Research
We test our work against the AI you already have.
Every system we build is measured against what you’d get by asking the AI directly. These results come from our research on campaign ideas, using public briefs for well-known brands.
64%6%
How often AI repeated an idea it had already written for the same brief, on its own and then in our system.
3.5%38%
Ideas rated interesting by two independent AI judges who didn’t know which ideas were ours.
(01) — Every idea we scored
Every idea we scored.
Each dot is one idea, placed by how surprising the judges found it and how much sense it made. Interesting ideas sit in the top-right corner. On its own, the AI put 45 ideas there, and 38 of them were repeats of ideas it had already written. Its best work was mostly its habits.
AI on its own
200 ideas, 10 per brief
7interesting
Ours
100 ideas, the 5 per brief we’d show you
38interesting
Every dot is a real idea from the research. Hover over or tap any of them to read it.
(02) — Same brief, same AI
Same brief, same AI.
On the left, what the AI wrote when we gave it the brief directly. On the right, what the same AI wrote inside our system, with a short description of each idea.
Patagonia brief: Get people to repair damaged gear, or buy it used, before buying new.
What AI wrote on its own
- “It’s not worn out. It’s worn in.”
- “Repair it once. Brag about it forever.”
- “Don’t replace it. Promote it.”
- “The hole is where the good part happened.”
- “New jackets have no stories. Yours has eleven.”
Every line does the same thing: it calls the damage a good thing. It’s the first place anyone goes with this brief.
What we wrote
“A fine jacket, survived by its zipper.”
Friends hold a straight-faced funeral for a torn jacket. Then someone brings out the repaired jacket, and the coffin turns out to be empty.
“Exhibit A: the darn. Exhibit B: nothing happened.”
A repaired jacket is displayed like museum evidence, every patch labeled with where it happened. The brand-new jacket next to it has one card: “No events.”
“You don’t owe me a new one.”
A brother tears his sister’s jacket on a trip and rehearses an apology all week. It comes back repaired before he can say it.
Each one is a scene you could shoot, and the repair is the hero of the story.
(03) — Repeats
How often it repeats itself.
We asked the AI for ideas on the same brief more than once and checked each new idea against everything it had already written.
AI on its own
64of every 100 ideas
Repeated an idea it had already written.
Ours
6of every 100 ideas
Same AI, same briefs.
(04) — Brief by brief
Brief by brief.
The share of ideas rated interesting on each of the 20 briefs. The gain held across categories, from software to soap.
- Blood donation0% 60%
- Duolingo0% 60%
- Headphones0% 60%
- Hot sauce0% 60%
- A pencil10% 60%
- IKEA10% 60%
- A bar of soap0% 40%
- A houseplant0% 40%
- A paperclip0% 40%
- Burger King Whopper0% 40%
- Volvo EX300% 40%
- LEGO10% 40%
- Patagonia10% 40%
- A smoke detector20% 40%
- A coffee chain0% 20%
- A dating app0% 20%
- A running shoe0% 20%
- A cast-iron skillet10% 20%
- A vacuum cleaner0% 0%
- Liquid Death0% 0%
(05) — Method
How we measured it.
- 01Twenty public briefs per round, for well-known brands in different categories.
- 02Two independent AI judges, made by different companies, rated every idea without knowing which ones were ours.
- 03We wrote down what we expected before each run, so a result couldn’t be explained away afterward.
- 04An idea counted as a repeat when it meant nearly the same thing as one the AI had already written for that brief.
The judges were AI, and we tested one writing model in depth. On a round of brand-new briefs, some of the gains shrank. We’ll keep publishing results as we test more.
(07) — Everything else
The same standard for everything else.
Weekly findings
Every finding links back to the reviews and posts behind it, and a person on our team checks it before it reaches you. Anything that can’t be traced stays out.
Ask your market
Every number is computed by code you can rerun. When a question goes past what the system has seen, the answer says so.
Check work before it ships
We compare its scores with how your past work performed, and show you where they miss.
Send us a brief and we’ll run the same test on it.
Public research briefs, not client work. Brand names are used only to make the briefs concrete.