signalo
Run your first test
How it works Proof Simulated Customers
Run your first test Log in

Proof

Simulated Customers answer like real people.

We re-run studies real people already answered. Here is how close the simulated customers (digital twins) come, misses included.

89% 41 of 46 message pairs called the same way as 1,536 real women. PLOS ONE study
See the studies Read the research

Thousands of simulated customers, each grounded in real people’s data

81%99 of 122 calls matched real people
8control studies against real people
3,885real people compared
78-93%health study, across 18 repeat runs

Signalo control studies

Real people answered first. Then the Simulated Customers.

A match: the simulated customers preferred what real people preferred.

Health messages · US

Which message encourages screening most?

41 / 46message pairs matched
Same call as real womenDifferent call
0.25average gap, 1-5 scale
18 repeat runs · pairs matched
3034384246

Halvorson-Fried et al., PLOS ONE, 2025 · 1,536 women · 864 simulated customers

Average rating per message
Real womenSimulated Customers
Your health is in your hands.3.63 3.20
Did you know? It is recommended...3.48 3.16
Did you know? Doctors recommend...3.46 3.09
There is a plan for your life...3.43 3.26
Did you know? Everyone with a cervix...3.40 3.11
Important even if not sexually active.3.35 3.02
Even if you received the HPV vaccine...3.29 2.85
Don’t put it off.3.27 2.98
Good news! Screening can prevent it.3.26 3.18
Monogamous or into hookups...3.12 2.90
Schedule your screening today.2.45 2.38
12345
Grocery prices · US

Which product wins at this price?

23 / 27price sets matched

Kansas State Meat Demand Monitor · 1,456 shoppers

Naming · US

Which company name wins?

19 / 25name pairs matched

Insights24 survey, Pollfish 2024 · 300 people

Product claims · US

Which claim sells protein powder?

4 / 4calls matched
PeopleSimulated Customers
Protein ahead of B
82% 100%
Protein ahead of C
64% 98%
Protein ahead of D
68% 100%
Protein ahead of E
68% 92%

PickFu poll, May 2026 · 50 people

Ad headlines · US

Which headline is unclear?

2 / 2differences found
MarketersSimulated Customers
E less clear than A
+18.3 +18.3
E less clear than D
+19.1 +13.7

Brand24 survey, Pollfish 2026 · 150 marketers

Image ads · Poland

Which Facebook ad got more clicks?

3 / 5pairs matched · 1 tie
Meta click gap · Simulated Customers picking the winner
1.6x
56% Match
1.8x
43% Miss
2.0x
66% Match
1.3x
50% Tie
3.3x
73% Match

Brand24 Meta ad results, 2023-2026

Brand names · US

Which brand name fits?

1 / 2polls matched

Two public PickFu polls · 150 people

Image ads · obvious pairs

Does it catch the obvious?

5 / 5checks passed
  • Same ad twice · no winner50 / 50
  • Sharp vs blurred · sharp69 / 31
  • Price vs no price · price82 / 18
  • Readable vs low contrast · readable89 / 11
  • Logo vs no logo · logo67 / 33

Made-up brand, one change per pair · 600 simulated customers

Published research

What leading journals found.

Harvard · Stanford · Nature 2026
85%agreement with 70 real survey experiments
119,330real people
469effects
90%on studies not yet public

Ashokkumar et al., Nature 656

How Simulated Customers are scored

Real people are the ceiling.

People, asked again81.7%
Their Simulated Customers71.7%
= 88% of the ceiling. Same answers given twice, Twin-2K panel.
88%
Columbia · Marketing Science 2025of the human ceiling, simulated customers built from 2,058 people. Signalo’s US simulated customers start here.

Toubia et al.

90%
PyMC Labs · Colgate-Palmolive · 2025of human reliability on 57 product concepts (preprint).

Maier et al.

83-86%
Stanford · Google DeepMind · 2024-26of people’s own consistency, simulated customers built from 1,052 people.

Park et al.

89%
BYU · Political Analysis 2023of real voters’ 2020 choices predicted.

Argyle et al.

75%+
Berkeley · Columbia · Marketing Science 2024agreement on brand perception maps.

Li et al.

73-81%
Tsinghua · Nature Comp. Sci. 2025of simple effects in 156 experiments replicated.

Cui et al.

71%of 3,000+ researchers expect most research to use synthetic answers within 3 years

Qualtrics researcher survey, 2024

25-80%less survey data needed when AI answers are blended in with a statistical correction

Marketing Science, 2026

2xhow much AI overstates effect sizes, so Signalo reads the winner, not the size

Nature, 2026

Across the industry

The whole category checks its work.

As each company published it.

AgreementWho ran itWhat was testedSource
95%EY with EvidenzaAI buyers compared with real buyers. EY’s Americas marketing chief: the conclusions were “95% the same”.

evidenza.ai

92%BCGAI shoppers predicted which drinks real consumers picked.

BCG, 2026

92%Electric TwinSimulated customers’ survey answers compared with real people’s. Real people asked twice agree with themselves 94% of the time.

electrictwin.com

92%AaruSimulated shoppers put 340 pairs of brands in the order real card purchases did. Published surveys of the same people managed 78%.

Aaru, 2026

90%EY with AaruSimulated customers built from 3,600 investors answered 53 questions; the order of their answers was compared with the real investors’.

EY, 2025

~90%Bain & CompanyA 1,500-person study rerun on simulated customers; the share of its key findings that came out the same.

Bain, 2026

88%EvidenzaAverage across more than 100 checks of its simulated customers against real surveys.

evidenza.ai

86%Artificial SocietiesSimulated customers’ answers compared with real ones across 1,000 surveys. Real people asked twice agree with themselves 91% of the time.

societies.ai

81%Salesforce with EvidenzaSimulated customers’ answers compared with Salesforce’s own surveys of CRM buyers.

evidenza.ai

75%Toluna with HormelSimulated customers ranked 26 product claims; the order was compared with real consumers’.

Toluna, 2025

How we count

A match means the simulated customers preferred the option real people preferred. Health-message pairs are those whose average ratings differed by 0.15 or more on a 1-5 scale. Where runs differ we show the one closest to today’s engine; misses stay in.

Not in the tally

Two image-ad studies run once, Walter 2024 and Lavertu 2025, both missed. Meta clicks are noisy (the same ad twice differed up to 1.7x), so pairs 1 and 4 are weak references.

Other companies’ figures

Their own tests or client quotes, each measured differently, so not comparable with each other or with ours.

What research warns

Simulated customers read groups better than individuals (Columbia, 2026) and overstate how big an effect is. Signalo reports the group’s call. A good hint, not proof: check with real customers before a big spend.

Before you spend

Test on Simulated Customers that have been checked.

Run your first test See how it works