The research notes
The notes behind The research behind it: the report and the three sets of notes it drew on, and the ownership audit of October 2026, published as they were written. Researched with Claude, an AI model made by Anthropic, at the request of Vinay Kambipura Sridhar, who directed the questions and reviewed the findings.
They are working notes. They were written for Huduku’s owner, in the project’s own terms (rounds, rules, file names); they quote their sources more than the research page does, always with a link; and they were written before the citation check of 5 October 2026, which corrected some claims. Where they differ, the research page is the checked version. What the check changed is listed below, and on each note.
- The report “Owner turns beat what audits found elsewhere”
What the three sets of notes add up to: audits elsewhere, three ways a list can lean, Huduku’s rules set against the evidence, what to do next, and India. Round 29 was built from it.
- Notes: audits of aggregators “Aggregator audits and practice: source concentration and publisher flooding”
Published audits of Google News, Google’s Top Stories, Apple News and others: who took the room, how it was measured, and what aggregators have said about freshness and flooding.
- Notes: fair ranking “Provider-side fairness and source concentration in ranking and recommender systems, as applied to news aggregation”
Popularity and exposure bias, fairness for providers, equal against merit-proportional exposure, grouping by owner, and gaming by volume.
- Notes: norms and regulation “Normative frameworks, standards and regulation for news diversity and algorithmic curation (with India)”
What “diverse” exposure means, how ownership concentration is measured, binding and soft law on curation, India’s rules, and professional guidance.
- The ownership audit “Ownership audit, October 2026: findings”
Who controls the outlets with no group on record: the audit that put Kannada Prabha with Asianet Suvarna News, and what it couldn’t check.
What the citation check changed
On 5 October 2026 three Claude agents, each given a third of the research page’s 53 sources and the sentences citing them, opened every link and judged each claim, and the research page was corrected. The notes are left as written; these are the differences.
- Diakopoulos and colleagues, “I Vote For”: The data ran from 31 May to 8 July 2016, mostly after the primaries had ended, not through them. In: The report, Notes: audits of aggregators.
- Trielli and Diakopoulos: The 83.5% is a share of impressions, not of articles. In: The report, Notes: audits of aggregators.
- Hernandes and Corsi: The unweighted top-ten share runs from 21% to 36%, not 22% to 36%. The 17.2 hours is the average from sixth place down; places two to five averaged 14.0 hours. In: The report.
- Nechushtai and Lewis: The 69% is an average of the top five organisations in each search, not a share of all recommendations. In: The report, Notes: audits of aggregators.
- Evans, Jackson and Murphy: A similar study in the UK, not a replication. In: The report.
- Robertson and colleagues, 2023: The Springer link in the notes goes to a login page; the paper is open at nature.com. In: The report, Notes: audits of aggregators.
- Fischer, Jaidka and Lelkes: National outlets dominated results unless people searched specifically for topics of local interest; the notes say local outlets ranked lower even for local topics. In: The report, Notes: audits of aggregators.
- Magnusson: Correcting the inequality measure lowered the measure by 56% on average; the report says it cut the effect. In: The report.
- Search Engine Roundtable: Google’s site-diversity change was reported to cover only the main web listings, not other search features; the article doesn’t mention Top Stories. In: The report, Notes: fair ranking.
- Search Engine Journal: A site’s subdomains are generally counted as the site, not always. In: The report, Notes: audits of aggregators, Notes: fair ranking.
- Ground News; SmartNews: Ground News’s left-to-right ratings come from third parties; SmartNews’s are its editors’. In: The report.
- Abdollahpouri and colleagues: The study is of ratings, not clicks, and describes no feedback loop: items that already have more ratings get recommended more. In: The report.
- Singh and Joachims: They offer exposure in proportion to merit as an alternative to equal exposure. In: The report, Notes: fair ranking.
- Morik and colleagues: The 2% relevance gap and the click feedback are separate results. FairCo adds each group’s accumulated shortfall in exposure, measured against its merit. In: The report.
- Biega and colleagues: The target is attention in proportion to relevance, summed over many rankings, not attention evened out. In: Notes: fair ranking.
- Manku and colleagues: The paper is about detecting near-duplicates, not clustering them. In: The report.
- Zehlike and colleagues, FA*IR: The anthology link is dead; the paper is on arXiv. FA*IR keeps a minimum share for a protected group at every point in the list. In: The report, Notes: fair ranking.
- Gómez, Boratto and colleagues: Fairness research groups providers by traits such as gender, age or where they come from. In: The report.
- CMPF: It says the data needed to measure concentration are often missing or unreliable; it doesn’t say the figures rest on estimates. In: The report.
- TRAI, 2014; RSF: The 1,000-point limit applies where both the TV and the newspaper markets are concentrated, across both. That the recommendation was “never enacted” couldn’t be confirmed; what is sourced is that Indian law still sets no such thresholds. In: The report, Notes: norms and regulation.
- Medienstaatsvertrag, § 60: The gesetze-bayern.de link failed (503); the official text is the media authorities’ copy. The law presumes dominant power over opinion at a 30% audience share; it isn’t a cap. In: The report, Notes: norms and regulation.
- Ofcom, legal framework: The page refused every fetch (403), so the UK’s 20% couldn’t be checked and was left out of the research page. In: The report, Notes: norms and regulation.
- Council of Europe, 2021: The five criteria are for deciding what counts as public-interest content to give prominence to, not for ranking in general. The site refused the fetch; the wording was read in two published copies. In: The report.
- Council of Europe, 2018: The 5% disclosure threshold couldn’t be checked (403) and was left out. In: The report, Notes: norms and regulation.
- India’s competition regulator (TechCrunch): The CCI said it appeared that publishers had no choice, in ordering an investigation; it was a first view, not a finding. In: The report, Notes: audits of aggregators.
- IFF; The Tribune: IFF reports the Bombay High Court’s stay; the Madras High Court’s is in The Tribune. In: The report.
- IFEX; SFLC.in: IFEX refused the fetch; SFLC.in publishes the same statement. In: The report, Notes: norms and regulation.
- European Media Freedom Act; DSA Observatory: The Act has no rules on recommender systems. Its Article 18 covers how very large platforms treat media content; explaining ranking is the Digital Services Act’s Article 27. In: The report, Notes: norms and regulation.