Methodology

How we collect data

We track AI startups using four data streams, each transparent in our database:

1. Editor curation. Every startup is reviewed by a human editor before publishing. Submissions land in a moderation queue. We reject companies that aren't building AI, are zombie domains, or have no demonstrable product.
2. Public filings. Funding rounds are extracted from SEC Form D filings, Crunchbase News, the company's own press releases, and the lead investor's portfolio listings. Every round has a source_url pointing back to the original.
3. Source-linked AI extraction. Profile fields (tagline, description, pricing tiers, agent-readiness) are populated by Anthropic Claude Opus 4.6 reading the company's own website, pricing page, docs, and changelog. The AI is instructed to attribute every contestable claim and never to fabricate.
4. Reader signals. Verified founders can claim their profile and submit edits. Edits go through a human-review audit log before publishing.

How we score companies

The NeuronScore (0-100) is a composite of seven weighted signals:

Score weights are reviewed quarterly. The Agent Readiness score is calculated separately and tracks MCP-server availability, public API, SDKs, webhooks, auth, and docs (max 100).

How we track funding rounds

• Reported rounds: extracted from press releases, SEC Form D, and Crunchbase News via our weekly auto-discovery pipeline (Apify + Firecrawl + Opus 4.6).
• Lead vs follow: where the article specifies the lead investor, we mark it; otherwise we leave it unattributed rather than guess.
• Currency: all amounts are stored in USD. Conversion uses the announcement-date FX rate.
• Round-type normalization: "Series A", "Series-A", "series a" all collapse to series-a.
• Dedupe: we never insert a round if (startup_id, round_type, announced_date) already exists.

How we generate AI-written copy

News summaries, alternatives pages, comparisons, and intelligence reports are written by frontier large language models (Anthropic Claude, Google Gemini Flash, and OpenAI GPT models on rotation depending on the task) under a tightly scoped system prompt that:

• Bans hype words ("revolutionary", "game-changer", "disruptive", "leverage" as a verb, etc.).
• Requires attribution for every contestable claim ("the company said", "according to the filing").
• Refuses to fabricate numbers, names, dates, or quotes — if a figure is missing, we omit it.
• Caps articles at 380 words with 8-12 short paragraphs for mobile readability.

Drafts are reviewed by a human editor before publishing for any article that mentions specific funding figures, individuals, or competitive claims.

Errors and corrections

Profile errors can be reported via the "Suggest edit" button on any profile page or by emailing [email protected]. We respond within 7 business days. Persistently inaccurate profiles are temporarily unpublished pending review.

If a published article contains a factual error, we issue a corrigendum at the bottom of the article rather than silently editing — readers should be able to see what changed.

What we don't do

• We don't scrape gated content (LinkedIn beyond public profiles, paywalled publications).
• We don't track individual employees beyond named founders + executives with public profiles.
• We don't sell user data.
• We don't accept paid placement on the NeuronScore, agent readiness, or rankings. Sponsored deals are clearly labelled DEAL.
• We don't use AI-generated images for company profiles. Logos are pulled from official sources (Clearbit, the company's own website) or supplied by the founder during a profile claim.

Independence

NeuronFeed is reader-supported via affiliate commissions on deals (clearly marked) and a forthcoming Pro subscription. We do not take payment from VCs, startups, or any party in exchange for inclusion in rankings, the NeuronScore, the news roll, or the front page. Editorial decisions are made by the editor.