Pre-signal outbound email: what differs between won and lost

Comparing outbound emails sent during the acquisition phase — before any welcome, invitation, or invoice signal on the opportunity — across closed-won and closed-lost opportunities in the trailing twelve months. 750 won and 1,500 lost emails were scored against a 10-dimension rubric; the full pre-signal populations were used for deterministic features and template usage.

Largest won-favoring gap: personalization beyond greeting (Δ +0.55). Only meaningful lost-favoring gap: social proof / peer specificity (Δ -0.32).

How the cohorts are defined

Each cohort is at the email level. For each opportunity, we keep [Out]-direction emails sent before the opportunity's first welcome / invitation / invoice / onboarding signal (detected in subject or body). Calendar invites and short scheduling-acknowledgement Tasks were filtered out — those are scheduling artifacts, not outreach. The W50-sent portion of each body was extracted by stripping reply chain tails; the rubric scores only that portion.

Cohort definitions exactly:

• Won. Tasks of Type=Email on opportunities present in the closed-won pull (stages including Invoice Paid, Invoice Sent, Backfill Converted, Invoice Paid - Partial, Invoice Created, Deal Closed - Won), where the email subject contains the Outreach.io marker [Out] and was sent before the opportunity's first signal marker.
• Lost. Same filters applied to opportunities in the closed-lost pull (stages: Prospect Contacted - Closed Lost, Deal Closed - Lost, and a small tail of New Member Cancelled / Backfill Declined). 84% of the lost pool falls in Prospect Contacted - Closed Lost (cold-outreach-that-never-engaged); 16% in Deal Closed - Lost (engaged-but-didn't-sign). Both are pooled per the analysis brief.

What we observe — headline rubric distributions

Each email scored 0/1/2 on ten dimensions. Bar = cohort mean (0–2 scale). Cell label = share of emails scoring the maximum (2). Δ > 0 = won-higher; Δ < 0 = lost-higher.

Observational read

• Won-higher on: personalization beyond greeting (0.88 vs 0.33, Δ +0.55; 14% vs 5% score 2); subject specificity (1.15 vs 0.75, Δ +0.40; 35% vs 18% score 2); risk reduction language (0.45 vs 0.22, Δ +0.24; 7% vs 7% score 2); named anchor / hook strength (1.11 vs 0.88, Δ +0.23; 37% vs 27% score 2); reason for reaching out (0.87 vs 0.66, Δ +0.20; 16% vs 11% score 2); trigger reference (0.51 vs 0.38, Δ +0.13; 16% vs 9% score 2).
• Lost-higher on: social proof / peer specificity (0.59 vs 0.91, Δ -0.32; 11% vs 27% score 2).
• Indistinguishable between cohorts (|Δ| < 0.12): cta specificity (Δ +0.08), length & scannability (Δ -0.01), member content density (Δ +0.02).
• Directional, not causal. Won opps may receive higher-rated outreach because polished outreach wins; OR because easier-to-close prospects attract more polished outreach; OR both. The sequence-position cuts below narrow but do not resolve this.

What we observe — by sequence position (1st outbound vs follow-ups)

Splitting each cohort by the email's position within its opportunity's outbound sequence. The first outbound is the cold first-touch; later positions are bumps and reconnects. Comparing won-1st to lost-1st (and won-bump to lost-bump) partially controls for the fact that won opps tend to engage quickly, so their pre-signal sequences are shorter.

* cell n < 10 — direction only, not magnitude.

What we observe — deterministic features

Computed mechanically on the W50-sent portion of each email. No LLM judgment involved.

What we observe — per-W50-Group cuts (top 12 by sample)

Rubric-total (sum of all 10 dimensions, max 20) by group. Use these as directional only when a cell shows small n.

What we observe — template usage

The Outreach.io sequences in both cohorts heavily reuse identical body templates. We fingerprinted each email body (normalizing whitespace, stripping greeting + sign-off, taking the first 250 characters of opener) and counted appearances per cohort. Per-1k rates are normalized against each cohort's full pre-signal pool (won = 3,154 emails, lost = 400,844 emails) so they are directly comparable. Lift = won_per_1k ÷ lost_per_1k — values above 1.0 indicate a template appears proportionally more often in won than in lost.

Two cuts: (A) the templates that dominate the lost pile by raw frequency, and (B) the templates that appear proportionally more in won than in lost.

Interpretation note: the most heavily-used templates in lost — voicemail-follow-up variants, "missed note?" bumps, peer-list cold pitches — collectively account for tens of thousands of sends with near-zero won representation. This does not mean these templates "cause" losses; opps that receive many follow-up bumps may already be hard targets. But it does establish that the bulk of templated lost outreach is not converting, and the templates that do appear in won are different ones (typically post-call follow-ups referencing a real conversation, or specific named referrals).

What the data supports examining next

• Pull the templates used most often in each cohort separately. If the same templates appear in both, the differentiator is targeting / timing / sequence cadence, not content.
• Layer Outreach.io engagement metadata (open rate, click rate, reply rate) onto this rubric-scored sample to test whether high-rubric emails actually drove engagement, or whether engagement was driven by other signals.
• For dimensions where won and lost are statistically indistinguishable, deprioritize as actionable craft levers.
• For dimensions where one cohort scores meaningfully higher, sample the top-decile and bottom-decile within the other cohort and read what's different in concrete terms.
• For Phase 2 (the never-reached-Call-1 question), re-run on the inbound-to-stalled cohort once that pull arrives, comparing it back to the won pre-signal cohort here.