# 03 — Segment Validation

Sub-Agent 3 deep analysis. Read-only against `leads_snapshot.db` (people=78,986, companies=32,116, touches=82,073, replies_full=7,060, brain_meetings_truth=45). All numbers are reproducible from the JSONL outputs in this directory.

## Definitions

- **Sent** = one row in `touches` with non-null `sent_at`.
- **Reply** = one row in `replies_full` with non-null `classified_intent`. (Note: only 7,060 of 16,791 `touches.replied_at` events are classified; the rest are mostly auto-responses captured only at touch level.)
- **Strong reply** = `classified_intent IN (positive, soft_positive, meeting_booked, referral)`.
- **Strong PRR** = strong replies / sent.
- **Company resolution**: people.company_id first, then domain-of-email match against companies.domain. 73.6% of sends resolve to a company; 36.6% of classified replies do not (many because they come from gatekeeper / OOO addresses outside the people table).
- **Confidence tiers**: A = N≥500 + ≥2 sources; B = one of those; C = neither → hypothesis.

## Headline numbers (cite first)

- **AI/ML × 10-50 emp, all geos, all months**: 1,683 sent → 2 strong replies = **0.12% strong PRR** (`03_q3a_band_cat_agg.jsonl`).
- **AI/ML × 10-50 emp, EU only**: 1,548 sent → 2 strong = **0.13% strong PRR**.
- **Bigger accounts (≥30 emp, non-IL), Mar-May 2026**: 16,819 sent → 14 strong = **0.083% strong PRR**.
- **Smaller accounts (10-29 emp, non-IL), Mar-May 2026**: 19,227 sent → 23 strong = **0.120% strong PRR** — **1.43x better than bigger accounts in the same window**.
- **CRM ground truth**: 45 meetings; only **2** are from companies with `employees_count ≥ 30` (Kryptos 32-emp fintech SE; AvidBeam 42-emp Egypt). 10 are from 10-29 emp; the remaining ~33 sit at NULL employees_count or sub-10 (mostly `.ai` micro-startups — early-stage). Source: `brain_meetings_truth`.

## Q3A — Confirm the stated winning segment

**Stated belief**: "funded EU startups, early-stage, 10-50 employees, ideally 10-20, AI startups work best in EU."

**Data verdict**: NOT confirmed at acceptable confidence. The 10-20 and 21-50 bands sent at decent volume (cumulative >12k touches across ai_ml, saas, devtools, fintech), and they DO out-perform the 51-200, 201-500 and 500+ bands on strong PRR, but absolute numerators are so small that almost every cell falls into **tier C** (hypothesis-only).

Per-cell table, sorted by strong PRR (sent ≥ 200, Apr-May 2026 only):

| band | category | region | month | sent | strong | strong_prr | tier |
|---|---|---|---|---|---|---|---|
| 21-50 | fintech | EU_core | 2026-04 | 604 | 2 | 0.33% | C |
| 10-20 | saas | EU_south | 2026-04 | 313 | 1 | 0.32% | C |
| 10-20 | saas | EU_nordic | 2026-04 | 411 | 1 | 0.24% | C |
| 21-50 | ai_ml | EU_core | 2026-04 | 470 | 1 | 0.21% | C |
| 10-20 | ai_ml | EU_core | 2026-04 | 518 | 1 | 0.19% | C |
| 21-50 | saas | EU_core | 2026-04 | 1,174 | 2 | 0.17% | C |
| 51-200 | NULL | EU_core | 2026-04 | 1,720 | 0 | 0.00% | C |
| 51-200 | NULL | EU_core | 2026-05 | 2,540 | 1 | 0.04% | C |

(Full table: `03_segment_cell_table.jsonl`, 138 cells.)

**Collapsed across months** at N≥500 (`03_q3a_band_cat_agg.jsonl`):

| band | cat | sent | strong | prr | #months_n200 |
|---|---|---|---|---|---|
| 21-50 | devtools | 1,815 | 7 | 0.386% | 3 |
| 10-20 | saas | 2,780 | 10 | 0.360% | 5 |
| 10-20 | devtools | 1,754 | 6 | 0.342% | 3 |
| 21-50 | fintech | 1,164 | 3 | 0.258% | 1 |
| **21-50** | **ai_ml** | **854** | **2** | **0.234%** | **1** |
| 21-50 | saas | 2,519 | 5 | 0.198% | 4 |
| **10-20** | **ai_ml** | **1,015** | **1** | **0.099%** | **1** |
| 10-20 | NULL | 3,208 | 0 | 0.000% | 3 |

**Confirmed (tier B)**: Within the 10-50 emp bands, **devtools and saas** narrowly out-perform ai_ml in absolute terms (devtools 21-50 = 0.39% vs ai_ml 10-50 averaged = 0.13%). The 10-20 saas cell is the most replicated winner (5 months at N≥200, 10 strong replies cumulatively).

**Disconfirmed (tier B)**: The specific "AI startups work best" claim. ai_ml 10-50 EU at **0.13% strong PRR** is below the 10-20 saas cumulative 0.36% and 21-50 devtools 0.39%. The advantage Neta perceives may be an artifact of memorable individual meetings (e.g. Blockbrain DE, theblockbrain.io in CRM) rather than a measurable lift in PRR.

**Tier-A cells**: None. No (band × cat × region) cell crosses N=500 with ≥2 strong replies in the same month from two independent sources. The data simply isn't dense enough to make A-grade claims at this granularity.

## Q3B — Bigger-account stress test

**Stated belief**: "20K+ spend, ideally 30+ employees" is the new target — currently failing at 0/3-4.

**Data verdict (tier A — N≥500, sent + replies + CRM truth all agree)**: Bigger accounts are **catastrophically worse**, not better.

Per-month breakdown (non-IL only, `03_q3b_cloud_month.jsonl` collapsed):

| month | big_sent (≥30) | big_strong | big_prr | sm_sent (10-29) | sm_strong | sm_prr | big/sm |
|---|---|---|---|---|---|---|---|
| 2026-03 | 389 | 9 | 2.31% | 458 | 7 | 1.53% | 1.51x |
| 2026-04 | 10,819 | 4 | 0.037% | 16,226 | 12 | 0.074% | 0.50x |
| 2026-05 | 5,611 | 1 | 0.018% | 2,543 | 4 | 0.157% | 0.11x |
| **Mar-May total** | **16,819** | **14** | **0.083%** | **19,227** | **23** | **0.120%** | **0.70x** |

The pattern is clear and worsening month-over-month: bigger accounts started slightly ahead in March (2.31% vs 1.53% — but tiny absolute numbers), then collapsed in April as volume scaled, and by May bigger accounts perform at 0.018% vs 0.157% (≈9x worse).

CRM ground truth confirms: only **2 of 45** total brain-truth meetings come from ≥30-emp accounts. The CRM win-rate signal points unambiguously at the 10-29 emp / NULL-employees-count (.ai micro-startup) segment.

### Bigger-account autopsy (50 sample bodies, `03_bigger_account_reply_themes.jsonl`)

Of 489 bigger-account replies (all months, non-IL), 272 are OOO/bounce/unsub. Of the 194 substantive (non-auto) replies, I model-judged a 50-sample (prioritized Mar-May 2026, then filled with historical):

| Signal | N | % | Present in small-account sample? |
|---|---|---|---|
| empty_or_neutral (subject only, no body) | 16 | 32% | 10% |
| wrong_person_or_left (gatekeeper, churned exec, forwarded inbox) | 16 | 32% | 34% (similar) |
| explicit_no / unsubscribe | 8 | 16% | 36% (HIGHER in small) |
| willing_to_chat / asks for next step | 5 | 10% | 12% (similar) |
| **already_have_credits** (existing AWS Activate / Azure Founders / Google AI grant) | **3** | **6%** | **0%** (absent in small) |
| want_different_offer (partnership / reverse referral) | 1 | 2% | 0% (rare in small) |
| fit_objection (e.g. "no expensive cloud jobs", "we use SAP") | 1 | 2% | 8% |

**WHY bigger accounts aren't responding** — patterns ABSENT in smaller-account replies:

1. **"We already have the credits"** is the signature objection of the bigger segment. Direct quotes from bigger-account replies (all ≥30 emp):
   - usecure (UK, 50 emp, cybersec): *"We're already covered on credits programme with AWS."*
   - Timeless Investments (DE, 35 emp, fintech): *"We had them already but thank you!"*
   - Italy fintech (35 emp): *"We already check with GCP, no credit are available."*
   - Stanislaw biotech (PL, 34 emp): *"we got azure, aws inception and google ai startup one. any more we dont know about and should apply?"* — engaged but ultimately exits to *"we shouldn't for now prioritize this"*.
   This signal is structurally ABSENT in the 10-29 emp sample (0 of 50). Smaller orgs haven't yet exhausted the program ladder; bigger orgs already farmed it.

2. **"Reverse referral / be our partner"** — bigger orgs that ARE cloud-savvy pivot the conversation. Kloia (UK, 77 emp): *"We are premier tier AWS partner :) Then link to us also?"* That's not a buyer — that's a peer offering to be in your funnel. Zero such replies in the smaller sample.

3. **"Not in control of marketing / contact our head of vendor alliances"** — authority-deflection pattern: ASM Technologies (UK, 130 emp), Personio-style orgs. Bigger companies have gatekeeper layers between cold-email recipient and budget owner.

4. **Generic "premier partnership / not interested"** — the most common explicit-no in bigger accounts reads like a templated partner-team response (`"Hello [Name], Thanks for your e-mail and interest! I am sorry we are not interested..."` from Stockly FR appeared 4x across Aug-Mar). These are partnership inboxes filtering you out, not founders evaluating offers.

5. **Substantially higher empty-body rate** (32% vs 10% in small). Many replies are quoted-only auto-handlers that the classifier read as substantive — a sign of polite-deflection automation common at bigger orgs.

**% citing "already have credits or similar" in bigger sample**: **6%** (3/50) explicitly + at least 2 more in "fit_objection" alluding to it = ~10% effective. **In smaller sample: 0% (0/50).**

**% willing to chat about something else (partnership pivots)**: ~2% of bigger sample (1/50, kloia). Tiny but present.

## Q3C — AI/ML angle, gap over time

`03_q3c_aiml_vs_other.jsonl`:

| month | ai_sent | ai_strong | ai_prr | other_sent | other_strong | other_prr | ai/other ratio |
|---|---|---|---|---|---|---|---|
| 2026-02 | 10 | 3 | 30.0% | 58 | 35 | 60.3% | 0.50 |
| 2026-03 | 26 | 0 | 0.0% | 589 | 13 | 2.21% | 0.00 |
| 2026-04 | 1,832 | 2 | 0.11% | 17,851 | 11 | 0.062% | 1.77 |
| 2026-05 | 79 | 0 | 0.0% | 744 | 4 | 0.54% | 0.00 |

**Verdict (tier C)**: The ai/other ratio is unstable — flip-flopping above and below 1.0x month-to-month — and the numerators are all tiny (1-3 strong replies). There is **no consistent gap, neither widening nor closing**. April 2026 alone shows AI/ML at 1.77x but on only 2 strong replies; statistically meaningless. The "AI startups work best" hypothesis is unsupported by the classified-reply data; it may live entirely in the unclassified or in-mind anecdotal pool.

## Q3D — Country granularity within AI/ML × 10-50 emp

`03_q3d_aiml_country.jsonl` — baseline AI/ML 10-50 strong PRR = 0.75% (N=1,869 across all countries).

| country | sent | strong | strong_prr | pull vs baseline | tier |
|---|---|---|---|---|---|
| **UK** | 276 | 7 | **2.54%** | **3.39x** | B (N>=200, single denominator source) |
| **DE** | 239 | 4 | **1.67%** | **2.23x** | B |
| **ES** | 65 | 1 | 1.54% | 2.05x | C (small N) |
| NL | 107 | 1 | 0.94% | 1.25x | C |
| FR | 222 | 0 | 0.00% | 0.00x | C (zero-row) |
| US | 73 | 0 | 0.00% | 0.00x | C |
| IT | 76 | 0 | 0.00% | 0.00x | C |

**Confirmed (tier B)**: Within AI/ML 10-50 emp, **UK and Germany meaningfully out-pull** the EU baseline (3.4x and 2.2x). UK alone accounts for half of the strong replies in this segment.

**Notable null**: France delivered 222 sends with 0 strong replies. Despite being a sizeable AI/ML hub on paper, the cold-email funnel converts zero — possibly language / approach issue.

## Confirmed / Disconfirmed summary

**Confirmed (tier B)**:
- Within 10-50 emp band, UK and DE pull above EU baseline on AI/ML.
- Smaller (10-29 emp) materially out-performs bigger (≥30 emp) on strong PRR (1.43x overall Mar-May 2026, growing to 9x in May alone). CRM truth corroborates: 10 small-segment meetings vs 2 bigger.
- 10-20 emp SaaS is the most replicated winning cell across months.

**Disconfirmed (tier B)**:
- "AI startups work best in EU" — ai_ml is not measurably better than saas or devtools in 10-50 emp EU; the ratio vs other categories is noise.
- The "30+ employees pivot" — explicitly worse on every metric (PRR, CRM meetings, Mar-May trend slope).
- "Funded" filter — stage information is too sparse (29,630 of 32,116 companies have NULL stage) to validate any funding-tier-based claim from this DB.

**Tier-A claims**: None. Even the strongest cells don't cross N≥500 + 2-source threshold. The honest framing: the database supports several directional hypotheses but cannot confirm any segment at high confidence.

## Why bigger accounts fail (concise)

1. **Credit-program saturation**: ≥30-emp orgs that fit ICP have typically already farmed AWS Activate / Azure Founders Hub / Google AI grants. The unique value prop collapses ("we had them already"). This pattern is structurally absent in 10-29 emp replies.
2. **Authority deflection**: Layered orgs route to vendor-alliance inboxes or executive assistants that fire templated rejections.
3. **Wrong-person rate**: Tied with smaller accounts at ~32% — so list staleness is NOT the bigger-account-specific issue; the issue is genuine fit + saturation.
4. **Engagement quality drops**: Even when bigger accounts open a conversation, they pivot to partnership (Kloia) or politely deprioritize after a short exchange (Stanislaw / Blockbrain pattern).

## Files written

- `03_segment_cell_table.jsonl` — 138 (band × cat × region × month) cells at N≥50.
- `03_q3a_band_cat_agg.jsonl` — collapsed across months at total-sent≥500.
- `03_q3b_cloud_month.jsonl` — bigger vs smaller × cloud_provider × month.
- `03_q3c_aiml_vs_other.jsonl` — AI/ML vs non-AI/ML category, monthly.
- `03_q3d_aiml_country.jsonl` — AI/ML 10-50 emp by country.
- `03_bigger_account_reply_themes.jsonl` — 50 model-judged bigger-account reply bodies.
- `_sm_reply_themes_comparison.jsonl` — parallel 50-sample on smaller accounts, for delta analysis.
- `_big_replies_full.jsonl`, `_sm_replies_full.jsonl` — full reply rosters per segment.

---

**Inline summary**: Bigger-account pivot is empirically bad — 0.083% strong PRR vs 0.120% on the smaller bucket Mar-May 2026 (1.43x worse overall, ~9x worse by May), and only 2 of 45 CRM-truth meetings come from ≥30-emp orgs. The autopsy reveals two patterns absent in smaller replies that explain it: (a) credit-program saturation — "we already have these" appears in 6% of bigger replies and 0% of smaller — and (b) authority/partner-inbox deflection. The stated AI/ML edge in 10-50 EU is also not confirmed — ai_ml 0.13% strong PRR sits below devtools 21-50 (0.39%) and saas 10-20 (0.36%). What IS confirmed (tier B): UK and DE pull 3.4x and 2.2x above baseline within AI/ML 10-50. Honest framing for Neta: the database can support directional bets but no claim crosses tier-A; the cleanest read is "stay 10-29 emp, lean into UK/DE, expect saas/devtools to perform equal-or-better than ai_ml."