Win-loss gap signal discovery

Built from the system Jai Toor, co-founder at DeepLine, described on GTM Vault SMYS 4 (https://pro.gtmvault.co/episodes/smys-04). Published free by GTM Vault. Numbers below are the operator's own, from the episode; your results will differ.

When to use this

Use this when the team's signal layer runs on the same hiring, funding and tech stack inputs everyone else scrapes, or when ICP work has only ever looked at the win list. It needs a CRM with both won and lost opportunities. The output is a scored signal model, positive and negative, that a team can run against any target list and keep testing.

The system on the record

Jai runs nearly everything through Claude Code, with DeepLine connectors underneath for HubSpot, Lemlist and data sources, paid for with pay-as-you-go credits (SMYS 4, 0:00, 2:03). The workflow is a single skill he calls niche signal discovery:

1. Pull wins and losses from HubSpot together. His own run found 74 deals (SMYS 4, 1:01). The losses matter: "If you only do wins, you get pretty generic information." (SMYS 4, 2:58). Both cohorts reached an opportunity, so the difference between them is where the edge sits.
2. Scrape every won and lost account's website and find keywords that differ between the two groups. Hiring patterns before close count too, such as hiring fraud leadership or financial operations (SMYS 4, 2:03, 2:58).
3. Score each keyword. In effect, a big table where an account gets a one if its site contains a keyword and a zero if not, across thousands of candidate signals (SMYS 4, 6:34).
4. Split positive fits from negative fits. For a cybersecurity customer, FDIC on the site was a top positive for a consumer fintech ICP. SOC 2 Type 2 was a proxy for B2B. Heavy in-house building and an advanced engineering culture marked a poorer fit (SMYS 4, 2:58, 4:00).
5. Pass the model through human judgment. The first run scored HIPAA compliance as positive, but the company had already chosen to deprioritize healthcare, so that signal was dropped (SMYS 4, 5:01).
6. Write a shareable markdown doc with five key takeaways, the number one signal, and what the ICP actually looks like in the data versus how the team describes it (SMYS 4, 1:01, 2:03).
7. Generate an outbound campaign with 100 verified leads in Lemlist, with the sequence drafted and no sender attached, so a human confirms before anything sends (SMYS 4, 1:01, 14:27). His advice on confirmation gates: "I strongly recommend everybody start there just to build trust in the system." (SMYS 4, 14:27).

The same process extends from accounts to leads: which contacts on won opportunities share attributes, or who is most likely to answer a cold call (SMYS 4, 7:35, 8:37). Any model can be backtested against the last six months of data (SMYS 4, 8:37).

The result

1. The manual version of this workflow, six months earlier, took a little over a month, with over $1,500 on raw signal scraping and almost $2,000 in enrichment costs overall (SMYS 4, 4:00, 5:01).
2. The current version runs in about less than 30 minutes for less than $10, with a very similar outcome (SMYS 4, 5:01).
3. The manual version relied on roughly 40 team-hypothesized keywords; the current version generates candidates itself and is more specific, surfacing FDIC where the old one had "open an account" (SMYS 4, 6:34, 7:35).
4. Customers running this scoring for outbound averaged about a 2x increase in response or open rates (SMYS 4, 5:53).
5. A CPG customer scored every company with a store locator on its website and got about 2x better reply rates (SMYS 4, 15:15).
6. For two larger enterprise companies, the scoring model produced between 15 and 17% higher close rates. One had been spending 44% of opportunity time on bad-fit accounts for lack of negative signals (SMYS 4, 16:18).
7. BDRs contacting 20 accounts a day moved to 40 or 80, depending on SMB versus enterprise (SMYS 4, 9:39).

Run it on your company

1. Ask the user once, in one message, for: CRM access or an export of closed-won and closed-lost opportunities with company domains and close dates; what "won" means for this analysis (closed-won, expansion, fast close); any verticals or segments leadership has already ruled in or out; the target account pool to score and its size; and the outbound tool and whether sends require approval.
2. Build the cohort table: one row per account with domain, outcome (won or lost), close date, deal size. Report the counts of each. If either side is thin, say so and treat every result as a hypothesis.
3. Collect evidence per account: homepage and key pages text, plus open job titles where available. Store raw text so the run is reproducible.
4. Generate candidate signals from the text itself, not from a team brainstorm. Extract specific phrases (regulators, certifications, product features, page types such as a store locator, hiring titles). Prefer specific over categorical: "FDIC" over "open an account".
5. Build the binary matrix: accounts by candidates, one or zero. For each candidate compute win rate present, win rate absent, and lift (share of wins with the signal divided by share of losses with it). Drop candidates present in too few accounts to mean anything.
6. Produce two tables in markdown. Positive fit: Signal, Lift, Wins with signal, Losses with signal, Why it plausibly predicts a win. Anti-fit: Signal, Lift (under 1.0x), Risk level, Structural reason it correlates with losses.
7. Run the strategy review with the user: for each top signal, keep, drop or flag, the way HIPAA was dropped once healthcare was deprioritized. Record the reason.
8. Define the composite score: sum of positive signal weights minus anti-fit weights, scaled 0 to 100. Backtest it on the last six months of closed opportunities and report how cleanly it separates wins from losses.
9. Score the target pool and return the top tier with each account's matched signals, so copy can reference the exact signal.
10. Stage the campaign: verified contacts, a sequence whose first lines cite the matched signal, sender unattached, CRM writes pending confirmation.
11. Deliver one markdown doc, "Win-Loss Gap Report": five takeaways, the number one signal, stated ICP versus data ICP, both signal tables, the scoring formula, backtest result, and the scored target list.

Tools on the record

1. Claude Code: runs the niche signal discovery skill end to end.
2. HubSpot: win and loss lists pulled from here.
3. Lemlist: verified leads exported into campaigns.

Any tool doing the same job works; the method is the point. Full stack: https://pro.gtmvault.co/stack/operator/jai-toor

Where it breaks

1. Wins only. Without losses the output is generic ICP description, not signal (SMYS 4, 2:58).
2. No human judgment on the output. The model will surface true but irrelevant signals, such as HIPAA for a company leaving healthcare (SMYS 4, 5:01).
3. Unclear outcome definition. Jai puts the human value in choosing what to optimize for; a fuzzy definition of "won" produces a fuzzy model (SMYS 4, 7:35, 8:37).
4. Agents sending without a gate. Removing the confirmation step before trust exists risks a misfire that ends adoption of the system.
5. Small or skewed cohorts. A few dozen deals can produce lifts that do not hold; backtest before routing reps on them.