Methodology — The 2026 AI Founder Brief

Founder's AI Blueprint, Issue 001. This page is the full methodological backing for the main article and Quick Take. It exists separately because the article should be readable in 8-12 minutes and the methodology takes another 4-5 minutes. Readers who want to scrutinize the sample frame, the AI-native classification logic, the data-quality decisions, or the source citations come here.

Who should read this page: anyone who wants to verify a specific number from the article, replicate the analysis, propose a correction, or assess whether our universe-coverage caveats apply to their own use of the findings.

---

Sample frame

The working universe is the union of two PitchBook saved-search snapshots, dated March 4, 2026 and May 25, 2026. The snapshots were pulled with different saved-search criteria — only 32% of companies appear in both — so we union them by company ID and deal ID, then dedupe (NEW snapshot wins per field where both have a value; OLD fills NULLs). The combined universe is 18,352 unique companies and 67,609 unique deals.

The universe is curated and Europe-weighted: 44% EU, 23% US, 10% UK, 1% Israel, 23% other. This differs from PitchBook's typical US-dominant balance for venture-capital pulls. Readers should treat the geographic mix as a property of this dataset, not of global AI funding. When we report a geographic number, we frame it as "in our universe" rather than as a global statistic.

Within this universe, the observed 24-month seed-to-Series-A graduation rate is approximately 6% — well below the ~15% Carta reports for the most recent measurable seed cohorts (Q1 2022 vintage; Carta's Q2 2024 cohort at the same checkpoint tracks similarly). The industry-wide rate has declined from ~30% (Q1 2018 cohort) to ~15% (Q1 2022 cohort) over the past six years. Our universe runs below even the contemporary baseline. We attribute the additional gap to under-capture of subsequent Series A rounds in our saved-search universe — a coverage artifact, not a finding about AI funding pace. Whenever the article makes a conversion or pace claim, the cited source is Carta or AngelList directly, not our universe. PitchBook supplies sector composition, deal sizes, valuation distributions, investor activity, and company-level patterns.

AI-native classification

A company is classified as AI-native if:

(a) Its Description, DescriptionShort, or Keywords contains a whole-word match for one of an explicit AI/ML term set: artificial intelligence, machine learning, generative AI, LLM, large language model, foundation model, deep learning, neural network, natural language processing, computer vision, AI-native, AI agent, AI assistant, AI copilot, AI platform, AI infrastructure; AND

(b) It survives manual Stage-3 review applying the core moat test: removing the AI/ML model from the product breaks the core value proposition.

We exclude foundation labs raising $100M or more in a single round (OpenAI, Anthropic, Mistral, xAI, Cohere, AI21, Inflection, Stability, Together, Reka, Adept, Black Forest Labs) because their fundraising dynamics distort the distribution and they are not peers for most readers.

17 additional companies passed the keyword filter but failed the core moat test and were manually excluded: ByteDance, Baidu, AMD, PTC, DigitalOcean, Genpact, Caris Life Sciences, View (San Jose), Quantum (Software Services), Zeta Global Holdings, Planet Labs, Kraken Technologies, Elastic (Database Software), Astera Labs, Ayar Labs, Replit, Xiaoyu Intelligent Manufacturing.

The working AI-native population is 1,223 companies, of which 491 had priced rounds in our 2023-01-01 to 2026-03-31 analytical window.

Sub-sector classification

Companies are tagged into seven sub-sectors via a description-only regex heuristic with manual overrides applied for the 19 article-cited companies (full override list available in the codebase). The regex went through three iterations to address word-boundary errors in earlier versions; the locked v3 distribution is:

The ai_agents sub-sector has a ~35% audit miss rate because the phrases "ai agent" / "agentic ai" appear in long Description text of many vertical-AI companies that mention agents tangentially (e.g., cybersecurity products describing "agentic AI for threat detection"). Tightening further would shrink the bucket below n=10 entirely. Article disclosure: every agents-specific finding cites its n inline (n=6 at seed; n=13 at incorporation-to-first-round) and is treated as "signal to watch, not conclusion."

Investor lead verification

PitchBook investor lead-attribution data has a non-trivial error rate. All named-investor claims in the article were cross-verified against Crunchbase records and company press announcements where available. We found at least one significant discrepancy and corrected it: the lead investor on Paid (UK)'s $21M September 2025 seed was incorrectly attributed in PitchBook to a follow-on investor; external sources (TechCrunch, EQT Ventures blog) confirm Lightspeed Venture Partners as the actual lead, with EQT Ventures leading the earlier pre-seed. The article reflects the corrected attribution.

Data quality flags

Two companies that PitchBook identified as outlier-seed candidates had data conflicts with external sources and are excluded from the article's named analysis:

Both companies remain in the aggregate counts (universe membership, sub-sector tagging) but are not cited by name in Part 3's outlier roster.

External sources cited in the article

Carta:

• State of Private Markets Q3 2025
• State of Private Markets Q4 2025
• State of Pre-Seed Q1 2026

PitchBook & Crunchbase Q1 2026 venture reports:

• PitchBook-NVCA Venture Monitor Q1 2026
• Crunchbase Q1 2026 global VC report

Live API pricing (Q2 2026):

• OpenAI | Anthropic | Google Gemini | Pinecone | Weaviate | Modal | Replicate

Company engineering blogs and stack disclosures:

• Sierra — Constellation of Models
• The New Stack on the Cursor/Claude Code/Codex coding stack
• Perplexity technical deep dive at RankStudio
• Crusoe on 2026 AI infrastructure

Playbook sources:

• YC Library — Playbook for Building an AI-Native Company
• Lenny's Newsletter — Counterintuitive advice for building AI products
• Stratechery — Agents Over Bubbles
• Tomasz Tunguz — Communication tax of small orgs
• Epoch AI — Revenue per employee at AI companies
• SaaStr on Anthropic's headcount efficiency
• Braintrust — LangSmith vs Braintrust comparison (Notion case study)
• Valtorian — AI Moats in 2026
• TechCrunch — Paid raises $21M seed
• TechCrunch — Meridian raises $17M seed
• Weco — Seed announcement

Reproducibility

The full analysis pipeline lives in the project repository (PitchBook merge → sub-sector classification → chart-stats CSVs → Calculator backend). All scripts use deterministic random seeds for bootstrap CIs (seed=42, 10,000 resamples for chart stats; 2,000 resamples for Calculator cells). The merged CSVs, stage-classifier rules, manual override list, and chart-stats outputs are all reproducible from the published PitchBook export plus the documented exclusion lists.

What this analysis does NOT support

The methodology supports descriptive findings on the specific universe we measured at the specific time we measured it. It does not support:

• Causal claims (companies with higher valuations may have stronger fundamentals or better fundraising operators; we cannot distinguish)
• Prospective claims (the article describes 2023-Q1 2026 funding; it does not predict 2026-Q2 or later)
• Global generalization (our universe is curated and Europe-weighted; the patterns describe the segments we observe)
• Per-founder identification (we report aggregate patterns; we do not call out individual founders by name in pattern findings unless they have made the information public themselves)

Questions about the methodology, suggested corrections, or notable discrepancies between our findings and your own data: reach out via the survey link in the main article's closing.