A portion of page visits is very brief according to visit duration data, not according to eyetracking measurements. Attention varies strongly depending on platforms: short formats dominate on TikTok, while YouTube also accommodates intermediate and long durations. Intermediate formats face obstacles in algorithmic distribution and retention.

The essentials

  • Survey data on video consumption shows contrasting duration preferences depending on the platform. Intermediate formats face obstacles in algorithmic distribution and retention.
  • Navigation data, not eyetracking measurements, document very early page abandonment; studies measure short average attention spans (Gloria Mark, SciLine).
  • In 2023, 16.2% of people aged 15 and over read for personal interest on an average day according to the American Time Use Survey, a decline from 2003. Readers devoted approximately 97 minutes to reading for personal interest on days when they read.
  • Recommendation systems can use engagement and satisfaction signals, but no general rule establishing a preference for extreme durations is established.
  • By 2030, it remains to be determined which economic models allow diverse creation to survive outside the extremes that the algorithm recognizes.

The algorithm does not reward the middle

Studies on online attention document a short average duration of engagement. But this figure masks something more precise. It does not describe a homogeneous population of distracted people. It describes the result of a sorting process.

The acceleration of filtering based on the initial seconds of content is clearly observable. The acceleration is sharp. Content that does not capture attention in the first few seconds is buried by the recommendation algorithm before the user has been able to decide for themselves. The early abandonment signal is treated as a negative signal: less distribution, less visibility, a downward spiral. This mechanism does not punish mediocrity.

Recommendation systems place significant weight on the initial moments of content.

Those who pass the filter enter a different regime. In 2023, 16.2% of people aged 15 and over read for personal interest on an average day according to the American Time Use Survey. Readers devoted approximately 97 minutes to reading for personal interest on days when they read. Long-form podcast subscribers stay until the end. Newsletter readers from niche sources have open rates that mainstream media has not seen in thirty years.

Long-form immersion exists, and it persists.

Long-form immersion can concentrate on certain niches or editorial formats.

The barbell and its effects on creators

Bruno Patino, editorial director of Arte, has documented the attention economy as an infrastructure of power. In his work, he raises the point that certain recommendation algorithms place significant weight on the initial moments of content and on prolonged retention. Recommendation systems can use signals of choice, retention, and satisfaction, but YouTube’s documentation does not describe an inherent penalization of intermediate formats.

The barbell, the bodybuilder’s shape, wide at the extremes, empty in the middle, describes a market structure. An article of intermediate duration can encounter variable challenges depending on the platform and its audience. It can encounter potential obstacles in developing subscriber retention. It occupies an intermediate position where recommendation systems can use variable signals depending on format and context.

This phenomenon affects different sectors with striking consistency. Intermediate-length podcasts encounter increased competition from very short formats and long-form content. Standard-length articles may encounter variable full-read rates depending on context. General-interest in-depth newsletters suffer competitive pressure from ultra-short bulletins and highly specialized thematic publications. Podtrac and Spotify publish audience metrics, but the distribution by content duration is not entirely documented by available sources.

Some economic models organize themselves around high-volume short formats and long-form subscription formats. Intermediate formats can encounter funding challenges in this architecture. A freelance journalist producing in-depth articles without a strong editorial brand finds themselves both too slow for virality and not retentive enough for the subscription model. Distribution structures influence visibility, but quality characteristics such as readability can also affect engagement.

Is attention educable: a counter-reading

The barbell thesis deserves to be confronted with a competing reading. Daniel Susskind, an economist at Oxford who works on the future of work and human capacities, raises in a different register a question applicable here: the constraints we attribute to structures are sometimes constraints we impose on ourselves. The demand for short formats may be linked to platform design choices, and other choices could modify this balance.

This reading is honest, and the data partially bear it out. Certain Spotify indicators attest to interest in longer content and listening sessions. It can persist depending on environmental and design conditions. The argument for the educability of attention—that we can reestablish conditions in which the middle becomes viable again—is defensible.

Navigation behavior data raise an important tension. Early abandonments may measure behavior rather than declared preference. They measure behavior in a highly competitive environment. In this environment, even a reader who would prefer in-depth content does not pause if the hook signal is insufficient. Attention can adapt to different contexts depending on environmental and design conditions.

The distinction between these two contexts is precisely what regulation should address.

A genuine tension remains between the two readings: if demand for long-form immersion persists despite the algorithmic environment, then the barbell is not an anthropological inevitability. It may be a market consequence, in the sense that current incentives can produce and reinforce it. This diagnosis is shareable by both readings, and it is on this ground that the prospective question plays out.

Economic models attempting to resist

A few formats and a few actors have found stable positions outside the dominant algorithmic logic. They merit examination not as exceptions that prove the rule, but as experiments whose mechanisms are instructive.

Newsletters allow direct relationships with existing subscribers, without by themselves proving a relationship of trust or total independence from platforms. The creator is distributed via email, not through a feed. Email has its own constraints, competition in the inbox is real, but it does not impose a seven-second filter on the survival of content. The viability of a freelance journalist depends notably on revenue, conversion rates, pricing, costs, and clicks; a declared open rate is not sufficient to determine it. It is also a model that requires time to build a list, time that the algorithmic platform does not allow.

Editorial cooperatives, several European regional newspapers have adopted this model in recent years, pooling production costs and financing part of revenues directly from readers. In France, titles like Le Média or Blast (contributive model) show that activist loyalty can partially substitute for algorithmic virality. The limitation is that these models rely on a preexisting community or on a strongly marked editorial identity. They do not work for service journalism or niche content without a constituted community.

Patronage and journalism endowment funds, an American model with organizations like ProPublica or The Marshall Project, make it possible to finance long investigations without depending on algorithmic advertising revenues. But they remain marginal in volume and assume fundraising capacities that most freelance creators do not have. These models show that alternatives exist, without demonstrating that they can expand to the point of filling the void left by the collapse of the editorial middle ground.

The decisions the 2030 horizon demands

The question that poses itself is not whether intermediate formats will survive in their current form. The viability of intermediate formats assumes modifications in economic models and distribution systems. The question is what form diverse and remunerated creation can take at the decade’s horizon.

Two trajectories are emerging, and signals to monitor in the coming years will allow us to differentiate between them.

In the first, polarization stabilizes around two viable markets. Platforms and economic models emerge to directly finance long-form immersion, subscriptions, patronage, cooperatives, while short-form continues to function in its volume-based economy. The barbell could produce two distinct markets rather than a single unified market. Regional press, niche podcasts, and in-depth newsletters would find a viable economy in direct relationships with their audiences, without occupying the majority of attention. This scenario assumes that long-form formats continue to recruit first-time readers despite distribution difficulties, and that subscription models become accessible enough not to remain reserved for affluent audiences.

The question of access to content for audiences who cannot afford premium subscriptions remains entirely open in this scenario.

In the second, an algorithmic optimization toward the ultra-short could accelerate without regulation or alternative models. Editorial creation could concentrate in extreme formats. This scenario represents an extrapolation of the concentration trends observed on certain platforms. It poses a problem that exceeds creative economy: a scarcity of formats capable of developing complex arguments on the mass market could affect the quality of public debate.

Deficits in training for long-form skills that algorithmic platforms produce aggravate this dynamic.

Two types of signals will make it possible to see which of the two trajectories predominates. The first: the evolution of the abandonment threshold in Tobii data during upcoming measurements. If early content abandonment continues to increase, selection by ultra-short filtering would intensify. If it stabilizes, the equilibrium hypothesis becomes more plausible. The second: the share of freelance creators’ revenues coming from sources outside recommendation algorithms.

If revenues for creators from sources outside algorithms increase via direct subscription, patronage, or cooperatives, this would indicate a rise in alternative economic models. If this share stagnates, algorithmic dependency would persist.

The scope of regulatory action

Regulating the attention economy is the rising concern. The European Union has paved the way with the Digital Services Act, which imposes transparency obligations on very large platforms regarding their recommendation systems. But transparency does not change incentives. Knowing how the algorithm works is not enough to make it work differently.

Two more direct avenues merit examination. The first is an obligation of enhanced algorithmic transparency, which would allow regulators to measure the effect of recommendation on the polarization of attention, not merely to understand mechanisms in theory, but to measure actual effects on user populations. This measurement exists in part within the DSA framework; it should be extended and standardized to allow comparisons between platforms and over time.

The second is more controversial: a fund supporting intermediate editorial creation, financed by a tax on platform advertising revenues, explicitly targeting formats between three and twenty minutes documented as in decline. The precedent exists in audiovisual with obligations to finance independent production imposed on broadcasters. The question is whether this model is transposable to the digital realm without creating rent-seeking effects for artificially protected formats. The answer depends on the mechanism’s design: a fund targeting editorial diversity and plurality of voices, not protection of particular formats, would be defensible.

What is less defensible is inaction. Prolonged attention polarization risks reducing the supply of service journalism and content developing complex arguments. This cost will not be visible in a quarter. It will be over a decade.

Sources

  1. SQ Magazine & Clutch 2026, Social Media Attention Span Statistics
  2. Bruno Patino, Work on Attention and Algorithm
  3. Gloria Mark, SciLine, Average Attention: 47 seconds (August 2026)
  4. Tobii, Eyetracking (February 2026): early content abandonment data
  5. American Time Use Survey, Bureau of Labor Statistics, Pleasure Reading USA 2003-2023
  6. Podtrac / Spotify, Podcast Listening Data 2025-2026
  7. Digital Services Act, European Commission