Skip to content
Issue 1,847 · Est. 2017

METHODOLOGY · LAST REVISED MARCH 2025

How a changelog becomes a briefing.

BestUpdate reads 14,200 software products so you don't have to. Here's the engineering — five pipeline stages, a 99.2% field-extraction benchmark, and one editor at the end of the chain deciding what makes the cut.

Get the Daily Drop Read the pipeline ↓

Cited by TechCrunch, The Verge & Wired · 38,400+ subscribers · bootstrapped in San Francisco since 2017.

THE PARSING PIPELINE

Five stages. One editor. Zero guesswork.

Every release note that lands in the Daily Drop has been pulled, parsed, tagged, scored, and read by a human. The pipeline below is the same one we audited against four competing trackers in 2025.

  1. 01

    DISCOVERY

    Finding the feeds nobody indexes.

    We seed our product graph from public registries, GitHub org membership, and a corpus of 3.4M historical entries. When a vendor publishes a new feed on a custom domain, our crawler finds it within hours — not the weeks it takes RSS aggregators to notice.

    • Sources: RSS, Atom, GitHub Releases, custom JSON, vendor blogs
    • Coverage: 14,200 products across 312 categories
    • Crawl cadence: hourly for active vendors; daily for the long tail
  2. 02

    INGESTION

    Pulling the field — every field.

    A versioned parser strips each feed into structured fields: version, release date, channel, deprecations, migration steps. The parser is the part we benchmarked. Across RSS, Atom, GitHub Releases, and custom JSON feeds it hits 99.2% field-extraction accuracy — measured against a hand-labelled gold set of 4,800 entries.

    • Median detection: 38 minutes from publish to index
    • Field accuracy: 99.2% vs. four competing trackers
    • Failure modes: logged, retried, and human-repaired within 4h
  3. 03

    SEMANTIC TAGGING

    Knowing what a release actually means.

    A version string isn't a story. Our tagger applies semantic-version diff (major / minor / patch), infers change category from the prose, and cross-references each entry against the vendor's prior 90 days of releases. That lets us say "this is a breaking API change" instead of "this is a 1.4.0".

    • Tag taxonomy: 31 change categories, 9 severity bands
    • Vector model: fine-tuned on 3.4M historical entries
    • Review pass: every major release re-read by an editor
  4. 04

    EDITORIAL SELECTION

    A human decides what 38,400 readers get.

    The pipeline surfaces roughly 4% of indexed releases for editorial review. A rotating on-call editor — drawn from our 22-person team distributed across 9 countries — reads every flagged entry and writes the 6–10 paragraphs that ship each morning at 7am PT. No automated summaries. No LLM rewrites. A byline, every time.

    • Editorial window: 04:00–06:30 PT, Mon–Fri
    • Coverage: 24/5 on-call for breaking releases
    • No sponsored placements · no pay-to-publish · no VC

Want the longer version? The full methodology doc — including the gold-set construction and the per-tracker benchmark scores — is published at /method/.

THE 2025 BENCHMARK

The numbers we'd put on the record.

In Q1 2025 an independent research firm measured BestUpdate against four competing changelog trackers on a hand-labelled gold set of 4,800 entries. We won on accuracy, latency, and tag precision. These are the figures.

99.2%

Field-extraction accuracy

Measured across RSS, Atom, GitHub Releases and custom JSON feeds against a 4,800-entry gold set. Highest of the five trackers tested.

14,200+

Products tracked

Across 312 categories — the largest dedicated changelog index on the public web. Hourly release detection for active vendors.

3.4M

Indexed changelog entries

Every entry since 2017 carries semantic-version diff, category tag, and migration-impact score. Full corpus is searchable.

38min

Median detection latency

Average time from a vendor publishing a release to it appearing in our index. Not instant — we don't promise that — but close.

Independent benchmark conducted by Cedar & Loom Research, January 2025. Full report available on request.

Get the Daily Drop

§ The Editorial Filter

Of the 14,200 products we watch, only 4% make it into your inbox.

Every release that crosses our desk — roughly 380 per day — is scored against four filters before it earns a line in the Daily Drop. This is the human layer that separates BestUpdate from the firehose aggregators, and the reason subscribers report a 67% reduction in time spent triaging competitor releases within their first 30 days.

  1. 01

    Relevance to working PMs and eng leaders

    Does the release change how a team builds, ships, prices, or positions a product? Routine bug-fix bumps and micro-patch increments are filed to the index but never promoted to the briefing. A desk editor reads every line of the release note before promoting it.

  2. 02

    Likelihood of a breaking change

    Major-version bumps, deprecation notices, API contract shifts, and migration-impact tagged entries are flagged for the editor's second pass. Semantic versioning is read literally — a major bump is treated as a breaking change until proven otherwise.

  3. 03

    Competitive signal across the category

    Does this release shift the landscape for adjacent products? A pricing change at one major IDE ripples into four other categories. Our editors maintain category maps across all 312 tracked verticals and flag cross-category consequences the source changelog never mentions.

  4. 04

    Editorial voice and concrete detail

    Would the entry hold up in a standup meeting? Vague lines like "various improvements" are sent back for a second read of the source release. A line earns its slot only when a PM could act on it before lunch.

"BestUpdate cut our weekly competitor triage from a half-day meeting to about 40 minutes. We trust the parse, we trust the curation, and we trust that nothing in the briefing is filler. It is the only subscription our PM team refused to cut in last year's budget review."

Maya Okafor Head of Product, verified subscriber at a Series B infrastructure company 67% reduction in triage time, measured across a 2024 cohort of 1,240 subscribers.