Zum Inhalt springen
Mera News

Diese Seite ist nur auf Englisch verfügbar.

Reading check

The reading check is a research project by Mera Labs. It is hosted on mera.news for now.

A tool opens a publication's newest article pages in a phone-sized browser, moves through each one the way a reader would, and scores how easy the page is to read. This page is the full method: what is measured, why, how to repeat it and how to improve a score.

It is not scraping or crawling, and it adds no news to Mera News. Mera's news comes only from the RSS feed a publisher publishes. See Feeds and the fetcher. The reading check opens article pages a reader would open, up to 3 per site.

Why it exists

Mera News shows a headline and sends the reader to your page. What the reader meets there decides whether they stay.

Readers leave pages that are hard to read. A good article page is how a publisher keeps a reader who could have asked an AI assistant instead. The reading check shows a publisher what a first-time reader on a phone meets, with screenshots, and points to the fix.

Ads are fine

The page is judged as you serve it, on a phone, with its ads. No finding tells a publisher to remove ads or to rely on reader mode. Reader mode is a browser view that strips a page down to its text, and it removes the ads publishers live on.

The tool uses a reader-mode parser only as a ruler. The parser tells the tool which part of the page is the article. The tool then judges the real page against that.

Only article pages are judged. A home page or a section page is refused and never scored.

How a visit goes

The tool takes the newest article links from your RSS feed, up to 3. It opens them in one browser session, so a meter that counts free articles is seen as a reader would see it. The browser is Chromium with the profile of a Pixel 7 phone.

Before any page is opened, the tool requests robots.txt and obeys it. It also looks for a reservation of text and data mining rights. See "Judgement calls" below.

Then, for each article:

  1. Open the page. A page that does not look like an article is refused.
  2. Wait. The tool waits 2 seconds, then until 6 seconds, and records what the first screen holds.
  3. Clear what covers the article. The tool picks a button and taps it, with at most 3 taps per box and 6 per visit. A box that stays is recorded as not closed.
  4. Scroll to the last paragraph. It scrolls 80% of a screen at a time. It taps a "continue reading" button or a next-page link when one hides part of the article, and counts those taps.
  5. Tap plain article text once. Nothing should open.
  6. Scroll back up and pause. The tool waits until the page is quiet, 10 seconds at most.
  7. Narrow the screen to 320 px. It checks whether the page scrolls sideways.

A screenshot is taken at each step. Once per site, on the first article that was read to its end, the tool runs the translation checks.

The browser's processor is slowed to a quarter of its speed, so the visit is closer to a mid-range phone than to a server.

The three lead numbers

A report shows these first, because they are what a reader feels. They are not scores.

NumberWhat it is
Taps to readTaps needed between opening the article and reaching its last paragraph. 0 is the goal. It is the path the tool found, so it can be higher than the true minimum. Every tap is listed. A consent notice's tap counts here even when the notice costs no points. Across a site's articles, the highest count is shown
First screenHow the first screen divides between the article, ads, overlays and the site's own header and bars, and whether the headline is visible. Measured 6 seconds after the page opens
Scroll costScreens of scrolling needed to finish, against screens of actual article

The five criteria

Each criterion is scored from 0 to 100 for each article visit. A site's score for a criterion is the average over its visits. No criterion goes below 0.

CriterionWeight
Pop-ups and overlays25
Ads20
Paywall on first open20
Page structure20
Translatability15

1. Pop-ups and overlays (weight 25)

An overlay is anything that sits on top of the article and covers more than 30% of the screen. Reports call it a box.

Start at 100. Subtract for each overlay that is not a compliant consent notice:

What happenedPoints
It was there when the page opened35
It appeared on scroll30
It appeared while the reader was idle25
It took more than one tap to close, or could not be closed at all15 more
It came back after being closed15 more

Also subtract:

What happenedPoints
Tapping plain article text opened a new tab, went to another page or raised an overlay40
The page asked for notification permission before the reader did anything15
The page raised an app install prompt10

Consent notices. A cookie or age notice that the law requires is not a pop-up, as long as it offers a way to refuse on its first screen. That is a compliant consent notice. It costs no points. Its tap still counts in taps to read.

A notice is treated as an ordinary overlay when it has no way to refuse, when it is "pay or accept", when it carries a promotion, or when it sends the reader to another page.

Fix: Pop-ups and consent.

2. Ads (weight 20)

Only intrusive ads cost points.

Ad density is the total height of ads inside the article, divided by the height of the article. Ads above the headline, below the last paragraph and in the site's header or footer are not counted. A sticky ad, one that stays fixed on the screen while the reader scrolls, counts once.

Ad densityScore
Up to 30%100
From 30% to 60%Falls in a straight line from 100 to 0
Above 60%0

Then subtract:

What happenedPoints
A sticky ad covers more than 30% of the screen30
A video plays with sound without being asked to30
A video detaches and follows the reader while scrolling15

Ads that were not served are not "no ads". A tool that visits a page is often shown fewer ads than a reader. The tool reads the ad slots the page declares and counts how many were filled. If fewer than half were filled, Ads is not rated for that visit. The reason given is ads_not_served. A page that declares no ad slots is rated.

The tool recognises ads with a public filter list (EasyList). It never blocks an ad.

Fix: Ads.

3. Paywall on first open (weight 20)

A paywall is not a fault. Meeting one on the very first article is, because no reader subscribes for a month to read one story.

What a first-time reader getsScore
The full article100
The full article, with a notice that free articles are limited90
The article only after a free sign-up50
The opening lines, then a wall20
A wall and nothing else0

"Limited free articles" is recorded only when the page says so or declares it in its markup. Markup that honestly declares paid content (isAccessibleForFree) is noted in the result.

Only a wall that was actually detected is scored as a wall. A wall the decision model is not sure of is not rated.

Fix: Paywalls.

4. Page structure (weight 20)

Start at 100 and subtract:

What was foundPoints
The first paragraph starts more than 1.5 screens down20
Less than half of what the reader scrolls through is the article20
Less than a third of it is the article (replaces the row above)35
Other blocks interrupt the article more than 3 times10
A "continue reading" button hides part of the article10 each, 20 at most
The article is split over several pages10
Where the article ends is unclear, on a scale from 0 (clear) to 4 (very unclear)5 per step
The page scrolls sideways at phone width15
The page scrolls sideways at 320 px wide5
Body text is smaller than 16 px10, or 20 below 14 px
Body text fails the contrast minimum10
More than 5 tap targets in the article are too small5

A tap target is anything a reader taps, such as a link or a button. Contrast and tap target size are tested with the axe-core accessibility engine. When it cannot run on a page, those two rows are skipped.

Fix: Page structure.

5. Translatability (weight 15)

Many readers read news in a language that is not their own. This criterion asks whether a browser or a translation tool can translate the article.

Start at 100 and subtract:

What was foundPoints
The page or the article is marked not to be translated60
The page does not declare its language15
The declared language is not the language of the text15
More than 20% of the article is images of text30
Text cannot be selected20
Google Translate could not translate the article40

The first five rows are read from the page's own markup. The last row is the translate pass: the tool opens Google Translate's view of the article. A page that is not in English is tested into English. An English page is tested into Hindi.

If the translate pass was not run, nothing is subtracted for it, and the result says the pass was not run.

In Mera's runs the translate pass is switched off for now. Google's robots.txt disallows the path the pass uses. So in published reports, Translatability is scored from the page's own markup. The tool can still run the pass when you run it yourself.

Fix: Page structure, in the section on translation.

The overall score

The overall score is the weighted mean of the criteria that were rated. The weights are rescaled over those criteria. There is an overall score whenever at least one criterion was rated.

The cap. The overall score is capped at 40 when, on 2 or more of the site's visits, either of these happened:

  • A blocking overlay took more than one tap to close, or could not be closed.
  • The reader met a wall with nothing readable.

What is scored and what is not

If the page opened, it is scored as it opened. A box that cannot be closed costs points, because that is what a reader meets.

When an overlay locks the page, or still covers the article at the end, and no way to close it was found, the visit's outcome is could_not_clear. It is scored under Pop-ups: the row for when the box appeared, plus the 15 points for a box that could not be closed. It counts toward the cap. Paywall is not rated for that visit, because nothing was read and no wall was identified.

A box that stays while the article is still read to its end is charged by the row for when it appeared, and no more.

A site has no score only for one of these reasons:

StatusMeaning
blockedThe site could not be reached, refused the visit, showed a bot challenge, or its robots.txt disallows the visit
no_articlesNo article links could be found
not_an_articleThe page given is a home page or a section page
excludedThe site was left out of the run
errorThe tool failed

A site that was reached and had article pages always has a score. Within it, a criterion is not rated for a visit only for its own reason:

  • Ads, when the page's ad slots did not fill (ads_not_served).
  • Paywall, on a could_not_clear visit.
  • The translate pass, when it was not run. The other Translatability rows still count.

Speed is never scored. Load time, layout shift and tap response are measured in a slowed-down test browser. They are shown in the result as diagnostics.

The standards behind the rules

Each rule rests on a published standard where one exists.

StandardWhat rests on it
Better Ads Standards for the mobile web, from the Coalition for Better AdsPop-ups: overlays at load, on scroll and while idle (pop-up ads, prestitial ads, postitial ads with countdown, full-screen scrollover ads). Ads: the 30% density line, the sticky ad over 30% of the screen, video that plays with sound
Google's guidance on intrusive interstitialsPop-ups: boxes that cover the article, app install prompts, and the exemption for notices the law requires
Google's flexible sampling guidancePaywall: the order of the five levels, from the full article to a wall and nothing else
WCAG 2.2, 1.4.10 ReflowStructure: sideways scroll at 320 px
WCAG 2.2, 1.4.3 Contrast (Minimum)Structure: body text contrast
WCAG 2.2, 2.5.8 Target Size (Minimum)Structure: tap targets that are too small
WCAG 2.2, 3.1.1 Language of PageTranslatability: a missing or wrong language declaration
WCAG 2.2, 1.4.5 Images of TextTranslatability: images of text
W3C internationalisation guidance on the translate attributeTranslatability: content marked not to be translated

WCAG is the Web Content Accessibility Guidelines, published by the W3C.

What is the project's own choice. No standard sets these, and they can be argued with:

  • Every point value, the five weights and the cap at 40.
  • The 30% of the screen that makes something an overlay.
  • The 60% ad density at which Ads reaches 0, and the 15 points for a video that follows the reader.
  • The scores for each paywall level.
  • In Page structure: the 1.5 screens, the half and the third, the 3 interruptions, the 16 px and 14 px text sizes, the count of 5 tap targets, and the rows for "continue reading" buttons, split articles and an unclear end.
  • In Translatability: the 20% share for images of text, and the rows for text that cannot be selected and for the translate pass.

Judgement calls

Most of a result is measured: heights, counts, positions, taps. A few calls need reading a box or a wall:

  • What kind of box this is.
  • Which button closes it.
  • What kind of wall stopped the reader.
  • How clear the article's end is.

Built-in rules make these calls first. When the rules are not certain, a small decision model is asked. A decision model is an AI model that can only pick between fixed answers it is given. It cannot write text or take an action of its own. It can be sent a screenshot and the text of the box and its buttons.

Each answer comes with a probability. An answer the model was not sure of is listed in the result as low confidence. With no decision model at all, the tool uses its built-in rules alone and marks those calls low confidence. Each result names what judged the page.

Pages that reserve rights. A page can reserve text and data mining rights with a tdm-reservation header, a meta tag or a /.well-known/tdmrep.json file. Such a page is never sent to a model. It is judged by the built-in rules alone, and the result says so.

Mera trains no model on any of it.

Good manners

  • It names itself. The tool sends a phone browser's user agent with its own name and a link to this page added. A user agent is the line a browser sends to say what it is.
  • It obeys robots.txt. See "Opt out" below.
  • It never hides. There is no stealth mode, and it does not retry from other addresses to get past a refusal.
  • It never solves a CAPTCHA. A site that shows a bot challenge is marked blocked and gets no score.
  • It doesn't sign in.
  • It keeps no article text beyond the headline. See "What is kept" below.
  • It visits rarely. Up to 3 article pages per site, at most once every 90 days.

How to recognise it

In Mera's runs, the user agent is Chrome on Android and ends with:

(compatible; MeraReadingCheck/1.0; +https://mera.news/en/docs/reading-check)
PropertyValue
Name in robots.txtMeraReadingCheck
How oftenUp to 3 article pages per site, at most once every 90 days
Where visits come fromDirectly from Google Cloud in Belgium, with no proxy

Match on the user agent. The network address changes.

Because visits come from Belgium, your site may show that region's consent notices and ads. Each report states its check date. The result records how the site was reached: directly, or through a proxy in a named country. A direct visit shows no country on the report, so this page is where the place is stated.

Opt out

To stop the check, add this to robots.txt:

User-agent: MeraReadingCheck
Disallow: /

No reason is needed. The site is then marked blocked and gets no score.

Opting out of the check doesn't remove you from Mera News. To leave Mera News, see Feeds and the fetcher.

What a report contains

Each checked publication has a report page on mera.news, linked from the publications directory. It shows:

  • The overall score, the number of articles checked and the check date. If a visit went through a proxy, the country it was seen from is shown too.
  • The three lead numbers.
  • The five criteria, each with what was found.
  • Up to 6 findings, the ones that cost the most points first. Each links to the guide chapter that covers the fix.
  • The screenshots of the reader's path, as evidence. They show your page as it was served, for this research only. All rights stay with the publisher. They are kept for 180 days.
  • The tool version, the rubric version and what judged the page.
  • A JSON download of that publication's result.

The download leaves out the feed address. Article links in it are cut at the end of the path, so anything after a ? is dropped.

Reports are in English only. Your browser can translate the page.

A machine made the report, and it can be wrong. The scores change nothing in the Mera News app.

Improve a score

Fix the page, using the guide chapter each finding links to. The next check picks the change up. A result cannot be changed any other way.

CriterionGuide chapter
Pop-ups and overlaysPop-ups and consent
AdsAds
Paywall on first openPaywalls
Page structurePage structure
TranslatabilityPage structure, in the section on translation

What is kept

KeptNot kept
Scores and measured facts, such as how many taps it took to reach the endThe article's text, beyond its headline
The text of boxes and buttonsAnything about your readers
Screenshots of the page as a reader sees it, for 180 days

Scores are stored in Mera's database. Screenshots are stored in Mera's site storage. See What's stored where.

How the method was checked

Before publishing, Mera Labs compared the tool's results on 22 sites, from different countries and in different languages, with a second opinion. A separate AI model rated the same screenshots the way an expert reviewer would. It did not see the tool's scores.

After three rounds of fixes to the tool:

  • The outcomes matched on 21 of the 22 sites.
  • The mean score difference was under 4 points on every criterion.
CriterionMean difference, in points
Pop-ups and overlays1.0
Ads0
Paywall on first open0
Page structureAbout 3
TranslatabilityAbout 2 at most

This is a check with limits. It is not proof. The sample is 22 sites, the second opinion is also a machine, and it saw screenshots, not the live page. No person rated the pages.

Scope of the first run

The first run covers about 6,600 publications that Mera News reads. Government and regulator sites are left out, because they are not publishers.

Known limits

  • One visit, one place, one day. Each article is visited once, from one place, on one day. A rerun on another day or from another country gives different numbers. Ads, consent notices and meters all change with place and time.
  • Images of text on another host cannot be measured. The browser does not let a page's script read a picture served from a different host, so such a picture never counts.
  • A close button can be missed. The tool may not find a button a person would find. That now costs the page points, because a box that is not closed is scored as it stayed.
  • Sound is detected, not heard. Things that happen only in motion, such as a video that plays with sound, are reported by the browser. A screenshot cannot show them.
  • Some boxes cannot be classified without a model. On a page that reserves rights, or when no decision model is set, the built-in rules judge alone. They are marked low confidence when they are not certain.
  • Taps to read is the path the tool found. A reader may find a shorter one.
  • Ads may not be served to the tool. Ads is then not rated for that visit.

Run it yourself

The tool's code will be public on GitHub under AGPL-3.0, so anyone can run it on a page and compare. The repository is not public yet.

Once it is, a run needs Node 22 or newer, no account and no outside service:

npm install
npm run check -- https://example.com/news/some-article

Pass an article address, or --feed with a feed address. The run writes a report with screenshots and a JSON result in the same shape as the download on a report page.

Your numbers will differ from the published ones for these reasons:

  • You visit on another day, from another place.
  • Your run uses the tool's default name, mera-news-reading-check, not MeraReadingCheck. A site may treat the two differently.
  • Your run includes the translate pass unless you pass --no-translate-pass.
  • With no decision model set, more calls are marked low confidence. A decision model is optional. It can run on your own machine.

Versions

This page describes rubric version 2. The rubric is the set of rules and point values above.

Every result carries the rubric version it was scored with and the version of the tool that made it. Any change to a threshold, a point value, a weight or the cap raises the rubric version.

Scores are computed from stored facts. So an old result can be scored again under a new rubric version without visiting the site again.