Methodology

Last updated: September 2026

AI Stream Online runs on automated pipelines and openly available data, and automated checks watch the result: after every change and once a day the live pages are loaded in a browser to catch errors, empty sections and stale data, and the search-engine health of the site is checked every week. This page explains how every number and story on the site is produced, where it comes from, and — just as important — what it does not capture.

1. News aggregation

Articles are pulled from public RSS/Atom feeds on an automated schedule that runs every 3 hours. The current sources are the AI sections of TechCrunch, Ars Technica, MIT Technology Review and The Verge; the AI-focused outlets AI News, The Decoder, AI Business and IEEE Spectrum; the official blogs of OpenAI, Google Research, Google DeepMind, Hugging Face, NVIDIA, Meta's engineering team and MIT News; and Rest of World and the South China Morning Post for coverage from outside the usual US/EU sources. Feeds are read on a schedule with an identified user agent, and when a story has no image in its feed we check the site's robots.txt before opening the article page for its preview image. Each run merges newly fetched articles into a rolling archive (capped at 100 current items), deduplicated by article URL. We link to and credit the original source for every story; we never republish full article text. The one-sentence summary under each headline is written by Anthropic's Claude model from the headline and the feed's short excerpt, in its own words and using only facts stated there; it can still contain mistakes, so the original article is always one click away. Preview images load from the publisher's own site, and images whose address credits a photo agency (such as Getty Images, AP or Reuters) are left out.

Feeds that are not AI-only — the research blogs and the general-interest outlets — pass through a keyword relevance check first, so a chip earnings report or a phone review doesn't land here as AI news.

2. Story clustering

When two articles from different feeds share at least one tagged company and have substantially similar titles on the same day, they're merged into a single story with an "Also covered by" credit to the other outlet(s), instead of appearing as unrelated duplicates. Company overlap alone isn't enough on its own to trigger a merge, so two unrelated stories that happen to mention the same company aren't grouped together.

3. Per-article intelligence tagging

Every article is automatically tagged with an impact level, an automated sentiment score, and mentioned companies, countries, and technologies (labeled "Related companies & topics" on each card — they're extracted name matches, not an analysis of who is actually affected). This is done with a keyword/pattern-based heuristic engine, not a large language model. Country tagging only fires when an article's text explicitly names a country, so per-country news coverage is sparser than per-company coverage — a story about a company isn't necessarily about the country it's headquartered in.

Impact level: High and Medium are assigned only when specific keyword signals are found. When neither matches, the article shows no impact badge at all — a "Not assessed" state, not an assessed "Low."

Automated sentiment (Positive/Negative/Neutral) is a keyword count, not a factual verdict or a prediction of real-world impact — it says the engine spotted more positive- or negative-coded words in the text, nothing more.

4. AI companies tracked by country

Company counts shown per country come from Wikidata's public SPARQL endpoint — a free, live, structured, community-edited database. Companies are identified by their AI-industry tag and grouped by headquarters country.

Limitation: Wikidata is edited by volunteers, so coverage is uneven and skews toward companies that English-speaking contributors happened to document. It is not a census. We label this metric "AI companies tracked on Wikidata," never "total AI companies," for exactly this reason. Refreshed automatically every 3 hours.

A country gets its own page (and each of its companies a page) once it has at least 3 tracked companies, which filters out a small amount of data-quality noise below that floor. Entities with no usable name on Wikidata are also excluded from every count and list.

Reusing this data: the facts come from Wikidata, which releases them under CC0 (no rights reserved). Our compilation (the selection, grouping and counts) is published for reading and reuse at ai_companies.json; if you reuse it, please credit AI Stream Online and Wikidata, and keep your requests gentle (it refreshes every 3 hours, so once a day is plenty). Every company and country page also carries the same facts as schema.org structured data (Organization and ItemList), and /companies describes the whole file as a Dataset.

5. Explore AI by Country

This is an unranked, unordered-by-us grid of every country with at least 3 tracked companies (the number changes as Wikidata does), sorted only by that real Wikidata count — never described as a measure of "AI leadership." Each entry links to that country's full page (recent tagged stories, the same company list, always live).

The shaded world map above the list is the same dataset, not a second metric: each tracked country is placed into one of six real, data-derived count ranges and colored from dark brown to bright gold accordingly, with a legend showing which color maps to which range. It's a visual index into the exact numbers already in the list below — hovering or clicking a country shows its precise count. The base map itself is a public-domain map from Wikimedia Commons, simplified for file size; country shapes are for visual reference only and carry no data of their own beyond the coloring we apply.

6. GitHub activity (open-source AI sample)

The GitHub numbers (stars, open issues) are fetched from GitHub's official public API for a fixed list of widely-used open-source AI projects — a sample of open-source AI activity, not a measure of AI activity overall, and we say so next to the numbers. We show the last successfully fetched value and how long ago it was fetched, rather than a continuously ticking extrapolated estimate between refreshes. If a fetch fails, the whole widget hides rather than show a stale or zeroed-out number as current.

7. Videos and podcasts

Featured videos are selected via the YouTube Data API across several distinct AI sub-topics, keeping only English-language videos (by the uploader's declared language, the script of the title and description, and common non-English markers), taking the top result from each by view count so one viral topic can't crowd out the rest. Podcast episodes are pulled directly from each show's own public RSS feed.

8. Reader poll ("How Will This Affect People?")

Each story shows two independent things, kept visually separate: an automated sentiment read (see section 3) and a reader poll asking "How will this affect people?" with Positive / Negative / Not Sure as answers.

The poll is backed by a real shared counter, with server-side validation. When that backend isn't attached, the widget shows "Your reaction: [your answer]" instead of a fabricated or purely-local "community" figure — your own answer is never presented as if it represents other readers. Honest limitation: this is a lightweight reaction poll, not a scientifically controlled survey, and it reflects only readers who chose to answer.

Reader Mood. Above the poll, Reader Mood adds up every answer from the last 7 days: the share of Positive, Not sure and Negative, and a net score (the Positive share minus the Negative share, from −100 to +100). It only shows percentages once at least 20 votes have been cast that week, and it highlights the most divisive and most hopeful stories only among stories with at least 3 votes. Like the poll itself, it reflects readers who chose to answer, not the public at large.

The index measures how much attention a subject is getting right now compared with its own normal — not how big it is. Every signal is converted into momentum against that same subject's own history, which puts large and small players on one comparable scale.

Signals. Up to three, all from free, official sources: news — Hacker News stories mentioning the subject; Wikipedia views of its company and product pages; and developer downloads of its SDKs from PyPI and npm. Each registry is scored separately and the two are averaged, so an outage at one registry doesn't void the signal.

Formula. For each signal, a recent value is compared with a baseline:

  • Weekly: the average of the last 7 days against the average of the 28 days before them.
  • Daily: the latest complete day against the average of the same weekday in each of the four prior weeks — downloads and pageviews dip every weekend, so a plain average would read every Sunday as a collapse.

Each signal becomes a ratio of its recent activity to its baseline, taken on a log scale so that doubling and halving count as equal and opposite, and limited in both directions so a single freak day can't dominate. A subject's combined momentum is the plain average of its qualifying signals, mapped onto a 0–100 scale. 50 means exactly normal for that subject; above means more attention than usual, below means less.

When a signal is left out. A signal only counts when its baseline clears a minimum daily volume and enough days of data are present; otherwise it is excluded, never counted as zero. A subject needs at least two qualifying signals to be scored, which is why a company can appear in the weekly view but not the daily one.

Weighting. Signals are weighted equally.

Limitations: Wikipedia views are a proxy for public search interest. Hacker News skews toward a technical, English-speaking audience. Download counts include automated installs from build systems, so they track developer activity rather than individual people. The index measures attention, not quality, sentiment or market value.

Not financial advice. Nothing in the index — or anywhere on this site — is investment, financial, or trading advice, and a rising or falling score is not a signal to buy or sell anything.

Live Pulse. Under the index, a live panel counts mentions of the same companies in the last hour, as they happen: Hacker News stories and comments (checked every minute while the section is on screen) and, only after a visitor presses Go Live, Bluesky's public post stream for two minutes. It matches company and product names such as "Claude" or "Gemini", so an occasional unrelated use of a word is counted too. Only counts are shown, never the text of posts or comments, and the pulse does not feed into the daily index score.

10. AI on YouTube

Once a day, for each of about a dozen AI topics, we ask the official YouTube Data API for the most-viewed videos published in the last 7 days that match the topic's search terms, keep the English-language ones whose title, description or tags really mention the topic, and show the top one (preferring videos longer than a minute) with the view count YouTube reports for it. Topics are listed in a fixed order, not ranked.

No scores or trends. YouTube's API policies don't allow derived metrics without their approval, so we show their numbers as they report them, without computing scores or keeping a history.

What it can't tell you. YouTube search results depend on the search terms and on YouTube's own ranking, and the most-watched video isn't necessarily the best or most accurate one. View counts show attention, not quality or importance, and nothing here is financial advice.

11. Daily AI Brief

Three times a day — at 06:00, 13:00 and 22:00 Italy time, each covering the hours since the previous edition, so together they cover the whole day — the AI stories from our news archive — headline, source and a short excerpt, nothing about readers — are sent to Anthropic's Claude model, which writes a short summary of the 3 to 5 most important stories. It is instructed to use only facts stated in those stories, to write in its own words, and to avoid predictions and financial advice. The model only picks which stories to cover: every original headline, source name and link on the page comes from our own archive, so a link can't be invented. AI summaries can still contain mistakes, which is why every item links to its original sources.

12. AI Money Moves

The dashboard's AI Money Moves strip lists funding rounds, valuations, IPO plans, acquisitions and revenue milestones named in the past 7 days of our news feed. Anthropic's Claude model reads each headline and summary once and pulls out the company and the figure as reported; every item links to its story, so the figure can be checked at the source.

Live stock prices come from TradingView's widget, which is licensed for it. We don't copy quotes from finance sites.

13. AI Radar

The AI Radar page scans four open sources once a day through their official data feeds, never by copying their web pages, and identifies itself with a named User-Agent and a contact address. Research: the newest AI papers on arXiv (titles, authors, categories and links; arXiv metadata is released as CC0). Models: notable AI models from Epoch AI (licensed CC BY; credit Epoch AI and its authors). Policy: US federal documents whose title mentions AI, from the Federal Register (US government work, public domain). Attention: Wikipedia page views of AI topics over the last 7 days, from Wikimedia Analytics (CC0); the bars use a square-root scale so smaller topics stay visible next to the biggest one.

We show titles, names, numbers and links, and never the papers themselves or their abstracts. Every item links to its original source. Honest limitation: the Federal Register list only includes documents with AI in the title, so it misses documents that discuss AI inside a longer text; the model list depends on what Epoch AI has added so far, and its newest entries often have no training-compute figure yet; Wikipedia views measure curiosity, not usage.

14. AI Signals

AI Signals shows how this week compares with the usual, using only data this site already holds. It adds no new source and no AI. Attention: the change in Wikipedia page views of AI topics, taken from the AI Trends Index. Developers: the change in downloads of AI packages, also from the AI Trends Index. New models: how many notable models Epoch AI lists for the last 30 days, against this year’s average month. Deals in the news: how many stories in our archive are about funding rounds, acquisitions, IPOs or valuations this week, against the weekly average of the four weeks before. News tone: the average sentiment of the AI stories in our archive (the sentiment is written by AI), this week against the four weeks before; it reads “steady” unless the average moves clearly.

Every row is a change against a normal, drawn from a centre line. It is not a score out of 100, and up or down is not good or bad. Honest limitations: a row is left out until its history is long enough, and our archive keeps about twenty stories a day, so the deals and tone rows describe our coverage, not the whole market. The numbers are recalculated with every 3-hour refresh and are not financial advice.

15. Update frequency

News, AI Money Moves, videos, GitHub activity, podcasts, and AI company counts all refresh automatically every 3 hours via a scheduled job. Stock prices in the ticker at the top of the home page come live from TradingView's widget, not from us. The AI Trends Index is recalculated once a day, since its underlying sources publish daily figures. The Top AI Platforms / Top AI Voices lists are editorial and updated periodically by hand, not on a fixed schedule.

16. What we deliberately don't do

We don't fabricate metrics to fill a gap: where no reliable free data source exists, the metric is left out.

17. Questions or corrections

If you spot a data error or have a source suggestion, contact us at contact@aistreamonline.com.