Methodology

Bellwether collects what developers post, publish, install and hire for. Conversation comes from Hacker News (every feed), Lobsters, developer subreddits, Bluesky and Mastodon. Adoption comes from the package registries (npm, PyPI, Packagist, Docker Hub) and GitHub. Hiring comes from job boards. Each item is stored once per source, deduplicated on its URL, and tagged with the entities it mentions.

Every count you see is a count of distinct stories. The same Hacker News story appears in several feeds; counting rows instead of stories would inflate everything by two to three times, so no Bellwether figure does that.

Mindshare is reported as a share, never a raw count: the percentage of that month's tagged stories that mention an entity. Volume drifts as sources come and go; share does not. The current month is always partial and every issue says so.

A product and its company are one conversation. "Claude" alone understates Anthropic, so family level figures group a vendor's products with the vendor.

Downloads, stars, forks and pulls are snapshots taken daily. They measure installs, not opinion, which is why an issue can show conversation falling while adoption rises. That divergence is the finding Bellwether exists to report, and it is only visible when both lines are measured separately.

Adoption history is still short, so an issue describes it as a snapshot until there are enough weeks to call it a trend.

No number appears without its query and result. Reddit engagement stays out of headline claims for licensing reasons. Salary samples are too small to quote. If a period is quiet, the issue says it was quiet rather than manufacturing a trend.

Each issue ends with dated, falsifiable calls, and a later issue grades them. The full record is on the bets page.