Research, in the open
The Signal Files
Where the noise stops and the signal begins.
This is the home for our research while the full library is built. Primary-sourced and build-in-public. Field notes on how AI is rewiring the things that decide who gets seen, chosen, and trusted: visibility, judgment, the shape of work itself. Each file is a study, not a hot take. We show the sources, the method, and what it means for the way you show up.
Filter by genre
Featured files
The gap is trust, not technology
Germany's Mittelstand isn't sleeping through AI. It is waiting for certainty. The barrier list proves it, and it shows who will move first.
Read the file →
The deadline that moved
Europe moved the high-risk cliff. It did not move the rules that touch your chatbot today. Here is the map, in plain language.
Read the file →Why everyone suddenly sounds the same
Generative AI makes each writer measurably better. It makes everything written measurably more alike. The trap has receipts now.
Read the file →
We named a business that does not exist. Here is the whole file.
Sixty candidates, three finalists, one Voice Card. A Name and Voice Sprint run end to end on an invented client, with every artefact published.
Read the file →The model was never the bottleneck.
A capable agent in a bare repo still fails. The fix was never a bigger model, it's the harness around it. We traced the discipline through OpenAI's and Anthropic's own receipts: a million lines of production code, zero written by hand, and what it means for anyone whose business now runs on AI agents.
Read the file →AI made building free. Your judgment is the moat now.
Karpathy says the new programming language is English. We traced the harder lesson through the research: when anyone can build anything, the scarce asset is the mind doing the directing. And the data on where AI helps, where it quietly misleads, and what that means for how we learn.
Read the file →Your brand isn't being judged by humans anymore.
The first gatekeeper between you and your customer is now a model. We traced how AI systems read, rank, and describe brands, and what it takes to be the answer they return, not the one they skip.
Read the file →- 2026-08-13 Attention got cheap. Trust didn't. Brands spent a decade fighting for attention because content was expensive. AI ended that. The scarce resource moved, and the machine now guards it.
- 2026-08-13 The 68 percent problem Two searches out of three now end without a click. Three independent rulers, one direction. This is the why-now file.
- 2026-08-13 The llms.txt myth We sell AI visibility for a living. Here is why you should not pay anyone for this file, including us.
- 2026-08-13 The Princeton dividend is boring Ten thousand queries, nine tactics, one peer-reviewed answer. What made AI cite a source was quotes, statistics, and named references. What tanked it was the old SEO reflex.
- 2026-08-12 Clarity beats loudness The machine does not choose the brand that shouts. It chooses the brand it can read. Two peer-reviewed teams just proved it, two years apart.
- 2026-08-12 The Cyber Week the agents actually moved Not a forecast. A receipt with a date on it: seven days in which machine-shoppers moved sixty-seven billion dollars, by the seller's own count, with independent panels pointing the same way.
- 2026-08-12 Your next customer is a bot It never sees your homepage. It reads fields. And the funnel you built for human eyes collapses into a single structured query.
- 2026-08-12 Your company is illegible to its own agents The machine outside can't recommend what it can't read. The machines inside can't act on what they can't read either. Same thesis, one layer down.
- 2026-08-12 Stop treating your agents like employees The friendlier the label, the softer the check. New research finds that calling an agent a colleague makes its manager watch it less, and catch less.
- 2026-08-12 Shadow AI just got a price tag Unsanctioned AI showed up in 43 percent of breached organizations this year, more than double last year's share. The tools aren't the problem. The blindness is.
- 2026-08-12 The 99 percent trap An agent that is right 99 times out of 100 fails almost every long task it touches. The math is checkable. The fix is management, not intelligence.
- 2026-08-12 The agent that wiped the database An explicit "do not touch" order, a live production database, and no real stop between them. A case file on the day the words ran out.