Across 2,583 AI industry events in Scoopkit's archive, the coverage traces back to 107 distinct source domains, and the distribution is far more lopsided than the raw count suggests. The single largest source carries 37.2% of all events on its own. The top 3 carry 54.0%, the top 10 carry 81.3%, and 71 of the 107 domains contributed exactly one event each. If you are building an AI news pipeline and wondering how many feeds you actually need, this is the shape of the answer: a small handful do most of the work, and the long tail is very long and very thin.
Pulled 2026-09-05 from the live production API. Sample: all 2,583 events in the archive, every category, reported between 2025-10-21 and 2026-09-05. Method: source domain parsed from each event's source_url, counted once per event.
Concentration
| Sources | Events covered | Cumulative share |
|---|---|---|
| Top 1 | 962 | 37.2% |
| Top 3 | 1,396 | 54.0% |
| Top 5 | 1,682 | 65.1% |
| Top 10 | 2,100 | 81.3% |
| Top 20 | 2,431 | 94.1% |
| Top 50 | 2,526 | 97.8% |
| All 107 | 2,583 | 100% |
The 10 domains doing most of the work
| Domain | Events | Share |
|---|---|---|
| techmeme.com | 962 | 37.2% |
| arxiv.org | 262 | 10.1% |
| techcrunch.com | 172 | 6.7% |
| prnewswire.com | 164 | 6.3% |
| aws.amazon.com | 122 | 4.7% |
| semafor.com | 93 | 3.6% |
| simonwillison.net | 91 | 3.5% |
| the-decoder.com | 90 | 3.5% |
| marktechpost.com | 81 | 3.1% |
| openai.com | 63 | 2.4% |
The top entry is an aggregator, and that is worth being direct about rather than quietly presenting Techmeme as a primary source. Techmeme is an edited link digest, so a large share of Scoopkit's discovery runs through somebody else's editorial judgment about what counts as industry news. The upside is real coverage breadth from one well-maintained feed. The cost is a dependency, and it is the main reason first-party lab sources get polled separately rather than waited for.
Two entries surprise people. simonwillison.net, one person's blog, produced 91 events, more than The Verge or Wired in the same window. prnewswire.com at 164 is a wire service, which is exactly the kind of source that generates a lot of near-duplicate coverage and makes deduplication load-bearing rather than decorative.
First-party lab and vendor domains, counted together (OpenAI, Anthropic, Google, DeepMind, Apple, Nvidia, Meta, Hugging Face, AWS, OpenRouter), account for 370 events, or 14.3% of the total. That is the fraction of AI industry news that comes straight from the company that made it happen. The other 85.7% arrives filtered through somebody else.
Why the long tail still matters
71 domains produced exactly one event each, and 78 produced three or fewer. Together the entire tail below the top 20 accounts for 5.9% of events.
That number invites a wrong conclusion, which is to cut the tail. The tail is where the events nobody else has come from. A single event from a niche domain is often the one that fires a webhook a customer actually cares about, and dropping sources on volume alone systematically removes the coverage that is hardest to get anywhere else. The right read is that the tail is cheap to keep, not that it is safe to cut.
What this means if you are building this yourself
Scoopkit's own configured inventory is 35 sources: 34 RSS feeds and 1 scraper, with 27 currently active, 7 paused, and 1 on trial. Those 35 configured inputs produce events across 107 domains, because aggregator and community feeds link outward to publishers that were never configured directly.
The gap between "35 things to poll" and "107 domains represented" is the useful part. You do not need to find and maintain 107 feeds. You need roughly 30, chosen so their outbound links cover the rest. The single scraper in that inventory exists because Anthropic publishes no RSS feed at all, which is a whole problem of its own.
The maintenance load is the part that gets underestimated. 7 of 35 sources sit paused right now, for a mix of reasons: feed quality, volume that overwhelmed the useful signal, and access terms. That is a 20% pause rate on a curated list, and it moves month to month. A feed list is not a thing you assemble once.
If you would rather query the result than maintain the inputs, the free tier covers 100 requests a day against all of it.
FAQ
What counts as one source here?
The registered domain of the event's source_url, with the www. prefix stripped. Subdomains are kept separate, so aws.amazon.com and blogs.nvidia.com are counted on their own rather than folded into their parent domains.
Does the top source getting 37% mean the data is dependent on Techmeme?
For discovery breadth, partly yes. For the events that matter most, less so, because first-party lab sources are polled directly on their own schedule and do not wait on aggregator pickup. A model release announced on anthropic.com enters through the scraper on that domain, not through a link digest.
Why is arXiv second when its feeds are paused?
The arXiv cs.AI, cs.CL, and cs.LG feeds were active earlier in the window and are paused now, so the 262 events are real history from when they ran. They were paused because raw paper volume swamped the ratio of genuinely notable research to routine preprints. That is a live example of the 20% pause rate above.
Can I see which source a given event came from?
Yes, source_url is a free-tier field on every event, along with article_count. You do not need a paid key to trace an event back to where it came from.