Fractl Research

Original research on how AI actually decides what to recommend

The tools inside Fractl Agents are built on measurements, not vibes — and we publish the measurements. Every piece here ships with its data files, its pre-registrations, and its full corrections log, because we'd rather you checked us than believed us.

AI Ideation · Content strategySeptember 20266 assistants · 12 industries · 14,400 ideas · 300 Google searches

Page one already wrote your list

Ask an AI for blog ideas and the first five are the ones it gives everyone. We asked six assistants for 14,400 ideas across twelve industries, then checked them against Google. When several assistants agree on an idea, a post like it is already in the top five results eight times in ten. Switching assistants gives you a longer list, not an unwritten one. And two in five #1 results aren't articles at all. Plus the one-search habit that fixes it, and the free check that runs it for you.

Read the research →
AI Ideation · Scientific reasoningSeptember 2026~3,100 hypotheses · 13 methods · matched controls throughout

An AI has about six ideas

Not six ideas — six ways of having ideas. We asked one to invent a fresh angle on a problem seventy-nine times and it handed back the same handful, renamed. Feeding it angles borrowed from human catalogues fixes this, but only past a size we initially got wrong: three borrowed angles ran dry as fast as the AI's own six, while twenty found nearly twice as many ideas. Also here — the follow-up question that tripled the yield, and why naming what you're grading makes the answers worse.

Read the research →
AI Ideation · CreativityAugust 2026~5,000 ideas · 20+ methods · 21,000 judgments · 230 human labels

Escaping the idea basin

Ask an AI for 40 campaign ideas and you get 14 — the rest are the same ideas re-worded. We tested every published fix, from temperature to weight-surgery on a 72B open model, and had a strategist blind-judge thousands of the results. The knobs don't work, AI judges can't score originality (43–55% agreement with a human), and the two things that actually widen the swath of good ideas are a sharper brief and an ensemble of unrelated models.

Read the research →
AI Visibility · GEOJuly 202611,573 answers · 15 industries · 6 engines

What AI actually recommends

We measured every brand and every source in 11,573 AI answers across 15 industries, then joined it to what's inside the models' training data. Which sites AI trusts, why 73% of what it cites doesn't rank on Google's page one, how new brands win on live visibility instead of fame, and why getting featured on the sites AI cites — earned media — just became how you win an AI recommendation.

Read the research →
AI Detection · ContentJuly 2026326-model census · 0.97-AUROC detector · GPU-trained rewriters

Can you actually make AI writing undetectable?

We built the detector ourselves, ran a census of 326 models, and trained our own rewriters on rented GPUs. Detecting AI writing is basically solved and the "sound more human" checklists are a myth — but the one-click "humanizer" promise runs into a real wall: you can evade detection, or keep the facts, not both. A field report on what's solved, what's snake oil, and why.

Read the research →
AI Visibility · GEOJuly 20269 models · 6,000+ controlled runs · 4T training tokens

Where AI recommendations actually come from

We read a real AI model's actual training data — the only model on earth where that's possible — and ran 6,000+ controlled retrieval experiments across nine models, including a company we invented. What memory buys, what a retrieved page buys, why mention-count "AI visibility" scores measure the English language, and why the newest models are becoming auditors that discount promotional content.

Read the research →