Fractl's research on AI in marketing work.

Five more studies, newest first.

Each study links to the full report, with the methods, supporting files and corrections.

1. An AI has about six ideas. Borrowed angles doubled them.

   Different ideas come from a deliberately different starting approach.

   ~3,100 AI answers

   Ideas

   Sep 2026

   Read the full study

   How we studied it, and the limits

   The question: Do outside perspectives give AI different ideas?

   The key result: On one puzzle, twenty borrowed angles found 38 distinct ideas. Angles the AI invented for itself found 19.

   We tested whether ways of thinking from other fields widen AI's range of explanations. A large set of borrowed approaches produced more distinct ideas than asking AI to invent perspectives.

   About 3,100 answers explored one biological puzzle. We also tested prompts with the borrowed approach removed, to examine whether that approach made a difference.

   Limits: The controlled task was a single scientific puzzle. The findings suggest methods to test in other work; they do not establish a universal improvement across marketing assignments.

   See how Fractl Agents works on ideas and data

2. We made 5,000 campaign ideas. Strategists kept the plain ones.

   A useful idea fits the assignment as well as sounding different.

   1,200+ human labels

   Ideas

   Aug–Sep 2026

   Read the full study

   How we studied it, and the limits

   The question: Does more variety mean more usable ideas?

   The key result: A panel of seven strategists kept 25 of 40 plain-prompt ideas, and 18 of 40 from a method built for variety.

   We tested ways to broaden campaign ideas and asked people to judge the results. Human review showed some methods added variety alongside ideas a strategist would reject.

   The research compares generated campaign ideas, automated judgments, and blind human reviews from 8 readers. The updated report includes a wider reader panel and a corrections log.

   Limits: Some findings still rely on a smaller set of briefs and reviewers. The full report distinguishes replicated findings from results still being checked.

   See how Fractl Agents works on ideas and data

3. AI recommends whoever the rest of the web is already writing about.

   Your best openings are publications in your category that mention competitors but not your brand.

   11,573 AI answers

   AI visibility

   Jul 2026

   Read the full study

   How we studied it, and the limits

   The question: Which sources do AI brand recommendations cite?

   The key result: Across 11,573 AI answers, 73% of citations came from outside Google's top ten results.

   We recorded brand recommendations and citations across 11,573 AI answers in 15 industries. The observations show differences between engines, questions, and cited sources that a single visibility score can conceal.

   The study ran commercial questions through six engines, repeated queries, and collected cited pages for analysis.

   Limits: The results describe the sampled questions and collection period. Observed relationships do not guarantee that earning a placement will cause a future recommendation.

   See how Fractl Agents works on search and AI visibility

4. You can beat the AI detectors. The facts don't survive it.

   A rewrite needs a check of meaning and facts afterwards.

   20 documents rewritten

   Writing

   Jul 2026

   Read the full study

   How we studied it, and the limits

   The question: Does rewriting AI text preserve the facts?

   The key result: One rewriter fooled both detectors on 20 of 20 documents. About 31% of the source facts survived.

   We studied AI-text detectors and rewrites meant to change what detectors decide. Aggressive rewrites exposed a tradeoff: moving a detector score could also remove or change information.

   The project compared detection methods, examined outputs from many models, and tested trained rewriters against external detectors.

   Limits: Detector results depend on the detector and evaluation material. This report does not establish who wrote an individual document or certify a detector as infallible.

   See how Fractl Agents works on content and creative

5. AI knows your brand. Naming it unprompted is a different bet.

   A full check covers what AI already seems to know about your brand and how answers change with new sources.

   6,000+ test runs

   AI visibility

   Jul 2026

   Read the full study

   How we studied it, and the limits

   The question: Where do AI brand recommendations come from?

   The key result: A chatbot recognized companies with under 1,000 training-data mentions 83% of the time, and named them unprompted 8%.

   We studied public training material and ran experiments that controlled which web pages AI could read.

   The research combines an analysis of public AI training material with more than 6,000 controlled runs across nine models.

   Limits: Public training material does not reveal everything a private AI model learned from. Supplying selected pages in an experiment differs from a consumer AI search service.

   See how Fractl Agents works on search and AI visibility