The popular framing of algorithmic recommendation, repeated across most contemporary writing on consumer technology, is that personalised suggestions surface preferences the consumer already has. The algorithm watches what someone does, identifies a pattern, and offers more of what the pattern predicts they will like. On this account, the algorithm is a useful but essentially neutral tool, one that accelerates discovery without changing what the consumer would have chosen if left to themselves.
The peer-reviewed experimental evidence accumulated over the past twenty years has substantially complicated this picture.
What the controlled studies have repeatedly shown is that algorithmic recommendations do not merely reflect existing preferences. They actively shape them. The brain treats the recommendation itself as a piece of information, an informational anchor, that influences subsequent judgment regardless of whether the underlying algorithm is accurate or not. The same effect persists when the recommendations are deliberately randomised, when they are intentionally incorrect, and when the consumer has been explicitly told that the recommendations may be unreliable. The mechanism is not driven by the algorithm being right. The mechanism is driven by the algorithm being present.
The Adomavicius experiments
The foundational peer-reviewed study of this effect was published in 2018 by Gediminas Adomavicius at the University of Minnesota, with Jesse Bockstedt at the University of Arizona, Shawn Curley at Minnesota, and Jingjing Zhang at Indiana University, in the journal Information Systems Research. The team designed three controlled experiments to test whether personalised recommendation ratings produced by a recommender system would influence the prices that participants would be willing to pay for digital music tracks.
The first experiment showed participants algorithmic recommendation ratings for a series of songs. The ratings were presented as if they had been personalised to each participant by a recommender system trained on their listening history. The ratings were, in fact, randomly assigned. Participants were then asked how much they would be willing to pay for each song. The team found that the randomly assigned ratings substantially altered the prices participants were willing to pay, even when the team controlled for the participants’ own underlying preferences and demographics. Higher random ratings produced higher willingness-to-pay figures. Lower random ratings produced lower figures. The shift was statistically robust across the sample.
The second experiment used actual system-generated recommendations from a real recommender system, but the team intentionally perturbed the recommendations by introducing controlled errors. Participants saw recommendations that were partially based on their preferences but deliberately shifted away from what an unperturbed algorithm would have produced. The team observed substantially the same effect. The participants’ willingness to pay tracked the perturbed recommendations rather than their underlying preferences.
The third experiment reduced preference uncertainty by giving participants substantial time and information to develop firm views about the songs before they were exposed to the recommendation ratings. The hypothesis the team tested was that participants who had already made up their minds about the music would be less susceptible to algorithmic influence. The hypothesis did not hold. Even participants with stable, well-developed preferences showed the same willingness-to-pay shifts in response to randomly assigned algorithmic recommendations.
The anchoring mechanism
The Adomavicius team’s interpretation, supported by the broader peer-reviewed literature on cognitive anchoring effects, was that the recommendation operates as a numerical anchor that the brain uses as a starting point for subsequent judgment. The classic anchoring effect, identified by Daniel Kahneman and Amos Tversky in the 1970s and replicated in thousands of subsequent studies, shows that arbitrary numerical values presented before a judgment systematically shift the judgment toward the anchor, even when the people making the judgment know the anchor is irrelevant.
Algorithmic recommendations function as anchors in the same way. A recommendation of four out of five stars sets a numerical reference point that influences the consumer’s subsequent valuation of the product, regardless of whether the four-star recommendation reflects anything meaningful about the consumer’s actual preferences. The consumer who has been told a song is worth four stars, by what they believe to be an accurate personalised algorithm, will offer a price closer to what a four-star song commands than they would have offered in the absence of the recommendation. The shift occurs whether the four-star rating is generated by an accurate algorithm, an inaccurate algorithm, or a random number generator.
The implication is that the source of the algorithmic influence is not the accuracy of the algorithm. The source is the presence of the algorithm. The consumer is being influenced by a numerical informational anchor that has been generated by a system the consumer trusts to be measuring their preferences. Whether the system is actually doing that is, on the peer-reviewed evidence, largely beside the point.
The Amazon and YouTube specifics
The peer-reviewed experimental findings sit alongside a substantial body of industry data on the actual scale of algorithmic influence in major consumer platforms. Amazon’s recommendation engine, on figures the company has publicly disclosed and that have been independently analysed by McKinsey and Company since 2013, drives approximately 35 per cent of all purchases made on the platform. The figure has held roughly steady for over a decade, despite extensive evolution in the underlying algorithms and substantial growth in Amazon’s product catalogue.
YouTube’s chief product officer Neal Mohan disclosed at the Consumer Electronics Show in January 2018 that approximately 70 per cent of all viewing time on YouTube comes from videos selected by the platform’s recommendation algorithm rather than from videos that users have actively searched for. The figure has been widely confirmed by subsequent industry analysis and remains the standard reference for the role of algorithmic recommendation in video discovery. YouTube now has approximately 2.5 billion monthly users.
The figures across other major platforms follow similar patterns. Pew Research Center has found that approximately 81 per cent of YouTube users say they at least occasionally watch recommended videos, with 15 per cent reporting that they do so regularly. The recommendation algorithms on TikTok and Instagram measure user engagement to the millisecond and adjust the content stream in real time based on which videos a user has lingered on, replayed, or scrolled past quickly. The effect of all this, on the available evidence, is that the majority of the content the average person encounters in the major consumer platforms is selected by algorithms rather than chosen by the person consuming it.
The manipulation framework
A 2019 peer-reviewed article by Daniel Susser at Pennsylvania State University, Beate Roessler at the University of Amsterdam, and Helen Nissenbaum at Cornell Tech, published in the Georgetown Law Technology Review, argued that the systematic deployment of algorithmic influence in digital platforms now meets the philosophical definition of manipulation rather than persuasion. The distinction matters because manipulation, in the philosophical literature, is defined as the deliberate influence of someone’s beliefs, desires, or actions through means that bypass their rational agency. Persuasion, by contrast, addresses someone’s rational agency directly by giving them reasons to change their beliefs or actions.
The Susser team’s argument was that algorithmic recommendations, as currently deployed, systematically bypass the rational agency of consumers. The consumer is not given reasons why a particular product is being recommended. The consumer is not told what factors the algorithm has weighted, or how heavily, or what alternatives have been excluded. The consumer is presented with a numerical or categorical anchor and expected to make a decision in response to it. The cognitive mechanism the recommendation activates, on the Adomavicius experimental evidence, is the anchoring effect, which is precisely the kind of non-rational decision-making bias that the philosophical literature has long identified as the substrate of manipulation.
The Susser argument has not produced consensus. Other scholars have argued that the line between manipulation and persuasion is blurry, that consumers retain meaningful agency to ignore or reject algorithmic recommendations, and that the consumer welfare benefits of algorithmic curation outweigh the autonomy costs. The peer-reviewed dispute is genuinely ongoing. What is not in dispute is that the cognitive mechanisms identified by the Adomavicius team operate in ways the consumer is generally unaware of, and that they produce substantial behavioural effects that the consumer would not necessarily endorse if they were aware of them.
The honest limitations
Several methodological caveats apply to the literature described above.
The Adomavicius experimental work has been conducted primarily in laboratory settings using digital music as the stimulus. The generalisability of the findings to other product categories, other recommendation contexts, and the open-ended environment of actual consumer life is plausible but not perfectly established. The effect sizes observed in the lab may be larger or smaller than the effect sizes that obtain in everyday digital commerce.
The industry figures for Amazon and YouTube are derived from corporate disclosures rather than from independent peer-reviewed measurement. The 35 per cent and 70 per cent figures are widely cited and broadly consistent with the available academic analyses, but they should be treated as informed estimates rather than as precise scientific measurements.
The Susser 2019 manipulation framework is a philosophical argument rather than an empirical finding. The conceptual definition of manipulation, and the application of that definition to algorithmic recommendation systems, remain genuinely contested in the peer-reviewed legal and philosophical literature. The argument is strong but is not the only available position.
The anchoring mechanism identified by the Adomavicius team is one of several cognitive effects that have been proposed to explain algorithmic influence on consumer behaviour. Other proposed mechanisms include social proof effects, decision fatigue and the deferral of judgment to perceived authority, identity-reinforcement effects in which recommendations mirror and amplify the consumer’s existing self-concept, and choice architecture effects in which the structure of the recommendation environment itself shapes which options the consumer considers. The relative contributions of these mechanisms are not yet fully understood.
What it means
Several things follow from the peer-reviewed evidence on algorithmic influence on consumer choice that are worth saying clearly.
The first is that the popular framing of algorithmic recommendation as a neutral discovery tool is not supported by the available experimental evidence. The recommendation is not a neutral observation of what the consumer would have chosen. It is an active intervention in the consumer’s decision-making process, operating through cognitive anchoring mechanisms that are substantially independent of the recommendation’s accuracy. The consumer who receives a recommendation is being measurably influenced by the fact of the recommendation, not just by its content.
The second is that the scale of algorithmic influence in modern consumer platforms is substantial. Approximately 35 per cent of Amazon purchases, approximately 70 per cent of YouTube viewing time, and the majority of content consumption on TikTok, Instagram, and the major streaming platforms now come from algorithm-driven recommendation rather than from user-initiated search. The figure means that, for most consumers in most digital environments, what they encounter is selected by algorithms rather than by them.
The third is that the cognitive mechanism producing the algorithmic influence is one the consumer is generally not aware of, and one the consumer cannot easily resist even when they have been told the recommendation may be unreliable. The Adomavicius team’s third experiment, in which participants with firm pre-existing preferences still showed the influence effect, suggests that informed scepticism about algorithms does not, by itself, neutralise the cognitive mechanisms by which algorithms shape decisions.
The fourth, on the strongest current reading of approximately twenty years of peer-reviewed evidence in behavioural economics, decision psychology, and the philosophy of digital technology, is that the consumer who believes their choices in algorithmically curated environments are predominantly their own is, on the available evidence, mistaken in a measurable way.
The choices are partly theirs.
They are also partly the algorithm’s.
The peer-reviewed evidence has not yet established the precise weighting, but it has established that the second contribution is larger than the popular framing has assumed.
