Directional · one question · 5 runs each · 23 October 2026
Same question, five times.
Ask a plain AI the same question five times and you get the same guess five times. Ask Outweigh and you hear different people, and the real spread of opinion. That is what a stress test for a decision should show you.
The test
"An Irish electricity supplier plans to raise its standard residential unit rate by 8% from January. How will residential customers react, and should it go ahead?"
Audience: Irish residential electricity bill-payers. The same wording, run one at a time, five times in each of three ways. No run failed and none was dropped.
- Outweigh today
20 separate perspectives, rebuilt each run - Same AI, asked directly
Gemini 3.8 Flash, one plain prompt - OpenAI model, asked directly
GPT-6 Astra, same plain prompt
The ChatGPT app itself can't be scripted, so the third arm is OpenAI's model through the same AI service.
Negative share, run by run
The plain prompts repeat one guess. Outweigh rebuilds its audience every run.
Outweigh today
20 separate perspectives, rebuilt each run
Same AI, asked directly
Gemini 3.8 Flash, one plain prompt
OpenAI model, asked directly
GPT-6 Astra, same plain prompt
40 people, or one guess?
Press "Ask again" and watch what changes.
The plain prompts were asked to imagine 40 people. Their counts came back at 31–32 negative out of 40 nearly every time. Real groups don't agree that neatly. Outweigh's perspectives move, especially in the middle ground.
Run 1 of 5
Plain AI prompt · "imagine 40 people"
The same pattern every time you ask. One guess, repeated.
Outweigh · 20 separate perspectives
A different mix each run, like asking a fresh group of people.
Where all three agree
Same direction. Outweigh shows more of the middle.
- Strong resistance in every run of every method.
- Almost nobody positive.
- The top reason everywhere: January timing, during peak winter bills.
Outweigh reached the same conclusion as the plain prompts, but it was less extreme (64% negative on average, against 74–78%) and found far more people undecided or indifferent (5–30%). Those are the people a pricing team can still win.
Full results
| Outweigh today | Same AI, asked directly | OpenAI model, asked directly | |
|---|---|---|---|
| Negative, lowest–highest | 55–75% | 77.5–80% | 72.5–75% |
| Negative, average | 64% | 78% | 74.5% |
| Run-to-run spread | 8.0 points | 1.0 point | 1.0 point |
| Mixed | 15–20% | 12.5–15% | 17.5–20% |
| Indifferent | 5–30% | 7.5–10% | 5% |
| Positive | 0% every run | 0% every run | 2.5% every run |
| Verdict | Not given (Quick check gives no go / don't go) | "Don't go" 4 of 5, "conditional" 1 of 5 | "Conditional" 5 of 5 |
| Time per run | about 110 s | about 8 s | about 8 s |
| People per run | 20 separate perspectives | 40 imagined in one reply | 40 imagined in one reply |
What Outweigh adds that a chat window doesn't
Separate perspectives
Each person is built and asked on their own, so you see disagreement instead of one averaged voice.
Official figures with sources
Prices, incomes and household data come from named official sources you can open.
Objections you can trace
Every reason is tied to who raised it and what evidence it rests on.
A record you can check later
What you expected is sealed before launch, then compared with what actually happened.
When to use it
Use Outweigh to
- Stress-test a price change or launch before it goes out
- Find the objections and who resists
- Spot the undecided middle you can still win
- Prepare for a board with evidence you can show
Don't use it as
- A headline survey figure
- A market-share forecast
- A replacement for a real poll
If you need one number, run it several times and quote the range.
Limits
- Five runs per method on one question. This shows how much answers move, not which answer is right.
- There is no real survey result to compare against.
- The plain prompts imagined 40 people; Outweigh used its usual 20. Shares are compared, not counts.
- A single Outweigh run varied by up to 20 points, so one run alone is not a figure to quote.
- Outweigh's perspectives are synthetic, not real respondents. Results are directional decision support.