The research found that the surveys with synthetic respondents were on average two to three percentage points away from the headline results generated by real respondents. 

For more complex research, such as studying subgroups, drivers, segments and changes over time, Strat7 said that synthetic data failed to create coherent respondents despite formulating convincing responses. 

The study, published in a paper titled ‘Synthetic data: Is this as good as it gets?’, concluded that on many functions, synthetic data was significantly different to the survey results gathered from real humans. 

The findings follow similar research by Strat7 in 2025, which found that surveys augmented by synthetic data ‘boosts’ aligned with real data at a basic level, but cited inconsistent segmentation, a lack of logical consistency, a ‘bunching effect’, whereby fewer responses were found at the extremes, and divergent outcomes in key driver analysis. 

The new study, conducted earlier this year, repeated the experiment, comparing a nationally representative sample of 3,000 with surveys run by two synthetic data companies, with three changes: a Van Westendorp pricing exercise and willingness-to-pay questions; a topical situational question on the 2026 World Cup; and a tracking test comparing real change between 2025 and 2026 against the change the synthetic data implied. One of the synthetic data firms was also replaced in the 2026 study. 

The results 

For consumers’ willingness to pay, Strat7 noted that improvements had been made by synthetic augmentation in 2026 compared with the 2025 research, with less bunching around whole numbers, such as £1 or £2, but prices given by synthetic respondents were generally 16% above those provided by real people. 

In a price ordering exercise, where respondents were asked to provide four prices for a product, ranging from ‘too cheap’ to cheap, then expensive and lastly ‘too expensive’, purely synthetic respondents broke the logical order 68% of the time, dropping to 32.8% when blended with real data.  

Two problems persisted – survey logic and subgroups, where there were often large gaps between what real and synthetic consumers said, with the paper warning that greater care was needed if synthetic boosting was used for more specific questions.  

Synthetic surveys also had a 47% success rate in accurately tracking changes in numbers between the two studies – essentially no better than a coin toss, according to the paper.  

Strat7 added that if synthetic augmentation cannot reliably detect changes in numbers over time, it is the “wrong tool” for that job.  

Overall, synthetic data could help with headline key performance indicators, basic pricing exploration and boosting genuinely hard-to-reach groups, but the researchers recommended either avoiding it for, or exercising caution with using it in: tracking studies and market-mix modelling, new segmentations and demand-space work, subgroup and multivariate analysis, situational questions; and research on new entrants or brands with low market penetration. 

Analysis 

Speaking to Research Live, Hasdeep Sethi, group AI lead at Strat7, said that almost four years on from the release of ChatGPT 3.5, his feeling is that using a survey setting in a bid to simulate human responses to survey questions “is probably not the right use of the technology”. 

“There are use cases for synthetic boosts,” said Sethi. “But to try to simulate at the individual level what a person would respond to across the whole survey is very difficult.” 

He said that in reality, the study highlighted the need for genuine primary research. “If your use case is just to understand how much to charge for this product, then [synthetic data will] give you what you need. But then as you dig deeper, things start to get a little bit ropey. 

“If you want precision, then you still need to ask real people, and you should be careful about the boost.” 

Sethi argued that digital twins are “a better application of the technology”, and said the study underlined that it was important to retain human-based research, with synthetic respondents capped at 10% to 20% of the total sample for acceptable results. “Don’t replace people,” he added. “Or if you do, be careful about the percentage of respondents that you’re replacing with boosts. 

“I do think it is a square plank in a round hole, because if you can’t use [synthetic data] in the same way that you plan a survey in all of its complexity, then it kind of falls down. Even the economics of it doesn’t add up sometimes.” 

Sethi described the use of synthetic data to boost surveys as “a photo versus a sketch”. “It gets some big details right, but beyond that, it’s not really a substitute for a survey,” he added. “The ability to talk to a reconstruction of your segments or your twins is probably a better use of the technology because that doesn’t depend on forcing it into a survey.” 

Using synthetic data to augment surveys could be more promising with hard-to-reach or niche audiences, Sethi added, but he said there was not currently a lot of demand in the market for synthetic boost.  

“We are seeing more and more of digital twins and synthetic personas, but that still relies on primary research data. I think the path that we’re all navigating is [determining] the right blend of primary research and this kind of conversational synthetic.”  

For all the promises made about the efficacy of synthetic data technology, Sethi argued, “clearly, primary research still has an important role, and it’s our job to push back on where people say primary research is not important at all”.