Your Neutral Algorithm Is Not Neutral. Your Follower Graph Decides That.

Based on the research ofMorini, Pansanella, Pappalardo, Pedreschi, and Rossetti, "Joint Effects of Recommender Systems and Network Structure on the Visibility of Content and Creators," arXiv preprint, 2026

Seven Rankers, Two Networks, One Uncomfortable Result

Virginia Morini, Valentina Pansanella, Luca Pappalardo, Dino Pedreschi, and Giulio Rossetti built the test that most platform post-mortems never get to run: a controlled one. Using YSocial, an open agent-based "virtual twin" of a microblogging platform, they seeded 1,000 agents onto a social graph, disabled the large-language-model layer that normally drives agent behavior, and let 1,440 hourly rounds (60 simulated days) play out under a fixed action distribution, 30% posting, 50% commenting, 20% reacting, so that whatever happened to visibility could be traced to the ranking logic and the network, not to noisy content generation.

They ran seven feed-generation strategies. Four are global, ranking from the platform's entire active catalogue: reverse-chronological (the baseline), popularity (ranked by cumulative reactions), and two collaborative filters, one on item similarity and one on user similarity. Three are network-aware, filling 60% of the feed from followed accounts before ranking: a follower-chronological baseline, a follower-plus-popularity hybrid, and a multi-signal linear ranker that blends recency, follower status, and collaborative signals, modeled explicitly on the architecture X (formerly Twitter) published when it open-sourced its own recommendation pipeline. Each combination ran on two network topologies: a scale-free graph built by preferential attachment, and a random Erdős, Rényi graph with the same density but no hubs.

The headline number: under global popularity ranking, content-level inequality (Gini) hits 0.989, against 0.680 for plain chronological ordering, and only 1.8% of the active catalogue is ever recommended to anyone, down from 38%. Half of all creators (50.9% coverage, a 49-point drop from baseline) never get a single piece of content surfaced. Swap in collaborative filtering and the picture inverts: content Gini falls to roughly 0.19 and 99.7% of the catalogue circulates. Recommender logic, on its own, is exactly as powerful as platform critics assume it is.

When the Follower Graph Enters the Ranking, the Story Flips

Here is where the paper earns its title. Feed the same popularity signal through a follower-graph filter instead of the global catalogue, and the numbers move in a way that a single-variable audit would miss entirely. Follower-plus-popularity concentrates creator-level visibility (Gini 0.721) almost as severely as pure global popularity did (0.728), but the mechanism is different. Under global popularity, whoever caught an early reaction wins, largely regardless of their position in the network. Under follower-plus-popularity, the paper's degree-resolved analysis shows exposure and reach rising monotonically with a creator's number of followers: visibility is redirected specifically toward creators who were already socially popular before the simulation even started. Coverage even holds at 100% of creators, nobody is formally excluded, while the actual volume of attention pools overwhelmingly at the hubs. Everyone is technically reachable; almost nobody but the hubs is actually seen.

The multi-signal linear ranker, closer to what a modern feed looks like in production, splits the difference: creator Gini rises only to 0.423, a fraction of what either single-signal strategy produces on its own, because blending recency, affinity, and collaborative weight dilutes any one channel's grip on the outcome. And the topology matters, but only as a volume knob, not a direction changer: run the identical seven strategies on a random graph with no hubs, and the follower-plus-popularity creator Gini drops from 0.721 to 0.434, real, but the qualitative story (network filtering redirects popularity toward whoever is already well-connected) holds in both graphs. Scale-free networks, the kind preferential attachment produces and the kind most real social platforms actually resemble, simply amplify a mechanism that random graphs also contain.

Real Platforms Are Already Running This Experiment

You do not need a simulator to see the two regimes; production platforms have already chosen sides. TikTok's own engineering post on the For You feed states the design choice directly: "neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system," even while acknowledging that more-followed accounts still see more views "by virtue of that account having built up a larger follower base." That is close to the paper's network-blind, global-catalogue condition, ranking that does not filter candidates through the social graph, and it is a large part of why TikTok became famous for zero-follower virality in a way that follower-graph-native feeds structurally cannot replicate.

Instagram sits at the other end. Adam Mosseri has said publicly that reverting to a chronological feed makes the experience "much worse" because it lets accounts that post most often, companies, brands, publishers, flood the timeline, and that "creators get drowned out, and then your friends get drowned out even more"; he has cited an engagement drop of up to 50% in Instagram's own tests of chronological ordering. That is a platform executive describing, in his own words, a follower-graph-filtered feed doing exactly what this paper's follower-plus-popularity condition does in simulation: full nominal coverage, with volume pooling at whoever already has the most reach inside the graph.

YouTube offers the observational echo of the paper's temporal-reinforcement finding. Mozilla's RegretsReporter study, built from data volunteered by 37,380 users across 190 countries, found that 71% of regretted videos arrived via recommendation rather than search, and that those recommended videos accumulated a median of 5,794 views per day versus 3,312 for other videos users watched, attention compounding on content that had already caught some, the same reinforcement loop the paper traces mathematically inside YSocial. None of these companies ran the controlled experiment; they are simply living inside one cell of it at a time, at a scale the simulation is built to approximate safely.

Recommender logic sets the visibility regime. Network structure decides who inside that regime actually wins.

The Managerial Rule

The paper's own framing is a warning against single-lever thinking: visibility allocation, its authors write, "should be evaluated across content, creators, network position, and temporal reinforcement", not collapsed into one aggregate fairness score. A team that ships a demonstrably fairer content-level ranker, verified against a global candidate pool, can still leave creator-level inequality untouched or worse if the follower graph is doing the filtering before the new ranker ever sees a candidate. That is the mechanism chain in miniature: recommender logic plus network topology jointly sets whether visibility concentrates or spreads, and popular creators are the beneficiaries whenever the graph is left to do quiet work the ranking audit never inspected.

The uncomfortable implication for governance teams is that "fix the algorithm" and "fix the network" are not sequential fixes for the same problem; they are two separate levers that interact, and testing one while holding the other's effect invisible will produce a compliant-looking ranker sitting on top of an unchanged visibility monopoly. If your platform's follower graph is scale-free, and most organically grown social graphs are, that interaction is not a tail risk. It is the default behavior of any popularity-adjacent signal you have not explicitly tested against it.