A feedback loop from reader signals: which we trust, and which we do not

In journalism the question 'what does the reader actually want to read' has never had a clean answer. The 2010s analytics boom offered the illusion of one — pageviews, social shares, dwell time — but editors quickly learned those signals mislead. At Mediaorigo we have spent three years building a feedback loop that is deliberately conservative. Here is what we learned.

The three signals we collect

Dwell time — how long a visitor spent on the article, adjusted for scroll position and tab activity (an article left open in another tab is not an article being read). We measure plausibility with an 'average reading speed plus natural pauses' model — a 600-word article should hold 2:15 to 3:45 minutes; significantly less is scanning, significantly more is multitasking.

Scroll depth — how far into the article the reader got. The simple version is uninteresting; we added scroll-pause as a distinct event (5+ seconds parked on a paragraph), and scroll-pause has become a much stronger indicator of engaged reading than depth alone.

Share-with-quote — the reader highlighted a sentence or paragraph and shared it. This event is rare but carries very high signal value: the reader is telling us, in effect, 'this is the point of the article as I see it'.

One signal we do not collect

We do not collect explicit likes. We ran an A/B in 2024 that found 90% of likes arrived from readers who had not reached the bottom of the article — judgments made on the title. The correlation between likes and the returning-reader cohort was effectively zero. Likes are a popularity signal, not a quality signal. The two sometimes agree, often do not.

The signal we trust most

Returning-reader-within-7-days. An article succeeds when at least 12% of its new readers come back for something else on Mediaorigo within a week. We treat that number as our 'I won this reader' quasi-metric. It does not say the article was good; it says readers of this article were more likely to remain readers. That is a finer distinction than it sounds.

Why we trust it: long horizon, not manipulable by a single click-magnet, and it measures the relationship between reader and publisher rather than the click of the moment.

The signal we trust least

Immediate shares. 70% of shares in the first 60 minutes are context-poor — the sharer has not read past the lede, and the share intent belongs to the sharer (ideological, group-signaling, viral participation), not to the article. This is not a signal of article quality. It is a signal that the title was shareable. The two are not the same.

We do not ignore it — but it does not enter our internal importance index.

Editor's pick versus algorithm's pick

The homepage carries two modules: 'Editor's selection' and 'By reader interest'. The second is algorithmic, generated from the behavioral signals above. The two lists typically overlap 30–40%. We have watched the divergence for six months. The conclusion: we do not want them to converge.

The editor's pick is a significance judgment — pieces the newsroom believes readers should read. The algorithm's pick is an interest judgment — pieces readers, on average, want to read. Showing both side by side is more honest with the reader than blending them into a single ranked feed. The reader knows which is which and can choose.

A good feedback loop does not mean handing the decision to the algorithm. It means a substantive dialogue between editorial judgment and reader behavior, with each knowing its own role. That distinction is the difference between a publication and a feed.