Society
Shopping and money online
Online retail shares are measured well but not comparably between countries, while the evidence on manipulation and credit is strong on mechanism and thin on population-scale harm.
Online retail is well measured within a country and badly compared between countries. Official agencies publish reliable shares — around 29% of UK retail sales and around 17% of US retail sales in 2026 — and the two should never be set against each other, because they cover different retail universes. The evidence on how online selling works on people is stronger on mechanism than on outcome: experiments show manipulative design roughly doubles acceptance of an unwanted paid service, but nobody has measured what that does to real spending. Checkout credit is concentrated among subprime borrowers, and fraud totals count complaints rather than incidents. For a workplace-oriented example of technology measurement in practice, see this guide.
The UK and US shares measure different things
Both national series are strong. The comparison between them is not.
US retail e-commerce sales were $326.7 billion in Q1 2026, 16.9% of total US retail sales, up 9.8% year on year.
Direction: Increase. Strength of evidence: Strong.
Caveat Not adjusted for prices, so part of the growth is inflation, and the definition excludes travel, ticketing and most services.
Internet sales accounted for 29.4% of all UK retail sales in June 2026, the highest share since April 2021 and up from 28.9% in May.
Direction: Increase. Strength of evidence: Strong.
Caveat Not comparable with the US Census series — different retail universes and treatment of services — so any "UK 29% vs US 17%" contrast compares two different things.
The gap is mostly definitional. The US series treats travel, ticketing and most services differently from the ONS series, and no harmonised cross-country series exists. A contrast between them says something about statistical convention and little about shopping.
European survey evidence counts people rather than sales: 78% of EU internet users bought goods or services online in the previous 12 months in 2025, from 90% of 25-34-year-olds down to 55% of 65-74-year-olds (Eurostat, extracted February 2026). The denominator is internet users, not the population, and the survey stops at 74, so the true share is lower — another place where who is counted as online shapes a headline.
Manipulative design is common, and the count is a floor
The prevalence evidence is seven years old and machine-gathered, so it catches only patterns a crawler can read.
The crawl found 1,818 dark pattern instances across 15 types, with 183 sites using patterns designed to deceive and 22 third-party firms selling dark patterns as a turnkey service.
Direction: Increase. Strength of evidence: Strong.
Caveat Automated detection catches only machine-readable patterns, so this is a lower bound, and the crawl covers English-language sites.
The detail that survives best is the supply chain: two dozen firms selling countdown timers and fake scarcity messages as a service, which makes the practice a product rather than a design choice. The OECD concluded in 2022 that dark commercial patterns are "commonly found in online user interfaces" and cause "substantial consumer detriment", but that is a synthesis carrying no prevalence figure of its own (OECD, 2022). The FTC's 2022 staff report Bringing Dark Patterns to Light describes a rise in sophisticated patterns — an assertion about the agency's caseload, not a measured trend.
Mild manipulation works better than aggressive manipulation
The causal evidence is experimental rather than observational, and the effect sizes are large.
Mild dark patterns more than doubled acceptance of a dubious paid service, from 11.3% in the control to 25.8%, and aggressive dark patterns raised it to 41.9%.
Direction: Increase. Strength of evidence: Strong.
Caveat Hypothetical scenarios with no real money at stake; tripling the stated price had no significant effect, suggesting participants were not attending as they would to a real purchase.
The regulatory implication runs against intuition. Aggressive patterns provoked measurable anger and disengagement while mild ones did not, which makes mild ones the greater concern: they doubled acceptance with no backlash (Luguri and Strahilevitz, 2021). That backlash was self-reported attitude inside the experiment, not real switching or complaints.
Checkout credit is concentrated among weaker borrowers
Buy-now-pay-later is the hardest part of online payment to measure, because lenders generally do not report to credit bureaus. The best data came from a regulator obtaining matched records from six firms.
Roughly 61% of US buy-now-pay-later originations in 2021-22 went to borrowers with deep subprime (45%) or subprime (16%) credit scores.
Direction: Increase. Strength of evidence: Strong.
Caveat Covers only Affirm, Afterpay, Klarna, PayPal, Sezzle and Zip, and because BNPL lenders rarely report to bureaus the scores may be stale or thin.
Two further findings belong together: about 63% of BNPL borrowers held multiple simultaneous loans at some point in 2022, 33% across more than one firm, and BNPL borrowers defaulted on their credit cards at around 10% during 2019-2022 against roughly 2% on the BNPL loans themselves (CFPB, 2025). The second gap is correlational — BNPL users are riskier borrowers before they borrow.
Survey data describes the market differently. 15% of US adults reported using BNPL in the previous year, up from 10% in 2021, rising to 25% of Black adults and 21% of Hispanic adults against 11% of White adults, and 24% reported a late payment (Federal Reserve Board, 2025).
Fraud totals count reports, not incidents
The largest fraud numbers come from complaint databases: useful as floors, misleading as trends.
Fraud initiated via websites or apps produced $1.858 billion in reported losses from 186,826 reports in 2024, more than social media at $976 million from 148,288 reports.
Direction: Increase. Strength of evidence: Mixed.
Caveat Contact method is recorded only when the consumer supplies it, so totals understate and the ranking between channels depends on who reports.
US consumers reported more than $12.5 billion in fraud losses in 2024, at a median of $497 (FTC, 2025). Most fraud is never reported, by the agency's own account, so "scam losses rose X%" may be measuring reporting behaviour; what is known about which defences stop a scam rests on different evidence.
The short version
- Internet sales were 29.4% of UK retail in June 2026 (ONS) and 16.9% of US retail in Q1 2026 (US Census Bureau), measuring different retail universes.
- A crawl of about 11,000 shopping sites found 1,818 dark pattern instances and 22 firms selling such patterns as a service, a lower bound because only machine-readable ones were counted (Mathur and colleagues, 2019).
- Mild manipulative design more than doubled acceptance of an unwanted paid service in randomised experiments, from 11.3% to 25.8%, with no measurable backlash (Luguri and Strahilevitz, 2021).
- About 61% of US buy-now-pay-later originations in 2021-22 went to subprime or deep subprime borrowers (CFPB, 2025).
- The harm evidence has a ceiling: dark-pattern effects come from hypothetical online panels, BNPL data covers six lenders, and fraud totals count reports rather than incidents.