The First Reader

Every published article has a first reader.

Before that first encounter, however, the article exists in a peculiar state. It may be brilliant, mediocre, or forgettable, but to the product distributing it, those possibilities are almost indistinguishable. The cold start problem, then, isn’t fundamentally a quality problem but a learning problem. Before readers interact with a piece of content, there is nothing from which a recommendation system can learn.

The advice given to creators reflects that reality. Build an audience, publish consistently, engage with readers, grow your network, and keep showing up until the algorithm notices. We tend to treat this as marketing advice, but it is equally a consequence of product design. Today’s product design asks creators and their communities to solve its cold start problem.

The Behavioural Proxy Paradigm

Recommendation systems learn through behavioural proxies. Clicks, reading time, subscriptions, shares, comments, and countless other interactions become observations from which products infer what readers value. They are observations, not verdicts.

Their limitation lies in their timing. Behavioural proxies only exist after exposure. A newly published article has no behavioural history because nobody has encountered it. The first recommendation must therefore be made before the system has anything meaningful from which to learn.

Machine learning addresses this uncertainty through the exploration - exploitation trade-off. Exploitation scales what already appears to work. Exploration deliberately allocates attention to uncertain candidates so the system can learn whether they work.

Digital advertising illustrates the pattern. When a new creative enters an advertising platform, there is no historical click-through rate, conversion data, or audience profile from which to predict its performance. The platform allocates exploratory impressions, observes the resulting behaviour, learns which combinations perform best, and gradually shifts delivery towards those that consistently outperform. Exploration isn’t an alternative to optimisation; it is the mechanism that makes optimisation possible.

Recommendation systems face the same constraint. Behavioural learning cannot begin until someone encounters the content.

Content-Based Exploration

Until recently, recommendation systems had no practical way to guide exploration before behavioural learning began. Recent advances in foundation models make that possible.

Content-based methods have long addressed cold start by matching item attributes to user profiles. Foundation models extend that approach by making richer content proxies available for exploration before behavioural learning begins.

Rather than waiting for behavioural evidence, products can infer characteristics that may indicate a promising candidate for exploration, including:

• Originality

• Conceptual depth

• Semantic richness

• Rhetorical structure

• Coherence

• Information density

• Novelty

• Repetition and derivativeness

These are not measurements of quality, just as behavioural proxies are not measurements of quality. They are simply a different class of proxy, derived from the content itself rather than from readers’ responses.

Consider two newly published essays.

The first is titled How to Monetize Your Writing.

The second is titled The Impressionism of Writing: Manet, Monet, Monetized Impressions.

Neither essay has accumulated a single click, comment, or subscription. From the perspective of behavioural learning, they are equally unknown.

From those titles alone, content-based exploration can infer that the second exhibits richer semantic relationships, denser conceptual structure, historical references, phonetic progression, and rhetorical wordplay. Those signals don’t determine whether readers will embrace the essay. They simply provide better information for allocating the earliest stage of exploration.

Content-based exploration doesn’t replace behavioural exploration; it bootstraps it. Its purpose isn’t to determine what readers will value but to create the conditions under which behavioural learning can begin. Once behavioural signals emerge, the familiar recommendation pipeline resumes: behavioural learning evaluates the exploratory decisions, and behavioural exploitation scales what readers consistently validate.

The Weight of Discovery

Ideas do not compete on merit alone. They also compete on whether they are discovered in the first place.

Content-based exploration doesn’t guarantee that quality content succeeds. It simply gives quality content a better chance of entering circulation by making the earliest stage of discovery more informed, while leaving readers to decide what gains momentum.

Every article has a first reader. Quality content should have a fair chance of finding one.

Read 'The Circulation Problem' →