Reviews

In 1984, the psychologist Robert Cialdini published Influence: The Psychology of Persuasion, a book that synthesized decades of social psychology research into six principles by which one person convinces another. The third principle, social proof, has become the most consequential of the six in the digital era. Cialdini defined it simply: when uncertain about how to behave, people look to what others are doing and copy them. The principle is older than psychology (humans, like most social mammals, have always survived in part by following the herd) but Cialdini formalized it for the modern reader.

Social proof works because making decisions from scratch is cognitively expensive. Every shopping choice involves dozens of subdecisions about quality, fit, value, and risk. Looking at what other buyers have done is a heuristic (a mental shortcut) that saves enormous effort. If a thousand strangers bought the same item and rated it well, the probability that the item is at least adequate is higher than the probability that all those strangers were uniformly mistaken.

It is worth being precise about why this shortcut is rational rather than lazy. Gathering first hand evidence is expensive, and in most everyday decisions the cost of being slightly wrong is small. Deferring to the apparent consensus is therefore a sensible economy of attention, and it fails only in specific conditions: when the crowd had no better information than we do, when its members were all copying one another rather than observing independently, or when somebody has an incentive to manufacture the appearance of a crowd. Almost every problem discussed later in this page is one of those three conditions appearing in a new costume.

Online product reviews predate most of what we now call ecommerce. Bulletin board systems in the 1980s hosted user-written software reviews and game recommendations. Usenet newsgroups debated audio equipment, books, and movies in long threads that survived for years. But the structured review (a star rating, a written explanation, a verified-buyer indicator) was largely an invention of the late 1990s commerce platforms.

Amazon launched in 1995 and added user reviews almost immediately. The decision was internally controversial (why would a retailer let users publish negative opinions about products it was trying to sell?) but it proved transformative. Buyers trusted Amazon’s negative reviews precisely because they were negative; the visible existence of one-star reviews made the four-star and five-star reviews feel credible by contrast. eBay’s seller-feedback system, launched in 1996, applied the same logic to people instead of products. Yelp, TripAdvisor, Rotten Tomatoes, Steam reviews, App Store ratings: every digital marketplace of the past two decades inherited the basic Amazon-eBay template.

The decisive shift was from reputation held in a community to reputation held in a database. Word of mouth decayed, varied between listeners and could not be consulted by a stranger; a stored score does none of those things. That permanence is what made the format genuinely useful for one off transactions between people who will never meet again, and it is also what created every incentive problem that followed. A reputation that can be looked up can be invested in, and anything that can be invested in can be bought.

The psychology of why we trust anonymous reviewers is more interesting than the surface answer suggests. We do not, in fact, trust any individual reviewer very much. Solomon Asch’s classic conformity experiments in the 1950s demonstrated that people will conform to a group judgment even when they privately disagree with it; the strength of the conformity effect rises with group size and falls with the credibility of any single dissenter. Reviews work the same way. A product with three reviews is barely informative. A product with three thousand reviews tracking a stable distribution is statistically meaningful.

What we trust is the aggregate, not the source. The cognitive shortcut treats the rating as a sample from a population, and our intuitive statistics (though often informal) are reasonably good. We discount one extremely positive review and one extremely negative review. We weigh recent reviews more heavily than old ones. We factor in the rating distribution: a 4.2-star average from ten thousand reviews implies something different from a 4.2-star average from twelve reviews, and our attention to the difference is unconscious but real.

What readers are really doing is estimating whether an account could have been written without the experience. Specific, verifiable detail is persuasive because inventing it takes effort and invites contradiction; generic praise is unpersuasive because anyone could produce it for nothing. This is also why a review that mentions a genuine drawback and still recommends the thing is unusually effective: it demonstrates that the writer was willing to report inconvenient information, which is the cheapest available evidence of independence. Experienced readers are running an informal test of how costly the text would have been to fake.

Anyone expecting opinions to cluster around an average is surprised by real rating data. The characteristic shape is a J, with a large mass at the top, a smaller mass at the bottom and very little between. The cause is selection rather than sentiment: people who feel nothing rarely write anything. Writing a review costs time, and the two emotions that reliably pay that cost are delight and grievance. The resulting distribution is a record of who was motivated enough to speak, not a survey of what everyone thought.

This has a direct consequence for interpretation. A mean computed over a J shaped distribution describes a point where almost no individual sits, so two items with the same average can be entirely different propositions. One may be consistently acceptable to nearly everyone; the other may delight most people and badly fail a minority. The second is riskier even though the number is identical. Reading the shape rather than the summary is the single largest improvement most people can make, and it takes a glance at the bar chart.

The shape also moves with volume. The first handful of accounts of anything tend to come from enthusiasts and early adopters, which pushes the top bar up; broader participation later drags the average down, and that decline is routinely misread as deterioration. A fair comparison therefore holds volume roughly constant, or at least notes it. An item with forty accounts and one with four thousand are not measuring the same population, and the difference between them is usually larger than the difference in their scores.

The obvious method, dividing the sum of ratings by their count, has a defect that becomes obvious with small samples: a single top rating produces a perfect score. Ranking by that number puts items with one account above items with thousands, which is both wrong and easy to exploit. The standard remedy is to pull every score toward the overall average by an amount that depends on how little evidence there is, so one account barely moves from the global mean while a thousand accounts almost entirely determine their own position.

A second family of methods treats the question as estimating a proportion with uncertainty. Rather than asking what the observed share of positive accounts is, these ask what the plausible lower bound on that share is given the sample size, and rank by that bound. The effect is to rank by evidence rather than by enthusiasm: an item with a long clean record outranks one with a short flawless record, which is almost always the judgement a reader actually wants. The same arithmetic underlies ranking in several very different fields.

A third adjustment is weighting by recency, so that a product or a service that has genuinely changed is not permanently defined by its first year. This is desirable and introduces its own hazard, because a sufficiently aggressive decay lets a poor long term record be erased by a good recent month. Every platform picks a position on these three axes, almost none publish it, and the choice has more effect on what readers see than any individual account does. The visible number is a policy, not a measurement.

Beyond psychology, social proof has a sociological dimension. Online reviews function as a low-trust replacement for the high-trust networks of pre-digital commerce. Before ecommerce, you bought a refrigerator from a salesperson at a local appliance store, partly because someone you knew had bought one there. Reviews replace the personal recommendation with the aggregated experience of strangers. The mechanism is different, but the social function is the same: reducing risk by inheriting other people’s verdicts.

This shift has measurable second-order effects. Local merchants compete on review counts as much as on price. Restaurants design plates and seats for the photograph. Hotels train staff to ask for favorable reviews on TripAdvisor and Google. The review economy has restructured industries around the same logic: appearance in the aggregate matters more than reputation in any individual transaction. This is not necessarily a worse system than its predecessor (local reputation also had its biases) but it is structurally different, and understanding the difference is important for buyers and sellers alike.

Herding deserves emphasis because it breaks the assumption that makes aggregation work. An average is informative when its contributors formed views independently; it is far less informative when each contributor saw the running total first. Experiments that artificially add a single early positive to a randomly chosen item find the advantage persists and compounds, which means part of any mature score is a record of its own early luck. This is the second of the three failure conditions named at the outset, and it is the hardest to correct for, because nothing in the visible data distinguishes a consensus from a cascade.

The strongest structural defence against fabricated opinion is to require evidence that the writer actually transacted. The point is not that verified writers are more honest. It is that verification imposes a cost on each fake, turning an unlimited supply of invented accounts into a supply limited by how many real transactions somebody is prepared to fund. Defences of this kind, which change the economics rather than attempting to judge sincerity, tend to outlast defences that rely on spotting a liar from their prose.

The cost is coverage. Verification can only confirm transactions the platform observed, so a marketplace sees its own and nothing else, and plenty of legitimate experience happens elsewhere. Restricting display to verified accounts therefore trades a reduction in fraud for a narrower and sometimes biased picture, and excludes exactly the people whose experience was unusual enough to be worth reading. Most platforms compromise by labelling rather than filtering, which preserves coverage and leaves the weighting to the reader.

There is also a subtler distortion. When verification is tied to a transaction completed on one platform, it quietly privileges the parts of an experience that platform can see. A delivery it arranged is verifiable; durability over two years, a support interaction conducted elsewhere, or a problem that only appears at scale is not. Readers should notice that a strongly verified corpus can be systematically biased toward the first week of ownership, which is often the least informative period.

The intuitive approach, reading text and judging whether it sounds genuine, is the weakest available method and has become weaker as generated prose has improved. The durable signals are behavioural. Authentic accounts of a popular item arrive as a slow irregular trickle shaped by sales; purchased ones arrive in bursts, because they are bought in batches and delivered on a schedule. A sudden cluster after a long quiet period is the single most reliable indicator, and it requires no judgement about anyone’s sentence construction.

Network structure is the second signal. Fabricated accounts come from a limited pool of participants, so the same writers appear across unrelated items in patterns that are vanishingly unlikely by chance. Building a graph of who reviewed what and looking for unusually dense neighbourhoods exposes whole operations at once, including participants whose individual output looks entirely ordinary. This is why detection is done at the level of the population rather than the document; an operation is visible precisely because it is coordinated.

Defenders of a score also exploit a quieter route, which is suppression rather than fabrication. Discouraging dissatisfied people from writing, or resolving complaints on condition that an account is withdrawn, shifts a distribution without adding a single false entry. This leaves almost no forensic trace, since every remaining account is genuine. It is a useful reminder that an aggregate can be manipulated by subtraction, and that a suspiciously clean record deserves the same scepticism as a suspiciously enthusiastic one.

The same heuristic that makes reviews useful makes them exploitable in commercial terms. Fake reviews are now estimated to cost the global ecommerce economy tens of billions of dollars annually. Review farms, paid testimonial networks, incentivized review programs, and AI-generated review text have all proliferated. Major platforms invest heavily in detection systems, but the contest between fraud and detection is real and ongoing. A review system that has no fraud problem is one whose marketplace is too small to be worth manipulating.

Negative bias is the other failure mode. Psychological research consistently shows that humans weigh negative information more heavily than positive information of equivalent magnitude. A single one-star review with vivid details about a defect can outweigh a hundred five-star reviews in a buyer’s mental ledger. Sellers know this and respond accordingly, sometimes legitimately by addressing the underlying defect, and sometimes less honestly by trying to suppress or counter-balance the review with strategic positivity.

The deeper issue is that a single visible score becomes a target, and a measure under optimisation stops measuring what it did. Once a number drives revenue, effort flows toward moving the number, which includes legitimate improvement and also soliciting selectively, timing requests after good moments, and discouraging complaint. All of these raise the score; only one raises quality. A reader cannot distinguish them from the aggregate alone, which is the strongest practical argument for reading the distribution, the recent accounts and the specific objections rather than the headline figure.

Rating behaviour varies systematically between populations in ways that have nothing to do with the thing being rated. Some groups treat the top of a scale as reserved for the exceptional and the middle as genuine approval; others treat anything short of the top as a complaint. Research on survey response styles finds persistent differences in willingness to use extremes at all, with some populations clustering toward the centre and others toward the ends. A score aggregated across a mixed audience is a blend of these habits as much as of any underlying experience.

Language adds a second layer. Politeness conventions determine how directly dissatisfaction is expressed, so an account that reads as mild in one idiom may be a serious complaint in another, and sentiment tools trained mostly on one language inherit that bias. The practical effect is that automated summaries of multilingual corpora can systematically understate problems reported in the more indirect idioms, which is a quiet and rarely examined source of error.

The useful discipline is to compare within a population rather than across populations. Whether an item scores above or below its local peers is informative; whether its absolute number exceeds something measured elsewhere is often not. This applies to any aggregate read across regions, and it is a reason to be sceptical of league tables that rank items drawn from different audiences on a single scale without saying so.

Enforcement has shifted in the past decade from occasional action against the most flagrant operators to explicit rules about disclosure and fabrication. The common elements across jurisdictions are consistent: an undisclosed material connection between a reviewer and a seller is treated as deceptive, trading in fabricated accounts is treated as unlawful rather than merely unsporting, and suppressing genuine negative accounts while displaying positive ones is treated as a misrepresentation in its own right. The last of those is the most significant, because it names the subtraction problem described earlier.

A second strand concerns the aggregate rather than the individual account. If a displayed score is the product of a filtering or weighting policy, regulators increasingly regard a description of it as material information, on the straightforward argument that a shopper reading a number assumes it summarises everyone who spoke. Requirements to state whether accounts are verified, whether any were removed and how the figure is computed follow from that argument, and they are the reason more platforms now publish a methodology note.

Enforcement remains the weak point. Operations selling fabricated accounts are cheap to start, frequently sit outside the jurisdiction of the authority concerned, and reconstitute quickly after action. The realistic effect of regulation is therefore less deterrence than a shifting of obligation onto the platforms, which have the data to detect coordination and now bear a documented duty to try. For a reader, the practical consequence is modest but real: a published methodology is something to look for, and its absence is informative.

For buyers, the practical lessons distill to a few habits. Look at the distribution, not just the average. Read the most recent reviews, not the most popular ones. Check whether reviews are tagged as verified buyers. Pay particular attention to the median negative review (what one-star reviewers consistently complain about) because those criticisms are usually accurate, even when they are minor. Be wary of products with few reviews or sudden spikes in five-star activity. And remember that a four-star average is often more reliable than a five-star average, because no real product is perfect, and the absence of any negative reviews is itself a warning sign.

For Cosmo Strategy Guides readers, the same logic applies to gaming purchases. Steam reviews, Metacritic scores, and Reddit threads all draw on the same psychology. A game with mixed reviews and a strong fan base is often more interesting than a game with universal praise and a small audience. Reading reviews well is a skill, and it pays dividends across every kind of buying decision.

A compact procedure follows from everything above. Look at the distribution first and treat a polarised shape as a warning that the average is meaningless. Check the volume and the dates, since a thin or stale corpus describes a different product than the one on sale. Read the middle ratings rather than the extremes, because they are where people weigh competing considerations. Then search for repetition of a specific concrete fault, which is the one pattern that fabrication and suppression both struggle to produce. That sequence takes about a minute and extracts most of the information the aggregate throws away.