout, and what a Fokals intent score gives you shows how to read a row. The methodology describes each dataset.

Frequently asked questions

What this page says about the products it names was checked against their public documentation on 7 October 2026. Product and company names are trademarks of their owners. Fokals is not affiliated with them or endorsed by them.