New Coke vs. the Domino's turnaround: we replayed both recipe changes blind with synthetic consumers

Coca-Cola and Domino's made the same move. Each one threw out its recipe and told the public. In 1985 the New Coke reformulation was such a disaster that the original was back on shelves 79 days later. In 2010 the Domino's pizza turnaround became one of the best-known comebacks in fast food. Same move, opposite endings.

So we asked a narrow question. If you put both decisions to a population of synthetic consumers, blind, locked in the year the decision was made and with no way to look up how it ended, can they tell the two apart? These are the research notes from that test: what we asked, what the synthetic town answered, what really happened, and where the evidence is weak.

Coca-Cola had been losing ground to Pepsi, and Pepsi's blind sip tests showed people preferring a sweeter cola. So Coca-Cola retired its 99-year-old formula and replaced it with a sweeter one. The catch: in blind sip tests a majority, roughly 53 to 57%, really did prefer the new formula. New Coke did not fail on taste. It failed because people felt their Coke was being taken away.

Domino's had the opposite problem. Its pizza was the thing people complained about. So it said so, in its own ads, quoting real focus-group verdicts, and rebuilt the recipe from the crust up.

Each case went through two engines. TownFolk polled 200 synthetic consumers, each with an age, income, region, job and personality, in eight groups of 25. townSquare put five synthetic stakeholders in a room for twelve speaking turns and asked for a verdict: go, no-go or revise.

The test only means something if nobody can peek at the ending. These were the rules:

Every case also carried three checks. A trick question (an A/A placebo): two groups of 60 personas got the exact same neutral choice, and a gap of 8 points or more would void the case, because it would mean the crowd finds differences that are not there. A recognition probe: afterwards, 20 personas were asked whether the exercise reminded them of anything, and the share who named the real event is published below. And an adversarial verifier: a separate reviewer hunted for hindsight leaking into the reasoning.

The real-brand round ran on 21 July 2026. It used a faithful copy of the two engines' prompts in a test harness rather than the live product, because the live debate engine has web search switched on. The harness was the stricter way to keep everyone blind.

The taste paradox: 60% liked New Coke, 43.5% would accept it

We also gave the same population a second, taste-only question: two unlabelled cups, one a little sweeter and smoother, which do you prefer? That is the question Coca-Cola's own research asked. The answer reproduced the trap the real sip tests fell into.

Reading the gap. The direction matched: history rejected the replacement and so did the synthetic town. The poll was softer than history, though. 43.5% saying they would keep buying looks like a split crowd, while the real revolt was close to total. The debate caught what the poll softened, by landing on the one move history was forced into. In this case the debate was the better predictor.

Checks: the trick question passed (51.7% against 46.7%, a 5.0-point gap against a limit of 8). The verifier rated hindsight leakage low. But 17 of 20 probed personas (85%) named New Coke unprompted, which limits what this case can claim.

Checks: the trick question passed (51.7% against 48.3%, a 3.3-point gap). 13 of 20 probed personas (65%) named or clearly described the real campaign. The verifier did not return a result for this case (a formatting failure in its output), so the poll and debate stand without that extra review.

The obvious objection is that a language model has read about New Coke. So on 27 July 2026, six days later, both decisions ran again under fictional brands as part of a ten-case disguised round. Coca-Cola became Keystone Cola, a 108-year-old soda brand retiring its original formula. Domino's became Romano's Pizza, a national chain publishing its own worst reviews and rebuilding its recipe. Each poll used 150 new personas on the same five-step scale, from strongly support to strongly oppose.

Both answers held under a different name, and the verifier found no persona naming the real company. Read that carefully. The disguise was thin: a century-old cola losing sip tests to a sweeter rival is recognisable whatever it is called, and neither recorded recognition figure (60% for Keystone, 65% for Romano's) is a clean measurement. Across all ten disguised cases the round missed its own pre-set pass rule, and we published that too. For how we think about validating a synthetic population at all, see the validation problem.

It shows that on two real decisions with opposite outcomes, the synthetic town's verdicts matched history under blind conditions, its trick-question checks came back clean, and the answers held when the brand names changed. It does not show that the system can forecast a decision nobody has seen before. Plainly:

• These are famous cases. 85% of probed personas recognised New Coke and 65% recognised the Domino's campaign. Read the result as consistent with history, not as a from-scratch forecast.
• Stated intent runs softer or hotter than behaviour. The Coca-Cola poll understated the real revolt; only the debate caught its severity. Trust the direction and the reasoning more than the exact percentages.
• One advance prediction missed. Lapsed Domino's customers were expected to respond most and responded least.
• Some checks were incomplete. The Domino's verifier returned nothing, and the other verifiers reviewed only part of the persona rationales.
• The samples are simulation-sized. 200 or 150 personas per poll, five panelists per debate, 60 per trick-question group, 20 per recognition probe.
• Two cases are two cases. They are two of five in a real-brand round that passed on direction five out of five, and two of ten in a disguised round that failed its overall pass rule.

The pattern is still worth having. The synthetic town punished a change that took away something people loved and rewarded one that fixed something they had complained about, and the debate surfaced the conditions under which each would work. That is the case for rehearsing a decision before making it, and for reading synthetic agreement with care.

Why did New Coke fail if people preferred the taste?

Because the market was not asking which cola tastes better. It was asking whether people would let their Coke be taken away. In our replay 60% of synthetic consumers preferred the sweeter cola on taste alone, but only 43.5% would accept it as a replacement, and lifelong loyalists accepted it at 19%.

Why did the Domino's turnaround work?

Domino's fixed something customers had already complained about, changed the product before the ads ran, and invited people to judge for themselves. The synthetic panel set the same conditions on its own: remake first, confess second, and back it with a no-risk way to taste it.

Can synthetic personas predict how consumers will react to a product change?

On these two cases they called the direction of both outcomes correctly, under blind conditions, and held when the brands were disguised. Because most personas recognised the cases, that is evidence of consistency, not proof of forecasting. Use synthetic consumer research for direction and reasoning, and treat exact percentages with caution.

What is a time-locked synthetic persona?

A synthetic person who reasons as someone living in a given year, is told it knows nothing after it, and has web search turned off, so it cannot look up how a story ended.

If you have a product change you would rather rehearse than find out about in the market, we should have a conversation. Thirty minutes, no commitment.