Growth · 10 min read · 2026-09-14
Account Intent Scores: Weighting Signals Without Lying
Account intent scoring promises something every B2B SaaS growth team wants: a single number that tells sales which accounts are in-market right now, so limited rep attention lands on the accounts most likely to buy. The mechanics are seductive — pull intent signals from third-party networks, your own website and product, review sites, and ad engagement; resolve them to the account; weight them; output a 0–100 score with a "sales-ready" threshold. The problem is that a score assembled carelessly measures your data coverage, not buying intent, and hands sales a worklist ranked by which accounts your vendors happen to see rather than which accounts are actually researching.
This is an analytics-validity problem wearing a sales-enablement costume. The score is a predictive model, and like any model it can be well-calibrated or garbage, honest about its weights or a black box. This post takes the vendor-neutral view: how to weight first- versus third-party signals defensibly, why deduplication is the step everyone skips, and — most importantly — how to validate that the score predicts buying instead of assuming it does. It sits alongside the broader question of account-level versus user-level analytics, because an intent score is the sharpest example of rolling messy user-level signals up to an account-level number.
What the score actually aggregates
An account intent score is a composite of signals from three broad sources, and understanding what each contributes is the prerequisite to weighting them. Third-party intent networks observe research activity across the web — an account's employees reading buyer guides, comparing vendors, visiting review sites — and report it at the account level. First-party behavior is what you observe directly: website visits, content downloads, pricing-page views, and product usage. Engagement signals sit between them: ad clicks, email opens, and review-site visits you can attribute. As one aggregation framework describes it, the score "synthesizes intent signals from multiple third-party data sources, first-party behavioral systems, and engagement channels into a unified account-level intent score" (Saber).
The step that makes aggregation possible — and the one most likely to go wrong — is entity resolution: mapping individual research activities to the parent account. An analyst at a 5,000-person company reads three guides from three different offices on three devices; unless all of that resolves to one account, you either fragment the signal across phantom accounts or misattribute it. This is the same identity-resolution discipline that underpins product analytics, applied to the buying committee. Without it, the score's inputs are wrong before any weighting begins, and no clever weighting recovers a broken rollup.
Weighting: predictive value, not data volume
The weighting question — how much should each signal contribute — should be answered by predictive value and confidence, not by which data you happen to have the most of. The two properties trade off. Third-party intent tends to be more predictive of near-term buying because it captures research happening off your properties, where most of the buying journey occurs; first-party behavior carries higher confidence because you observed it directly rather than inferring it from a network. A signal can be highly predictive but low-confidence, or the reverse, and a good weighting accounts for both rather than treating all inputs as interchangeable points.
A defensible starting point exists in the literature: weight research activity highest, engagement lower, and add a bonus when both are present — one published model uses roughly 40% research, 30% engagement, and 30% for the combination, on the logic that "research activity is more predictive of near-term buying than engagement alone, but the combination is more powerful than either data source independently" (DemandScience). Treat those numbers as a hypothesis, not a truth. The specific weights matter far less than whether you have validated them against outcomes, which is where most scoring programs quietly fail.
Deduplication: the step everyone skips
Before you sum weighted signals, you have to deduplicate the ones that measure the same underlying behavior, or the score inflates without adding information. The canonical failure: an account visits a G2 review page, your first-party pixel records it, and a third-party intent network that also observes G2 reports the same visit. Sum them and one act of research contributes twice, so the account looks hotter than it is — purely because that behavior was visible to two of your data sources. Accounts visible to more sources score systematically higher as an artifact of coverage, which is exactly the bias an intent score is supposed to eliminate.
The fix is the same discipline as event deduplication: resolve each signal to a canonical event with an identity and a timestamp, collapse near-duplicate observations of the same behavior across sources, and only then apply weights and sum. This is unglamorous plumbing and it is the difference between a score that ranks accounts by intent and one that ranks them by how many of your vendors can see them. If you only fix one thing in an existing scoring program, fix this — it is the most common reason a score correlates with data coverage instead of buying, and it is invisible until you go looking for it.
Validate against outcomes or it is a vanity number
An intent score is a predictive model, so validate it like one: check whether accounts that scored high actually converted, measured against a baseline. Look back at accounts that crossed your sales-ready threshold and compare their opportunity-creation and win rates to accounts that did not. If high-intent accounts do not convert materially better than the rest, the score is not measuring intent — it is measuring coverage or noise, and every rep hour spent on its rankings is misallocated. This backward-looking check is cheap and most teams never run it, which is why so many intent scores persist unchallenged despite predicting nothing.
Track two properties over time. Precision: of the accounts the score flagged hot, how many actually bought — this tells you whether the threshold is meaningful. Calibration: do accounts scoring 70 convert roughly as often as a 70 should imply, or is the number decorative? A well-calibrated score lets sales trust the gradient, not just the threshold. This is the same demand for evidence behind disclosing the logic behind PQL and churn scores: a score that drives decisions should be explainable and validated, not an opaque number vendors hand you with a confident color scale.
How to actually use the score
Treat the validated score as a ranked worklist, not a verdict. Its legitimate job is to help sales spend finite attention on the accounts most likely to be researching a purchase — a prioritization aid, not a promise that a 72 will close. Used this way, even an imperfect score adds value, because ranking accounts roughly by intent beats working them alphabetically or by list order. Overreach starts when teams treat the number as fact, auto-routing accounts and forecasting pipeline off a score that has never been validated against a single closed deal. The score earns trust by being directionally right and honest about its uncertainty, not by pretending to a precision it does not have.
Pair the score with the transparency to explain any given account's ranking, because sales will — and should — ask why an account scored where it did. Being able to answer "this account scored 74 on heavy third-party research plus two pricing-page visits, deduplicated against the network's G2 signal" is what makes reps act on the number instead of ignoring it. That explainability also closes the loop with the rest of your acquisition analytics: intent scoring is one more attempt to illuminate the dark funnel of research that happens before an account ever identifies itself, and it feeds the same product-led sales instrumentation that decides which accounts get worked. Build it as a validated, deduplicated, explainable model and it prioritizes real attention; build it by naively summing whatever signals you can buy and it just tells you which accounts your vendors can see.
FAQ
What is an account intent score?
An account intent score is a single number, usually 0–100, that summarizes how much buying research an account is doing by aggregating signals from multiple sources into one account-level measure. It combines third-party intent (from networks like Bombora or TechTarget that observe research activity across the web), first-party behavior (your website visits, content downloads, product usage), and engagement signals (ad clicks, email, review-site visits), then applies entity resolution to roll individual activities up to the parent account and weighting to produce the composite. The score exists to help sales prioritize which accounts to work; most B2B SaaS teams set a 'sales-ready' threshold somewhere between 60 and 75 on a 100-point scale.
How should you weight first-party vs third-party intent signals?
Weight by predictive value and confidence, not by which data you have more of. Third-party intent tends to be more predictive of near-term buying because it captures research happening off your properties, while first-party behavior carries higher confidence because you observed it directly. A defensible starting split is to weight research activity highest, engagement lower, and add a bonus when both are present — one published model uses roughly 40% research, 30% engagement, and 30% for the combination. But the specific weights matter less than validating them: any weighting is a hypothesis about what predicts buying, and it should be checked against whether high-scoring accounts actually convert, then adjusted.
How do you avoid double-counting overlapping intent signals?
Deduplicate signals that measure the same underlying behavior before you sum them, because overlapping sources inflate the score without adding information. The common failure is counting a G2 review-site visit once from your first-party pixel and again from a third-party intent network that also observed it, so one act of research contributes twice and the account looks hotter than it is. The fix is the same discipline as event deduplication: resolve signals to a canonical event with an identity and timestamp, collapse near-duplicate observations of the same behavior, and only then apply weights. Without this pass, accounts that happen to be visible to more of your data sources score higher purely as an artifact of coverage.
How do you validate that an intent score predicts buying?
Validate it the way you validate any predictive model: check whether accounts that scored high actually converted, and compare that to a baseline. Concretely, look back at accounts that crossed your sales-ready threshold and measure their conversion-to-opportunity and win rates against accounts that did not — if high-intent accounts do not convert materially better, the score is not measuring intent, it is measuring data coverage or noise. Track the score's precision (of accounts flagged hot, how many bought) and calibration (do 70-scoring accounts convert roughly as often as the model implies) over time. A score nobody has validated against outcomes is a vanity number that misallocates sales attention.
Are account intent scores accurate enough to act on?
They are useful for prioritization but not precise enough to treat as fact, and the accuracy depends heavily on your weighting and deduplication discipline. Combining multiple signal sources genuinely produces better prediction than any single source, but only if overlapping signals are deduplicated and the weights have been validated against real conversion outcomes; a naively summed score mostly measures which accounts your data vendors can see. Treat the score as a ranked worklist for sales — a way to spend limited attention on the accounts most likely researching — rather than a guarantee, and pair it with the transparency to explain why a given account scored where it did.