Do online ratings actually reflect independent customer opinions?

The standard implicit assumption when reading online ratings is that each rating is an independent observation of customer experience: average them and you have an estimate of product quality. Moe and Trusov's analysis decomposes observed ratings into a baseline ratings component (the consumer's "socially unbiased" evaluation), a social-dynamics component (the influence of previously-posted ratings), and an idiosyncratic error component, then models product sales as a function of these components.

The findings are nuanced. Substantial social dynamics exist in the ratings environment — previously-posted ratings influence subsequent ones. These dynamics have both direct effects on sales (changes in average rating drive immediate purchases) and indirect effects (today's ratings influence tomorrow's ratings, which affect future sales). Some of the indirect effects mitigate long-term impact: when opinion variance is high, the social-dynamics-induced shifts get averaged out over time.

But the headline conclusion is that observed ratings do not always accurately reflect product performance. Even before Schlosser's negativity-bias finding or Hu et al.'s self-selection result, this paper documents that ratings are influenced by prior ratings. Marketers, recognizing this, invest in creating favorable ratings environments — not because they expect to fool customers but because the system actually works that way.

For recommender systems consuming ratings as input: the data is socially-conditioned, not just preference-conditioned. Treating ratings as independent observations leads to biased estimates of product quality and consequently biased recommendations toward whatever ratings dynamics happened to favor early. The fix is structural, not statistical: model the social conditioning explicitly.