An always-on research desk for paid-newsletter operators: living beat wikis plus a daily citation-pinned delta brief, deduplicated against everything you have already published — turning 20 hours of weekly research into 2 without ever pitching a story you already covered.
Category: LLM wiki / auto-research (living documents + delta alerts, à la Karpathy)
An always-on research desk for paid-newsletter operators: living beat wikis plus a daily citation-pinned delta brief, deduplicated against everything you have already published — turning 20 hours of weekly research into 2 without ever pitching a story you already covered.
Solo and small-team paid-newsletter operators on Substack/Beehiiv/Ghost in business, finance, policy, and tech — from the 45+ newsletters above $1M ARR down to $5-50K/mo operators. The writer pays personally because research hours are their core cost of goods and their byline rides on every sourced claim.
Onboarding ingests the writer's full archive to build an 'already-covered' memory. The operator defines beats ('EU AI regulation', 'fintech infra M&A'); the agent maintains a living beat wiki — entities, narrative timelines, open questions — updated overnight via batch from 50-200 sources per beat (RSS, primary feeds, filings, earnings transcripts, niche forums, X lists) with diff-based ingestion. Flash/Haiku triage scores each item for novelty against the archive AND all prior briefs, with slop filtering and source-quality scoring. Each morning a delta brief lands: what changed, why this publication's readers care, three suggested angles, pull-quotes with pinned quote + URL + timestamp, and callbacks ('you covered this March 3 — the new development is X'). One-click pulls any wiki section with sources into a draft; the archive doubles as a fact-checked back-catalog and corrections log.
Daily email/Slack delta brief (the operator's native medium) + web workspace for beat wikis and draft pulls, with one-click export to Notion/Google Docs/Substack drafts. Overnight latency is irrelevant — exactly what makes batch economics work.
Substack's paid boom concentrated in exactly the business/finance/politics/tech verticals where research is heaviest; ChatGPT Pulse validated the morning-brief format but is $200/mo, mobile-only, lifestyle-skewed, with no export or citation workflow; 'AI slop' as 2025 Word of the Year means provenance-rich sourcing is a differentiator writers pass through to paying readers; picks-and-shovels sells to curators instead of competing with them.
Feedly's real monitoring tier ($14.4-19.2K/yr enterprise MI), Particle ($2.99 consumer catch-up), generic ChatGPT/Perplexity (one-shot, no beat memory, 37% citation misattribution — career-ending under your own byline), human VAs at $1-3K/mo.
- Feedly Market Intelligence — $1,600/mo Standard, $2,400/mo Advanced, billed annually ($19.2-28.8K/yr); the enterprise monitoring price ceiling
- Feedly Pro/Pro+ — ~$7-12/mo prosumer RSS monitoring; the low anchor for individual writers
- ChatGPT Pro with Pulse — $200/mo; 5-10 generic daily morning briefs, mobile-only, no export/citation workflow; rollout to $20/mo Plus planned
- Perplexity — Pro $20/mo, Enterprise $34/seat/mo; cited one-shot research, no beat memory or archive dedup
- Brand24 — from $79/mo; Mention Solo $41/mo; Awario $24/mo; Prowly $119/mo — mid-market media monitoring ladder
- Meltwater / Brandwatch / Talkwalker — custom quotes, typically $1,000+/mo enterprise media intelligence
- Human VA/researcher — $1-3K/mo, the manual alternative Stringer displaces
- Substack platform economics (payer health) — 8.4M paid subs Q1 2026 (+68% YoY), $45M platform revenue 2025 via 10% take, ~45-50 newsletters >$1M ARR
Ingestion: own RSS/Atom + sitemap diff poller (near-free) with content-hash dedup; Firecrawl or Zyte for JS-heavy/anti-bot pages; free primary feeds (SEC EDGAR full-text, Federal Register, EUR-Lex, GovInfo, court RSS); entry-tier transcript API (Finnhub/API Ninjas ~$50-150/mo shared) — skip X or pass through the customer's own API key; YouTube captions API. Archive memory: ingest public archive RSS/exports, chunk + embed (Voyage-3 or text-embedding-3-large), pgvector on Postgres (Neon/Supabase); novelty = cosine-sim against archive + all prior brief items with an LLM tiebreak. Triage: Gemini Flash or Claude Haiku 4.5 via Batch API (50% discount) with structured outputs for novelty/slop/source-quality scoring. Synthesis: Claude Sonnet 4.6 (or GPT-5-class) with prompt caching for the morning brief + wiki diffs. Citation grounding: verbatim-quote extraction with post-hoc exact string-match verification against fetched HTML — reject any quote that fails match; store quote + URL + timestamp + content hash (this gets precision to ~99%; recall remains unguaranteed and should never be promised). Orchestration: Inngest or Temporal nightly DAGs per customer timezone; SQS-grade queue fine for v1. Surface: Next.js workspace, Postmark/Resend email, Slack app; Notion/Google Docs export APIs; Substack draft export via browser automation (fragile, no official API). Unit economics per $99 pro user (15 beats, ~1,500 nightly source checks): crawl/proxy $10-40, batch triage $5-10, synthesis w/ caching $8-25, embeddings/storage $1-3, email/infra $2 = $25-80/mo COGS → 20-75% gross margin on commodity-source coverage; adding real X/transcript/paywall coverage pushes COGS past $100/user and breaks the model. Viable as a 2-3 person niche business at $49-99/mo targeting $1-3M ARR; not venture-defensible.