How should we rank target accounts, and how do we know it's right?

Our first three scoring rules tiered accounts by job-posting volume, by the marketing stack on their site, and by the GTM leaders they'd hired. We checked the tiers against the pursue or pass calls of our founder, Sam Gong, on 51 accounts. Of the 23 he chose to pursue, none sat in the top tier and 19 sat in the bottom.

Posting volume tracks headcount, so the biggest companies floated up. If your top tier is the logos and your sellers keep private lists, check what your score is really sorting by.

A hiring post for an AI transformation seat in marketing or GTM names the problem in the account's own words and the team that owns it, and it carries a date. On August 24, 2026, one account's job titles told us it had no AI work in GTM. The body of one of its postings described a team being built to move marketing work to agents.

Rank on fit and sequence on timing. If one number has to carry both, show its parts, or a seller can't tell whether an account ranks high because it fits or because something happened last week.

Every account on our list carries a band, its reasons in words, and a coverage state. In August 2026, 453 of the 612 accounts we scored read as thin: fewer than 25 job posts, too few for the full model to read what the company is hiring for. A thin account keeps the band free company data gives it, with the thin mark beside it. It moves up only on a signal or a person's read, and nothing marks it down for being unreadable.

That's the rule, not the whole result. Our outbound goes to the full model's top band, and a thin account reaches it only through a person's read. We don't yet count how many thin accounts got one, so in practice coverage still steers a lot of our attention. Count yours.

Our gate set aside 9 of 90 labeled accounts, and all 9 were passes. We left the catch-all industry codes ungated because they hold good accounts. Nine labeled passes don't show that the unlabeled accounts a gate removes are passes too. Read a sample of what any exclusion removes before it goes into a plan.

A large company's homepage may describe the parent when a division does the buying. Find the unit that buys before you grade it. Then ask RevOps what your model does with an account it can't see. If missing job posts lower a score by themselves, your bottom tier mixes poor fit with poor coverage. Before the plan locks, someone reads the thin accounts in the top territories by hand, or they stay marked as unread.

One rule nearly shipped: rank up companies hiring for three or more different GTM operations jobs. It looked right on 17 of 18 accounts. We sealed it, changing nothing, and ran it on 50 accounts it had never seen. Among the 23 software companies in that sample, where it had to work, it called 13 right. A rule that says “pursue” to everything called 16. It also missed all three accounts Sam rated highest: it said pass on one and couldn't measure the other two, because they post few jobs. Few posts means the screen can't see. It doesn't mean the account doesn't fit.

Keep three checks apart on your own list.

• Inspection. Can anyone say why an account sits where it does, and is the reason current?
• A past quarter. Score it using only what was knowable before it started, and compare it with what happened among accounts sellers worked. The last section walks through it.
• The next quarter. Freeze two rankings of the same list today, the score and the one you'd use without it, and date both. Before results arrive, set one cutoff for both from how many accounts your sellers can work. At quarter end, report for each shortlist how many accounts were worked, how much effort they got, and the meetings and opportunities that came of it. Track wins as they close. Unworked accounts stay unknown. It's the only check made before anyone knows the answer, and it still can't prove the ranking caused the results.

Closed deals also reflect execution, price and procurement, and small samples mislead. Our numbers say how a model did against one person's calls on our list. They say nothing about yours. The checks transfer. The numbers don't.

Here is how each band did against our founder's calls. The two models ran on different samples, and we chose their features and band cuts on these same accounts. Read the rates as a direction, not a contest between the models.

The cheap model ruled accounts out well in our sample: 6% of 17. It ruled them in weakly: 75% against a 51% base rate, on 8 accounts. We tried four bands first and dropped them, because the cheap model's top band scored 63% and its second band 79%. A model that ranks weakly should show fewer bands, so nobody reads more precision into it than it has.

We point our own outbound at the full model's top band only, 108 accounts on September 30, 2026. That's a working choice for a team our size, not a tested rule. Below it, nothing moves without a person's yes. Scoring every account is cheap and reading one deeply isn't, so deep reads go to the shortlist. We don't gate on size: a 1,000-employee floor would have deleted four accounts Sam chose to pursue.

We have no territories to cut, so this is the version for a team with sellers, and it stops at capacity. We haven't made a cut or an override ourselves, so we won't show one.

Fit sets the floor: which accounts are worth anyone's time. Deal size, relationships you already have and timing set the order inside it. RevOps sets the cutoff from time, not from the list. Take each seller's selling hours for the planning period, subtract what their current accounts already take, and divide by the hours the chosen play needs per new account. That's how many new accounts each seller can actively work, not every account they own.

Geography, accounts already owned and continuity then adjust the cut. Write every override down with its reason and its owner, so next year's check can tell an override from the model. Thin accounts sit in a review queue until someone reads them. RevOps owns the cut, the CRO owns the plays, and a seller disputes a ranking with a reason attached.

Timing refreshes when a signal lands. Fit gets refit on a schedule. Territories move at planning, not mid-quarter.

Our refit on September 9, 2026, with new data and new weights, moved 75 of the 614 accounts it had scored before, 12%, and grew the top band from 85 to 108. Eleven accounts now disagree with calls Sam had already made by hand. Those eleven are his review queue, not noise.

We got the dating wrong ourselves. By September 9, 2026, our live list held 831 accounts, and our account explorer still showed its August 24 snapshot of 614. It looked live for two weeks.

1. Pull last quarter's target list as it stood on the first day of the quarter. The dated copy, not today's.
2. Take ten accounts each from the top, the middle, the bottom, and the accounts the score couldn't read. Ask RevOps why each sits where it does.
3. Count the gaps: every account nobody can explain, every reason nobody can source or date, and every account scored low only because the model couldn't read it.
4. Take only the accounts your sellers actually worked last quarter. Rank them twice, by the score and by company size, each as it stood on the quarter's first day. Compare the top quarter of each on meetings, opportunities and wins. If size does as well, find out what the score adds before you build territories on it. Sellers chose which accounts to work, so this is a diagnostic, not proof.
5. If no dated copy exists, that's the first finding. Save both rankings this week, set one cutoff from seller capacity and a date to read them, and record work and results from today.