The research behind it

In short

Every published audit of a general news aggregator that this research found shows the same pattern. A few large outlets take a third to over half of the places on the page; the top place is more concentrated than the rest of the list; and fresh copy wins. The audits put this down to ranking by recency and by the size of an outlet, not to personalisation. So a page that is the same for everyone isn’t balanced unless it balances on purpose.

Huduku’s rules count owners rather than feeds, give every owner a turn before any gets a second, cap the cards any owner can have, and rank stories by how many owners carried them. The research supports that approach. It also shows where it was weaker: who goes first within a turn, how accurate the ownership data is, and the pages the rules didn’t reach. This page sets out the findings, how Huduku’s rules compare, and what was changed as a result.

What audits of other aggregators found

Between 31 May and 8 July 2016, as the US presidential primaries ended, CNN and The New York Times held 44.2% of the links in Google’s “In the News” box, and 64.4% of its first, most prominent slot (Diakopoulos and colleagues).

A 2017 audit of Google’s Top Stories found that 20 sources took more than half of all impressions, and the top fifth of 678 sources took 86%. Of the impressions, 83.5% were of articles less than a day old, and the authors suggested that outlets able to keep producing fresh copy may be more likely to be picked (the Tow Center’s summary; the paper).

A 2024 audit of the Google News tab in Brazil, the UK and the US found that 2.1% of outlets supplied half of all results. Counted plainly, the ten biggest outlets held 21–36% of results; weighted by position, about 77–81%. Articles in the top slot were 13.3 hours old on average, against 14.0 hours in places 2 to 5 and 17.2 hours from sixth place down (Hernandes and Corsi). That gap between plain and weighted shares matters most: whoever holds the top places holds most of the attention.

Personalisation isn’t the cause. In a US study, people of opposite political leanings saw the same Google News stories 99.9% of the time, and on average the five most recommended organisations made up 69% of what each search showed (Nechushtai and Lewis, summarised by JournalismAI). A similar UK study found nine London outlets supplying three-quarters of the links, with little sign of personalisation (Evans, Jackson and Murphy). A study of Google Search found that the partisan or unreliable news people saw and clicked on came more from their own choices than from the algorithm (Robertson and colleagues). Concentration is a property of the shared ranking, which is the part an aggregator controls.

Design choices move the numbers. In Apple News, the Trending section, chosen by algorithm, gave its top three sources 45.2% of its stories; the Top Stories section, chosen by editors, gave its top three 23.7% (Bandy and Diakopoulos).

How a measure is built matters too. One study found that national outlets dominated Google News results unless people searched specifically for topics of local interest (Fischer, Jaidka and Lelkes). A reply from a Google researcher argued that its measure of inequality was built wrongly, and that correcting it lowered the measure by 56% on average (Magnusson). So Huduku publishes its method alongside its numbers.

Published industry practice is thin. The one numbered rule against flooding that a major aggregator has disclosed is Google’s of June 2019: usually no more than two results from one site near the top, with a site’s subdomains generally counted as the site (Search Engine Journal). It was reported to apply only to the main web listings, not to other search features (Search Engine Roundtable). Google has said it gives significant original reporting more prominence (Google), and it later asked sites that republish others’ articles to keep those copies out of its search results, partly because a republished page can differ from the original (Search Engine Journal).

Some news apps balance by showing outlets rated from left to right, by third parties (Ground News) or by their own editors (SmartNews, reported by TechCrunch). Huduku’s rules rule that out: it gives no ratings. Ground News also records the owners of more than 2,200 outlets by hand, the nearest thing to Huduku’s owner facts.

Three ways a list can lean

Computer-science research treats this as fairness to the people whose work is ranked, not only to readers (Burke). Readers look at the top of a list far more than lower down. One standard model gives the item in place j a share of attention of 1 ÷ log(1 + j), so small differences in score become large differences in exposure (Singh and Joachims). In a news example, ranking by relevance alone gave a group of articles only 2% less relevant far less exposure; and ranking by clicks feeds on itself, as the items shown most collect the most clicks (Morik and colleagues).

Popularity. Items that already have more ratings get recommended more (Abdollahpouri and colleagues). Huduku uses no clicks or ratings, so nothing of the kind feeds back into its pages. The nearest thing in Huduku is supply: an owner with many titles looks like many voices unless counts go by owner. Top stories rank by how many owners carried a story, which closes that.

Volume. An outlet that publishes all the time fills any list sorted by newest. Turns by owner and a ceiling take away the reward for volume, and grouping headlines into stories stops one agency report from filling several places. Those match the standard defences: detecting near-duplicates (Manku and colleagues) and limits per site.

Freshness. This is the effect the audits measured most directly. Huduku’s weight for a card, its owners ÷ (1 + quiet hours ÷ 8), still favours recent stories, but it multiplies recency by how many owners carried a story, instead of letting the newest headline win.

Huduku’s rules, set against the research

Each rule as it stood after round 23 (October 2026), the nearest research or standard, and what that says about it.

1. Count owners, not feeds.

   Google counts a site’s subdomains as one site (Search Engine Journal). Antitrust’s concentration index adds up the squared shares of firms (US Department of Justice). India’s telecom regulator proposed measuring news concentration per owner in each language market (TRAI, 2014).

   It goes further than any disclosed aggregator rule this research found, which group by web address. It fails silently when the ownership data is wrong.

2. One turn per owner before any gets a second.

   A way of diversifying results that hands out places like seats in an election, here with every owner’s vote equal (Dang and Croft). The “open” kind of diversity, an equal share for every voice (Loecherbach and colleagues), close to the deliberative model’s equal representation (Vrijenhoek and colleagues).

   Every point down the list is balanced, much as the FA*IR method keeps a minimum share for a protected group at every point (Zehlike and colleagues). It takes the equal-exposure side of a live debate (below).

3. Within a turn, the highest weight first.

   Position bias of 1 ÷ log(1 + j) (Singh and Joachims); attention balanced against relevance across many rankings (Biega and colleagues).

   Ties and near-ties keep favouring the same owners; the research says to rotate or even them out. Round 29 changed this (below).

4. At most 2 cards per owner in a section, 3 in the lead section.

   Google’s usual limit of two results per site near the top.

   Close to the one public industry rule.

5. Past 10 cards on the edition, an owner waits for the others.

   A limit that relaxes, as Google’s does when its systems judge a result especially relevant. Legal thresholds for media ownership sit far higher: Germany presumes dominant power over opinion at 30% of the TV audience (Medienstaatsvertrag § 60).

   A design choice for a page, not a legal analogue.

6. A card’s weight: its owners ÷ (1 + quiet hours ÷ 8).

   Ranking by freshness may favour publishers that can keep producing fresh copy (the Tow Center).

   It weighs how widely a story is covered, a sign of merit that isn’t volume, and still favours recent stories. The 8 hours is Huduku’s choice; the research doesn’t give a number.

7. Top stories rank by how many owners carried them; each is shown by a reporter whose owner has the fewest cards.

   Google’s boost for original reporting (Google), and its trouble telling originals from republished copies (Search Engine Journal).

   Showing a story by room rather than by who was first gives up credit to the first reporter for balance. Who was first is hard to know anyway.

8. A check: the effective number of owners, and an alarm past 15% for one owner.

   HHI, top shares, Gini and evenness, as audits report them (Hernandes and Corsi; Bandy and Diakopoulos).

   Comparable with the research, but it counted cards, not where they sat. Round 29 added position.

9. Feeds fetched every hour; a time in the future becomes the time it was fetched.

   Google tells publishers not to give future dates (Google Search Central).

   Moving future times follows the guidance. Feeds reached only every two or three hours lose to hourly ones.

10. No ratings, no boosts, no personal signals.

    The Council of Europe’s guidance on giving prominence to public-interest content: deciding what counts as such content should rest on criteria that are “clear, non-discriminatory, viewpoint neutral, transparent, and objectively justifiable” (Council of Europe, 2021; quoted by Policy Review).

    Written for a different purpose, but the closest description of Huduku’s approach in guidance of this kind.

Equal turns: a choice, not a settled question

Researchers disagree about the goal. Singh and Joachims call equal exposure across groups “demographic parity”, warn that it can cost a lot when groups differ in relevance, and offer exposure in proportion to merit as an alternative (Singh and Joachims). Biega and colleagues also aim for attention in proportion to relevance (Biega and colleagues). Saito and Joachims then show that a rule in proportion to merit can leave some items worse off than a random order would, and see no principled reason for exposure to rise in a straight line with merit (Saito and Joachims). A survey of the field concludes that the right definition depends on the values chosen (Zehlike, Yang and Stoyanovich). Scholars of news diversity add a democratic argument: no single actor should be able to dominate public debate (the Dagstuhl manifesto).

Huduku’s answer: output volume isn’t merit, and what equal turns remove, as the audits show, is an advantage in the capacity to produce copy. Within each turn Huduku still orders by a sign of merit that isn’t volume: how many owners carried a story.

The cost is real, and Huduku says so. When one owner breaks several big stories, its second-best can sit below another owner’s weaker one, and a small owner with one weak item still gets a full turn. Equal turns are also about sources, not views: two owners can say the same thing, and what people actually read is up to them (Loecherbach and colleagues). Because the page is the same for everyone, it can’t split readers into separate agendas (Vrijenhoek and colleagues).

Where the rules leaked

Position. Equal turns fix how many cards each owner gets, not where they sit, and the first card in a run gets more attention than the last (Singh and Joachims). The research’s answers are to balance attention against relevance across many rankings (Biega and colleagues), to add each group’s accumulated shortfall in exposure, measured against its merit, to its items’ scores (Morik and colleagues), or to measure the exposure expected across rankings that vary (Diaz and colleagues).

Ownership data. Grouping by owner is only as good as the ownership table. Fairness research groups providers by traits such as gender, age or where they come from; this research found none that groups them by corporate owner (Gómez, Boratto and colleagues). One missing link gives a group two turns. In India, the Media Ownership Monitor found cross-shareholdings that obscure who owns what, and no standards for measuring concentration (Reporters Without Borders); the Centre for Media Pluralism and Media Freedom notes that the data needed to measure concentration are often missing or unreliable, even in Europe (CMPF). TRAI’s 2014 proposal counted control as 20% of the shares, or control in law or in fact (TRAI, 2014).

Freshness. It came back by three routes. Feeds reached only every two or three hours lose to hourly ones under a weight that fades with time. Search and topic pages listed the newest first, with little or no limit per owner. And outlets whose feeds land in one section compete for that section’s two places. None of these is a fault in a rule, but each undid a rule’s purpose on some pages.

What the research recommended, and what Huduku did

1. Fix the inputs. Kannada Prabha and Asianet Suvarna News are now counted as one owner. A May 2026 rating report lists Kannada Prabha’s company as a subsidiary of Asianet News Network (Acuité), which holds 51% of it (Newslaundry, 2024). Every other public outlet’s owner was checked for a parent company in company records, rating reports and the outlets’ own pages, and no other shared owner was found. Public feeds that aren’t news get an hourly run of their own.

2. Measure what the audits measure. Huduku’s balance check now reports the top three and top five owners’ shares; the owners of the top slots; shares weighted by position, with the 1 ÷ log₂(1 + rank) discount; HHI beside the effective number of owners, with antitrust’s bands of 1,000 and 1,800 as a familiar reference (US Department of Justice); and evenness. It runs every day. This research found no standard threshold for a news page, and the authors of the RADio measures say theirs are “not to serve as thresholds” (Vrijenhoek and colleagues, 2022). Huduku’s line of 15% for one owner is its own choice.

3. Even out position. Within a section, among neighbouring cards whose stories the same number of owners carried, the owner with fewer cards higher on the page now goes first. It is the same for everyone and can be checked against the page, and it never moves a card past one that more owners carried.

4. Take balance to the other lists. Search lists at most 8 headlines from one owner among its 100, and topic pages 6; the rest are behind a “More from …” line at the end. “N reports” counts outlets, so an outlet’s video and article on a story are one report.

5. Explain the rules. How Huduku keeps balance names each rule, its numbers and its reason, and shows today’s counts.

Lower on the list: balancing outlets within an owner as well as owners (Gómez, Boratto and colleagues), and checking that no owner fares worse than it would under a random order (Saito and Joachims).

India: concentrated supply, no audits

This research found no published audit of source concentration in Google News in India, Dailyhunt, Inshorts or Indian search results. These reach many readers: in the Reuters Institute’s 2022 survey of mainly English-speaking Indians online, 53% used Google News, 25% Dailyhunt and 19% Inshorts (Reuters Institute).

The supply they draw on is concentrated. Four Hindi dailies hold 76.45% of Hindi readership, the top two papers in each regional language hold over half of its readers, and Indian law sets no thresholds for concentration (Reporters Without Borders). An aggregator that mirrors that supply reproduces it, so balancing by owner matters at least as much in Kannada as in English. India’s competition regulator, ordering an investigation in 2022, said it appeared that news publishers had no choice but to accept Google’s terms (TechCrunch), and a study of the Google News Initiative’s programmes in India found that they favoured established publishers (Agarwal).

The nearest Indian standard is TRAI’s 2014 recommendation: measure news concentration with HHI in 12 markets by language and state, Kannada among them; treat a market above 1,800 as concentrated; and, where both the TV and the newspaper markets are concentrated, stop an owner from contributing more than 1,000 points to both (TRAI, 2014). Its unit, one owner in one language market, is the unit Huduku’s counts use. TRAI reopened the question in 2022 (TRAI, 2022), and Indian law still sets no such thresholds (Reporters Without Borders).

The legal position is unsettled. The IT Rules of 2021 name news aggregators among publishers bound by a code of ethics, and don’t define curation (MeitY). The Bombay High Court stayed two of their provisions in August 2021 (Internet Freedom Foundation) and the Madras High Court did the same in September (The Tribune). A draft amendment of March 2026 would widen that part of the Rules (IFEX; SFLC.in). This research couldn’t confirm where the stays and the draft stand now. Whether Huduku counts as a news aggregator under the Rules is a question for its pending legal review. No Indian rule requires or forbids balancing by owner.

What isn’t known

No study has tested whether readers find pages with equal turns per owner less useful, and no published computer-science paper evaluates turns by owner on news. The EU’s rules on how very large platforms treat media content (the European Media Freedom Act, Article 18) and on explaining how platforms rank (the Digital Services Act, Article 27, as described by the DSA Observatory) don’t bind Huduku; here they are a yardstick, not an obligation.

The full notes

The report and the notes it drew on, with every source, and the October 2026 ownership audit are published as they were written: The research notes. They are working notes from before the citation check; where they differ, this page is the checked version.