The Trust Stack: How Buyers Decided Whom to Believe, from the Maker's Mark to the Machine

Trust as a substitute for knowledge

A buyer almost never has the information a transaction actually requires. A shopper cannot assay the silver in a ring, run a compound interest calculation on a lender's real terms, or independently test whether a space heater will catch fire. The economist George Akerlof formalized what this gap does to a market in a 1970 paper built around the market for used cars, showing that when a seller knows more about a good's quality than a buyer does, buyers rationally discount every offer to protect against the worst case, good sellers exit rather than accept that discount, and the market can degrade toward the low-quality goods economists nicknamed lemons. Akerlof shared the 2001 Nobel Memorial Prize in Economic Sciences with Michael Spence and Joseph Stiglitz for extending this insight, and it explains why a trust signal is not a courtesy a seller extends to a buyer. It is closer to a repair for a market that would otherwise fail.

Communication researchers studying persuasion arrived at a compatible answer from a different direction. Carl Hovland's Yale program, formalized with Walter Weiss in a 1951 paper, found that a message's persuasive power tracked two properties of its source: expertise, the perceived competence to know the truth, and trustworthiness, the perceived absence of a motive to distort it. A stranger's claim about a stranger's product supplies neither on its own, which is exactly the vacuum a trust signal is built to fill. Three decades later, Richard Petty and John Cacioppo's 1986 Elaboration Likelihood Model explained the mechanism a buyer uses to act on a signal without redoing the underlying analysis. When a buyer has the motivation and the ability to scrutinize a claim closely, in Petty and Cacioppo's terms the central route, they weigh the argument's substance. When they do not, the far more common condition for an ordinary purchase, they fall back on the peripheral route: a badge, a star count, or a stranger's five-word review stands in for the analysis itself.

Read together, these two literatures describe the same object from economics and from psychology. A trust signal exists because information is unevenly distributed, and it works because a buyer can act on a compressed cue instead of reconstructing the missing knowledge firsthand. What follows is a history of which cues a market has trusted, in what order, and what happened once each cue became valuable enough to counterfeit.

The maker's mark and the guild's guarantee

The earliest documented version of this repair in English commercial life predates the word "brand" by centuries. In 1300, a statute of Edward I ordered that gold and silver be struck to a defined standard and required "les Gardeins du Mester," the Guardians of the Craft, to test each piece and stamp compliant work with a leopard's head. The statute survives in language precise enough to read as a modern consumer protection clause: no goldsmith was to make, or cause to be made, any vessel, jewel, or other item of gold or silver except of the true alloy. The Goldsmiths' Company, the London guild responsible for enforcing it, received a royal charter in 1327 formalizing its authority over the trade and its right to test and mark work above provincial goldsmiths' own claims. By 1478, following a further statute of Edward IV making the Company answerable for substandard wares, it had established a dedicated assay office at Goldsmiths' Hall in London, staffed by a full time assayer, and it is from that hall that the word "hallmark" descends.

The system did real, measurable work for a buyer who could not run a metallurgical test personally. Sterling silver was fixed at 92.5 percent pure and gold at the "touch of Paris," roughly 19.2 carats, and a buyer anywhere in England could treat the stamp as equivalent to having verified the metal directly. What made the mark trustworthy was not the symbol itself but the enforcement standing behind it: a false or counterfeit mark was prosecuted as fraud, and the guild's own commercial interest in the mark's credibility gave it a reason to police its members. In Hovland's later vocabulary, the leopard's head supplied expertise, since the Guardians actually tested the metal, and a maker's own personal stamp, which later medieval practice required alongside the guild mark, supplied trustworthiness by attaching a named individual's reputation to a specific object, seven centuries before either term entered the psychological literature.

The invention of the portable brand

Industrialization broke the condition the guild mark depended on: a buyer who might reasonably know, or know of, the individual maker. A factory in Ravenna, Ohio, or Cincinnati, selling oats or soap into a national market, could not rely on personal acquaintance, and manufacturers reached for the modern registered trademark, a mark whose entire function was to travel with a product to a buyer the maker would never meet. Britain's Trade Marks Registration Act 1875 created the first centralized, deliberately national register built for exactly this purpose, taking effect on 1 January 1876. A representative of Bass Brewery is said to have camped outside the registry's doors overnight to be first through them, and the brewery's red triangle mark is conventionally remembered as the United Kingdom's first registered trademark, the mark the UK Intellectual Property Office now honors as Trade Mark No. 1. The record adds a useful nuance worth keeping in a study built on precision: the surviving ledger actually numbers the Bass marks 914 through 916 among that day's filings, and the "No. 1" status reflects that Bass was first through the door, not first in the registry's own numbering sequence.

The United States saw a comparable shift in the same decade. Henry Seymour and William Heston's Quaker Mill Company registered the Quaker name on 4 September 1877, after Seymour concluded that the Religious Society of Friends carried associations of integrity, honesty, and purity worth attaching to a breakfast cereal, one of the earliest branded packaged foods sold under a registered mark of its kind. Two years later, Procter & Gamble's Harley Procter and James Gamble launched Ivory Soap with the slogan "99 44/100% Pure," turning a purity claim, the same category of claim a medieval hallmark once certified, into a nationally advertised, ownable line a competitor could not simply copy.

What a registered trademark did structurally was extend the guild mark's logic across distance and time. A buyer in a city the maker had never visited could extend the same trust to a bar of soap that a neighbor extended to a known local maker, on the promise that the identical name stood behind every unit, indefinitely, now backed by law rather than only by guild custom. In Hovland's terms, the brand manufactured trustworthiness at a scale personal acquaintance could never reach, provided the promise of consistency actually held over time.

The third-party certifiers

Between a personal referral and an impersonal brand, a third kind of institution emerged, one that tested a product itself and staked its own name on the verdict, independent of the maker's interest in the sale. Underwriters Laboratories was founded on 24 March 1894 by the electrical engineer William Henry Merrill, a direct outgrowth of the electrical safety inspection work he had carried out at the 1893 World's Columbian Exposition in Chicago, and the organization published its first written safety standard in 1903. The UL mark answers a narrow, testable question, whether a specific product met a specific safety standard, in the same structural form the hallmark used four centuries earlier.

Good Housekeeping took the same idea into the home. The magazine's Institute launched its Seal of Approval in December 1909, testing and evaluating products before allowing them to carry it, and by the end of the seal's first year nearly 200 products had qualified. Harvey Washington Wiley, who had led the federal Bureau of Chemistry and driven passage of the 1906 Pure Food and Drug Act as the country's first chief food and drug regulator, joined the Institute in 1912 and lent it a regulator's credibility rather than only an advertiser's. The seal still carries a limited warranty today, a replacement or refund of up to $2,000 if a sealed product proves defective within two years, a guarantee that converts the magazine's own money into evidence of its confidence.

Michelin arrived at independence by an odder route. André and Édouard Michelin distributed the first Michelin Guide free in 1900, roughly 35,000 copies, explicitly to encourage driving and therefore tire wear, a strategy self-interested from the start. The guide only became a credible arbiter once the brothers stopped giving it away. After André reportedly saw a stack of guides propping up a garage workbench, the company began charging seven francs for a new edition in 1920, on the reasoning the company has since attributed to him, that "man only truly respects what he pays for." The star ratings that made the guide a genuine independent authority followed in 1926, reaching the modern one to three star system by 1931. The sequence says something about credibility in general: a recommendation given away for free advertised the giver's own interest, and only once a customer paid for the verdict directly did the verdict start to read as independent of the businesses it rated.

The reviews revolution, where the crowd became the certifier

Amazon's decision in 1995 to let ordinary customers post reviews of the products the company itself sold looked, at the time, like a mistake. Letting a customer publish a negative opinion next to a listing the retailer wanted sold struck many observers as an odd thing to allow, let alone build a feature around. Three decades later, Amazon has said its stores draw roughly 125 million reviewing customers in a year, contributing nearly 1.5 billion reviews and ratings, on the order of 45 new reviews every second, and the company now treats the absence of reviews on a listing as close to a defect. Yelp, founded in 2004 by Jeremy Stoppelman and Russel Simmons after Stoppelman's own trouble finding a trustworthy local doctor, and TripAdvisor, launched in February 2000, generalized the same mechanic to local services and travel, categories that had never before come with a searchable public record of what previous buyers actually experienced.

By the 2010s, researchers could measure what a review was actually worth in revenue rather than only in sentiment. Michael Luca's Harvard Business School study paired Yelp ratings for Seattle restaurants with the Washington State Department of Revenue's own sales tax records, using the way Yelp rounds ratings to isolate the rating's effect from a restaurant's underlying quality, and found that a one star increase in Yelp rating was associated with a 5 to 9 percent increase in revenue. Crucially, Luca found the effect concentrated entirely among independent restaurants; chains, which already carried a national brand's own trust signal, showed none of it, a finding that lines up precisely with the logic of the sections above. A review supplies the trust a brand already supplies to a chain, which is why it does the most economic work for the business that has no other signal to offer.

Northwestern's Spiegel Research Center, working with PowerReviews across several online retailers, found that the likelihood of purchase for a product carrying five reviews was 270 percent higher than for an identical product with none, an effect that grew to a 380 percent increase in conversion for higher priced goods specifically. The same research found something a purely additive model of trust would not predict: purchase likelihood peaked for products rated between 4.0 and 4.7 stars and declined as the average approached a perfect 5.0, suggesting buyers read a flawless score not as stronger evidence but as weaker evidence, closer to a manufactured number than an aggregated real experience.

Verification, and the fraud it invited

Once a rating could move 5 to 9 percent of a restaurant's revenue, it became worth faking, and every major platform answered with some form of identity verification layered on top of the review itself. Amazon began flagging reviews from confirmed purchasers as "Verified Purchase," giving a reviewer's claim a second, harder to fake signal behind it, an actual transaction record. Twitter introduced its blue checkmark in 2009 for a narrower reason, to confirm that an account claiming to represent a public figure or organization actually belonged to them, after the retired baseball manager Tony La Russa threatened suit over an impersonating parody account. Google Business Profile verification moved, industry trackers reported through 2025, toward relying primarily on a short video walkthrough of a business's physical premises for most new listings rather than the mailed postcard the process once depended on, a shift toward a harder to fake, physical world check once a purely digital signal had proven too easy to manufacture at scale.

Twitter's checkmark is the cleanest illustration in this history of what happens when a verification signal is repriced without being redesigned. When new ownership converted the badge from an identity check into a feature of an eight dollar a month subscription in late 2022, removing the requirement to actually confirm who an applicant was, and then stripped badges from accounts that had earned them under the old system in April 2023, the mark did not become worthless so much as it began answering a different question than the one readers still assumed it answered. It no longer meant this account belongs to who it claims to be. It meant this account is paying eight dollars a month. A signal's credibility, this case shows plainly, depends entirely on what it actually verifies, not on how familiar its shape remains to the person reading it.

The financial incentive behind manufacturing trust signals is large enough to sustain an entire counter economy. A World Economic Forum analysis estimated that fake online reviews cost consumers $152 billion globally in 2021, the same year the organization estimated reviews influenced $3.8 trillion of global e-commerce spending, a modeled estimate this study treats as indicative of scale rather than as a precise, independently audited figure. Amazon's own account of fighting that economy is considerably more granular: the company has said it proactively stopped more than 250 million suspected fake reviews in 2023 alone, brought legal action against more than 150 alleged fake review brokers across the United States, China, and Europe that year, and had taken down 75 fake review broker websites by June 2024, backed by a stated 15,000 employees working on fraud and abuse globally, while maintaining, without independent verification available to this study, that more than 99 percent of the products customers view carry only authentic reviews.

Algorithmic corroboration, the newest shortcut

An AI answer engine compresses, into a single generated sentence, most of what this history describes as separate stages: it can draw on a brand's name recognition, the volume and tenor of a business's reviews, any certifications or citations that mention it, and a general sense of its reputation across the open web, then deliver a short list of businesses it will name as trustworthy, typically without showing the buyer the underlying evidence trail that produced the answer. Adoption of the tools capable of doing this has moved quickly. A Pew Research Center survey fielded in February 2026 found that 49 percent of US adults had used an AI chatbot such as ChatGPT, Gemini, or Copilot, with 44 percent specifically having used ChatGPT, more than double the 18 percent Pew recorded in 2023, about a quarter of adults using a chatbot at least daily, and 60 percent saying they now read the AI summary that appears above traditional results in a search engine.

The behavior change shows up in commercial data, not only in survey answers. Adobe Analytics, tracking more than 1 trillion visits to US retail websites, found that traffic referred by an AI tool to those sites grew 138 percent year over year in May 2026, roughly fourteen times its October 2024 level, and that this AI-referred traffic converted to a purchase 54 percent better than traffic arriving from non-AI sources, with visitors spending 53 percent more time on a retailer's site and viewing 23 percent more pages once they arrived through an AI referral. However a buyer arrives at an AI engine's recommendation, the engine's answer is, by this measure, doing more to move that buyer toward a purchase than a conventional search result currently does.

A discipline for earning that recommendation has already formed around a logic that echoes search engine optimization's own history. Researchers at Princeton and Georgia Tech named the practice Generative Engine Optimization in a 2023 paper, and their benchmark testing found that specific, deliberate changes to a source's content, adding citations, statistics, and clearer structure, could raise its visibility inside a generative engine's answers by up to 40 percent. The finding treats an AI-assembled answer as a new surface a business can be present on or absent from, structurally the same problem a business faced when it needed a registered trademark to be findable in a national market, or enough reviews to clear a buyer's star rating floor, solved this time by making content legible to a model rather than only to a person.

The verdict: what does not change, and what does

Every signal this history traces was built to solve the problem Akerlof described in 1970: a buyer cannot verify a seller's claim directly, so the market needs a cue that substitutes for that verification. And every signal, once it became valuable, was eventually formalized into law (the hallmark statute, the trademark register, the FTC's 2024 rule), professionalized into an independent institution (Underwriters Laboratories, Good Housekeeping, Michelin, Consumer Reports), or counterfeited at a scale large enough to force a new layer of verification on top of it (fake reviews, purchased checkmarks, manufactured engagement). None of the earlier signals actually disappeared once a newer one arrived. A buyer today can still recognize something like a maker's mark in a brand's logo, read something like a modern guild's testimony in a certification badge, hear word of mouth digitized as a review, and now ask a machine to compress all of it into a single sentence, which is the reason this study treats the whole apparatus as a stack that accumulates rather than a sequence of signals that simply replace one another.

The newest layer has a real limitation: nothing yet plays the Assay Office's role for it. Sterling silver either is or is not 92.5 percent pure, and a leopard's head could be checked against that fixed standard; a UL mark or a Good Housekeeping Seal rested on a documented test a buyer could, at least in principle, ask to see; the FTC's 2024 rule gives a regulator explicit statutory language to act on a fabricated review. An AI engine's judgment that one business deserves to be named as trustworthy and another does not currently rests on a private, largely undisclosed weighting of the same signals this history describes, tested by outside researchers so far only informally and, as the Tow Center's study shows, only for a narrower question, citation accuracy for news articles, than the trust question this study is actually about, which businesses an engine names as reliable. That is a genuine, stated gap in the evidence, not a settled finding in either direction, and this study does not resolve it.

What can be said as projection, clearly marked as such, is narrower and more defensible. Every earlier shortcut in this arc became valuable enough to imitate, was imitated, and was eventually answered with a verification or auditing layer once the imitation grew large enough to matter to a regulator or a market. Whether a comparable layer eventually attaches to algorithmic corroboration the way one attached to online reviews after nearly three decades, or to identity verification the way Twitter's 2023 collapse showed a signal can lose its meaning without losing its familiar shape, is not yet known. It is the open question this study leaves standing: at this point in 2026, it has not been decided yet.