Generate three, pick one

This is the one move in this part of the book with a proper controlled experiment behind it, so it is worth reporting carefully, including the parts that do not flatter the advice.

Dow and colleagues had participants design web advertisements for a real client. One group worked in parallel: three prototypes, feedback on all three, then a final. The other worked serially: one prototype, feedback, revise, repeat. The study “held constant the number of prototypes created, the amount of feedback provided, and the overall time allotted.”

Same hours, same number of designs, same amount of critique. The only difference was whether the alternatives existed at the same time.

The parallel ads got 445.0 clicks per million against 397.9. Over the first five days, “parallel ads had 79,800 impressions with 44 clicks and serial ads had 79,658 impressions with 26 clicks.” Seven blind expert raters, four magazine editors and three advertising professionals, scored them 24.4 against 21.7 out of 50. And independent workers judging similarity found the serial ads significantly more alike than the parallel ones.

The part that changes the advice

Now the finding that makes this chapter different from the version you have read elsewhere.

The same team ran the same manipulation on a physical egg-drop task and got nothing: 5.4 feet parallel against 5.5 feet serial. And in a follow-up, participants who created three designs but showed only their favourite did no better than those who created one. The authors’ conclusion is the sentence to keep:

“Simply creating multiple designs (without feedback) led to broader exploration, but not better results. The benefits were only realized if participants shared multiple designs.”

So generating three is not the mechanism. Sharing three is. If you produce three options and quietly pick your favourite before anyone else sees them, you have done the expensive half of this move and skipped the half that works.

Why it works, on people rather than pixels

The most interesting result is not about the designs at all.

In post-task interviews, “nearly half of serial participants reported negative reactions to critique of their prototypes, while no parallel participants reported this.” Not one.

And novices in the parallel condition gained 2.9 points of self-efficacy while serial novices lost 0.73.

That is the actual mechanism, and it explains the egg-drop null too. When you have made one thing, critique of it is critique of you, and you defend. When three of your things are on the table, critique becomes comparison, and comparison is a conversation you can have without flinching. The egg drop had objective feedback, the egg either broke or it did not, so there was no ego to route around in the first place.

WHAT YOU DID                    WHAT ACTUALLY HAPPENED

 built one, showed it            "so this bit here is a bit
                                  confusing"
                                 -> you explain why it is right
                                 -> feedback becomes negotiation

 built three, showed three       "the second one is confusing but
                                  the third one solves it"
                                 -> nobody is defending anything
                                 -> feedback becomes selection

 built three, showed             identical to building one.
   your favourite                measured. no benefit.

The move

Make three genuinely different versions, show all three, and let the person you are showing them to say which parts of each are wrong.

The word doing the work is show. Three in your head does not count. Three in a branch you did not share does not count.

Doing it in 2026

Generating three alternatives used to be the expensive part, which is why almost nobody did it. It is now close to free, and that changes the economics of this move more than anything else in this part of the book.

Which puts the cost somewhere new. The scarce thing is no longer producing the options, it is having the taste to tell them apart and the discipline to show all three when one of them is obviously weaker. I have not found any study testing parallel prototyping with generated alternatives, so whether the effect survives when you did not personally make the three is genuinely open.

What it costs

Three real options is three times the work, and two of them get thrown away. That is the straightforward bill, and it is only worth paying when the question is open. Do not generate three versions of something you already know the answer to.

And the effect has a documented boundary. It did not appear on the egg-drop task, where feedback was objective and immediate. The closer your problem is to the egg either breaks or it does not, the less this move buys you. Save it for questions where the answer is a judgment.

Try this week

Next time you would normally build one screen, build three, and make them differ in something structural rather than cosmetic. Not three colour schemes. Three arrangements of the same information.

Then show all three to one person, in one message, and ask the question that makes comparison possible:

Three options. Which bits of each are wrong?

Not which do you prefer. Which bits of each are wrong invites them to take the top of one and the bottom of another, which is where the fourth and better option usually comes from.