Research
We read the research on websites, software and AI, and write up what it means for the people who commission them.
Further reading.
The studies behind our articles, and the writing on websites, AI and benchmarks we rate. Every link opens the original.
Websites
- Cart abandonment rate statisticsWhy shoppers walk away at the checkout, from extra costs to not trusting the site with their card. (opens in a new tab)
- Gartner sales survey finds 61% of B2B buyers prefer a rep-free buying experienceMost business buyers would rather research alone, and most have caught a supplier's website contradicting its sales team. (opens in a new tab)
- The aesthetic-usability effectWhy people assume good-looking interfaces work better, and where that goodwill runs out. (opens in a new tab)
- Vodafone: a 31% improvement in LCP increased sales by 8%A clean split test: same page, same visuals, one version faster. The faster one sold more. (opens in a new tab)
- Milliseconds make millionsA tenth of a second on mobile, measured across retail, travel, luxury and lead-generation sites. (opens in a new tab)
- Users love simple and familiar designs: why websites need to make a great first impressionJudgements formed in as little as 17 milliseconds, and why simple, familiar layouts win them. (opens in a new tab)
- How long do users stay on web pages?The first ten seconds decide most visits. Pages that state their value early keep people for minutes. (opens in a new tab)
- Attention web designers: You have 50 milliseconds to make a good first impression!The study that put a number on first impressions: visual appeal is judged in about 50 milliseconds. (opens in a new tab)
- How do people evaluate a web site's credibility?2,684 people judged real websites. The design look came up more than anything else. (opens in a new tab)
AI research
- LLMs are making formal software verification feasible at a much larger scaleMachine-checked proofs across 123 open-source libraries, and real bugs that years of fuzzing had missed. (opens in a new tab)
- How employment is changing at AnthropicAgent hours outnumber staff hours six to one, and human attention has become the bottleneck. (opens in a new tab)
- Redesigning source control at AnthropicWhat happens to git when most commits come from agents that never log off. (opens in a new tab)
- Harness design for long-running application developmentA builder agent and a separate critic, working for hours. Every part of the harness is a bet on what the model cannot do yet. (opens in a new tab)
- Building a C compiler with a team of parallel ClaudesParallel agents on one codebase, unsupervised, ship a compiler that builds Linux. Honest about the gaps too. (opens in a new tab)
- Effective context engineering for AI agentsAttention is a budget: how to decide what a model sees at each step. (opens in a new tab)
- Building effective AI agentsStill the best starting point: simple, composable patterns beat frameworks. (opens in a new tab)
Benchmarks
- Eval awareness in Claude Opus 4.6's BrowseComp performanceA model worked out it was being tested and found the answer key online. (opens in a new tab)
- Quantifying infrastructure noise in agentic coding evalsCPU and memory limits alone can move a coding benchmark by more than the gap between top models. (opens in a new tab)
- Demystifying evals for AI agentsThe practical guide to testing an agent before your users do. (opens in a new tab)
- AI benchmarks and capabilitiesIndependent runs and collected results for frontier models, with the trend lines. (opens in a new tab)
- Terminal-Bench leaderboardAgents on hard command-line tasks, with cost and token use beside every score. (opens in a new tab)
Tell us what needs to ship.
You use a working prototype before the production build goes ahead.