Your Lead Scoring Model Only Works on People Who Already Found You

The only thing you can observe about a cold lead

Everything scoreable about an outbound lead sits outside your systems: firmographics, org structure, hiring activity, tech stack, funding, leadership changes. None of it is behavior. All of it is situation.

The useful question is not "how engaged is this lead", because the answer is always zero. It is: how confident am I that this account has the problem I fix, right now, and that there is a reachable human who owns it?

That reframe is the whole model. We run it as three tiers, and they do not mean what most scoring decks mean by tiers.

Tier 3: fit only

Firmographic fit and nothing else. For our own list that means B2B tech, UK or US headquarters with the bulk of employees and decision makers in the same market, 20 to 150 employees, a sales team of 1 to 10, and no RevOps function.

Tier 3 is a base, not a priority. It says an account could plausibly buy, and nothing about whether they will take the call. Most models stop here, call it a score out of 100, and wonder why the ranking does not predict anything.

Tier 2: fit plus a business signal

Something observable that raises probability without proving need. A new revenue leader in the last six months. Hiring SDRs or BDRs with no RevOps role posted alongside them. A PLG company bolting on a sales team. A recent CRM migration. A messy multi-tool stack with no integration owner. These are the "about to feel it" accounts: worth working, not worth your best hour.

Tier 1: a standalone intent signal

One signal that on its own makes you confident the problem is live. Three work for us:

1. Job ads. A company posting a RevOps, Sales Ops, GTM systems or CRM owner role is a buyer openly declaring the need. No inference required, and free with the stack we already run.
2. First-party "knows us". Attended an event, subscribed, viewed the profile, downloaded something, surfaced in visitor identification. This is the one place real behavior exists on a cold list, so capture it deliberately instead of letting it die inside a marketing tool.
3. Structural condition. Exactly one person in the function, no more and no fewer. A solo operator is structurally overstretched by default.

Tier 1 is not "high score". It is a category of evidence.

Tiers are independent paths, not stacked points

This is where most models break, and it is a maths problem rather than a philosophy problem. The instinct with points is to stack: fit plus signal plus intent, sum, threshold. Which means a Tier 1 account with a live job ad but average firmographics falls below a Tier 3 account with perfect firmographics and no evidence of anything. You just deprioritised the only account that told you it has the problem.

A Tier 1 signal does not need a business signal stacked under it. Requiring both collapses your addressable market and buries your best evidence. Run the tiers as separate paths into the same routing layer. Fit qualifies, signal prioritises, and they stay two numbers. Same discipline as the ICP_score and outreach_intent_score split we install in HubSpot: fit is stable and slow-moving, intent is volatile and fast-moving, and blending them lets a strong reading on one axis mask a zero on the other.

The signal that looks great and is not

Social listening is the one everyone reaches for first. Buy an intent tool, watch LinkedIn, catch people complaining about the problem you solve, call them. We built the case for it, then killed it, and the reasoning generalises.

Nobody posts their own mess. For a hyped problem, people post. For an embarrassing one they do not. "Our CRM is a mess and nobody owns it" is a confession, not a brag, so for operational problems the authored pain signal barely exists.

Engagement is your own profession. We pulled the reactor lists on two large RevOps posts, roughly 730 and 560 reactions. Both were overwhelmingly practitioners, GTM engineers, CRM consultants and competing agencies. Automating that signal would have automated surfacing our own industry, on a subscription.

The buyer you want is heads-down in the mess, not engaging with content about it. Pull the reactor list before you buy the intent tool.

Four rules that stop the model manufacturing false positives

An outbound scoring model does not fail loudly. It fails by confidently promoting accounts that do not qualify, and you find out on the call.

Exact title matching, never fuzzy. Fuzzy matching once returned an account with "17 RevOps people". It was four sales ops staff and a lot of near misses. Define the strong title set explicitly (Revenue Operations, RevOps, Sales Operations, GTM Operations, plus Manager, Analyst, Specialist) and treat ambiguous titles as ambiguous. GTM Engineer is the clearest case: hyped, poorly defined, and half the reqs are outbound campaign work rather than systems work. It counts only if the job description describes CRM architecture, data hygiene, lifecycle, integrations, forecasting or attribution.

Gate on data coverage. Unknown is not no. Headcount and title counts come from third-party indexes with uneven coverage. A poorly indexed company returns zero, and zero reads as "no function here", which is your highest value signal. A false zero manufactures your best tier out of a data gap. If the company is not properly indexed, drop it to base rather than guess. Same principle as our enrichment guardrail: mark a propensity signal "yes" only with a verifiable public source, otherwise "unknown", never "no". A guessed no silently deletes real buyers.

Require a reachable persona before the tier is real. A perfect signal on an account where you cannot reach a decision maker is research, not pipeline. On one of our own batches we produced 23 signal-qualified accounts and shipped 18, because five had no reachable leader persona. They went to a held file rather than getting padded in to hit a number.

Do not stack proxies as hard filters. Growth rate, funding, stack sprawl and sales team size are prioritisation notes on the row. Turn them into gates and volume collapses while you are still testing whether the primary signal works at all.

Those rules cost volume, and that is the point. On one run a structural signal yielded 7 qualifying accounts from 101 well indexed candidates, and 3 survived the persona requirement. Roughly 3 percent. That reads like failure until you price the alternative at 500 to 550 dials per booked meeting.

One last rule that sounds trivial and is not: once a tier label is defined, never redefine it. Add new sub-signals under new labels instead. If Tier 1 means something different in March than in June, you can never read which signal books meetings.

Wire it into the CRM or it is a spreadsheet

A ranked list is a snapshot. Three months later half of it has never been touched and nobody can say whether the rankings still hold. Scoring becomes a system only when it drives assignment:

• Intent written from the only real behavior a cold lead produces, the reply. High 10 for a demo or meeting request, Medium 5 for a pricing or lead magnet action, Low 1 for an irrelevant reply. Normalise it so routing is deterministic: a positive reply inside the window goes to the AE or SDR, a negative one suppresses, and high exposure with no reply pauses and recycles.
• Tier surfaced on the record as Hot, Warm, Consider or Ignore, so a rep sees a decision rather than a number they cannot explain.
• Tier riding along as a property before enrichment, so you enrich by tier and stop paying to enrich accounts you were never going to work. Qualification is the cheapest column, enrichment the most expensive. Put them in that order.
• Nothing writes to the CRM without a verified email. That gate protects sender reputation and the score itself.
• SLA enforced by automation, not rep vigilance. Untouched inside the window alerts and re-routes, nothing in 24 hours escalates, and every assignment is timestamped, so "I didn't see that one" becomes a measurable event instead of a shrug.
• First and last touch date, channel and campaign name synced live from the outbound tools and verified with a test contact. Without that capture layer you can score and route perfectly and never learn which signal produced revenue.

What good looks like

Two checks, and they are the only ones worth arguing about.

The backtest: your top 10 percent of scored leads should map to at least 80 percent of closed-won. Score the lead as it looked when it was a lead, not as the record looks today after the deal closed and everything got enriched. Train on the post-close record and you build a model that predicts the past perfectly and the future not at all.

The QA gate: 100 percent of new leads scored and routed inside the SLA, a 30-day monitoring log, and a manual audit of the top 50. If a human cannot explain why the top 50 are the top 50, reps will not trust the number, and an untrusted score is worse than no score because it still costs money to run.

We have built versions of this inside outbound rebuilds like Device Solutions and routing builds like Elesa+Ganter. The pattern holds both times: the scoring is the easy part. Deciding what counts as evidence, then refusing to guess when the data is missing, is the work.

If your outbound list is worked in export order because the scoring model had nothing to say about it, that is a wiring problem, fixable in weeks rather than quarters. That is what our Lead Scoring & Routing Engine module installs. If you want to know what the current order is costing you first, start with the 30-minute GTM Maturity and Revenue Leak Assessment.