Personalization Was Supposed to Help the Little Guy. It Did the Opposite.

Based on the research ofZhang, Tong, Luo, Lin, and Li, "AI orchestrator: How recommendation algorithms shape complementor strategy and market equality," Strategic Management Journal, 2026

That is the finding from a new study in *Strategic Management Journal*, and it inverts the story most platforms tell themselves about what "smarter" ranking is for. Researchers Zhang, Tong, Luo, Lin, and Li tracked a food-sharing platform through two sequential algorithm upgrades: first from a location-based baseline to a popularity-based system (PopRec), then from PopRec to a personalization-based system (PersRec) that tailors recommendations to each user's individual history. Across more than 1.7 million observations, they found that PopRec pushed sellers to concentrate on a handful of hit offerings, while PersRec pushed the same sellers to introduce new products and diversify. The two strategies could not coexist. And the redistributive effect ran backward from intuition: PopRec cut into superstar revenue and lifted long-tail sellers, narrowing inequality, while PersRec did the opposite, asymmetrically rewarding the sellers who were already winning.

The Promise Personalization Never Quite Keeps

The intuition every platform designer starts with is that personalization is the fairer technology. A ranking that shows the same popular items to everyone is, almost by construction, a rich-get-richer machine: whoever is popular today gets shown to more people tomorrow, which makes them more popular the day after. A ranking tuned to each user's individual taste is supposed to break that loop, surfacing niche items to the specific people who would actually want them rather than funneling everyone toward the same handful of bestsellers. A 2024 survey of the recommender-systems literature, by Klimashevskaia, Jannach, Elahi, and Trattner, states the promise directly: personalized recommenders are supposed to "increase the visibility of items in the long tail," the lesser-known items a catalogue-wide popularity ranking would bury. That is the sales pitch. It is also, per the same survey, not what tends to happen, the authors find that recommendation algorithms in practice often exhibit a popularity bias instead, concentrating on already-popular items and risking a self-reinforcing "Matthew effect" over time.

Two well-known studies of commercial recommender systems back this up empirically. Fleder and Hosanagar, writing in *Management Science* in 2009, modeled and simulated how standard recommenders behave and found that common designs can reduce sales diversity rather than expand it, because collaborative-filtering systems recommend based on existing sales and ratings data and cannot surface a product with a thin track record, even one users would rate favorably if they saw it. Lee and Hosanagar followed with a randomized field experiment across dozens of thousands of products at a major online retailer, examining collaborative filters like Amazon's "customers who bought this item also bought" feature; they too found these systems tend to pull aggregate sales toward concentration even when individual shoppers feel like they are exploring. A 2023 study of Spotify by Tofalvy and Koltai, published in *New Media & Society*, extends the pattern to music: the platform's recommendation system, they argue, reproduces the industry's existing core-periphery structure rather than flattening it, systematically favoring already-connected acts.

Why the Cruder Ranking Was the Fairer One

Against that backdrop, PopRec should have been the villain. It is, after all, a system that explicitly ranks by popularity, the textbook Matthew-effect design. But the food-sharing platform's location-based baseline was not neutral either. Proximity is its own kind of incumbency advantage: a seller who happens to be closest to a cluster of hungry users wins by geography, regardless of whether their food is any good, and a talented newcomer three blocks farther away simply never gets seen. Replacing that with a single, transparent popularity signal did something location-based ranking could not: it let word-of-mouth and demonstrated demand travel across the whole local market rather than staying trapped inside whoever happened to be nearest. A seller who was not the closest option, but was clearly the best-reviewed and most-ordered-from, could now out-rank a mediocre incumbent. The result, per the paper, was that PopRec's biggest beneficiaries were long-tail complementors, and its biggest losers were the platform's existing superstars, whose advantage under the old system had rested partly on physical position rather than on being demonstrably worth ordering from twice.

The algorithm marketed as fairer to each individual user turned out to be the one that made the market less fair to individual sellers.

PersRec, arriving next, is where the promise of personalization runs into a structural problem. A system trained on each user's own purchase history has, definitionally, more historical data about a superstar seller, more prior orders, more prior ratings, more signal to learn from, than it has about a seller who has only ever sold to a handful of people. When personalization tries to predict what a given user will like next, it is disproportionately equipped to make confident predictions about the sellers users have already interacted with, which tend to be the same big names PopRec had just started to dislodge. The Fleder-Hosanagar and Lee-Hosanagar findings about collaborative filtering describe almost exactly this dynamic in retail generally; the SMJ paper shows it playing out, with real revenue at stake, in a live two-sided marketplace.

Complementors Play Two Different Games

What makes the finding more than an algorithm-design footnote is how sellers responded, because the two responses were opposites and mutually exclusive. Facing PopRec, where a single popularity score determines visibility, complementors concentrated: they narrowed their offerings down to their one or two best-performing items, since a mediocre listing dragging down an aggregate score does more harm than good on a system that ranks by overall popularity. Facing PersRec, where the algorithm looks for a match to each user's specific taste, complementors did the reverse: they introduced new products, betting that a wider catalogue increases the odds of matching some slice of the audience the personalization engine is trying to serve. Neither strategy is irrational given the algorithm sellers are actually facing. But a seller cannot simultaneously concentrate and diversify, and the platform cannot ask for both at once without contradicting itself. Whichever algorithm is running, it is quietly selecting which strategic playbook the entire supplier base plays.

The Governance Question Platforms Keep Avoiding

The paper's framing is careful to call this an "AI orchestrator" problem rather than a purely technical one: the algorithm is not just sorting search results, it is allocating economic outcomes among the businesses that depend on the platform for their livelihood, and it is doing so with a direction that runs opposite to what the algorithm's own marketing implies. That is the same accountability gap that shows up whenever a platform swaps one automated system for a supposedly better one without asking who wins and loses in the reshuffle. Personalization is genuinely good at what it is built to do: matching individual users to items they are likely to want. What this study adds is that being good at matching demand is not the same as being neutral about who on the supply side gets to grow. The next algorithm upgrade a platform ships is also a policy decision about market structure, whether or not anyone frames it that way internally.

For any platform running a recommendation stack, the practical takeaway is to stop treating "more personalized" as synonymous with "more fair to sellers." The two properties are not the same axis, and this paper is direct evidence they can move in opposite directions inside the same market, on the same product, within the space of a single sequential rollout.