Do synthetic rows increase the independent human sample size?

No. A file with 60 human responses and 120 synthetic rows contains 180 records but still only 60 observed human responses. Any improvement in an estimate needs evidence and an appropriate account of uncertainty.

What should you compare a proposed boost against?

Use a pre-agreed human-only analysis and an untouched reference, with the same permitted inputs. Include weighting as a baseline if that is your current approach, and evaluate the answer shares, segment differences or rankings that matter to the decision.

How do you choose a synthetic sample-boosting provider?

Last updated: October 19, 2026

Choose a synthetic sample-boosting provider by the data it needs, the questionnaire rules it preserves, and the improvement it demonstrates over your original analysis. Ask how it measures uncertainty and distinguishes modeled records from human responses. A larger exported file, by itself, does not establish a more accurate estimate or a larger independent human sample.

By Lewsearch. Vendor descriptions checked October 19, 2026.

First, agree on what “boosting” means

Two proposals can use the same label and require different inputs. One may learn patterns from respondent records in your existing survey. Another may simulate additional profiles answering the same instrument. Ask the provider to explain which approach it proposes and what evidence supports that approach for your study.

For example, Fairgen says its Boost models train only on each study’s data and preserve questionnaire features such as skip logic and multi-selects. Ipsos describes boosting as augmenting limited or uneven samples and evaluates synthetic data through its SURE framework. These are the providers’ descriptions, not an independent comparison of performance. Fairgen’s research-provider workflow, Ipsos on synthetic data boosting.

If you need answers to questions absent from the original survey, that is a different evaluation problem. See synthetic audiences and new questions.

Put these five questions in the brief

1. What human evidence does this segment need?

Request the minimum usable input, the relevant subgroup requirements, and the conditions under which the provider declines a boost. Ask whether the source data contains the behaviors and relationships your analysis needs. Demographic overlap alone is a weak reason to assume a model understands a purchasing decision.

Kantar emphasizes the relevance and freshness of human training data, including whether a model’s knowledge fits the question being asked. Its guidance supports asking about the evidence behind the target segment rather than accepting a universal minimum sample size. Kantar’s synthetic-data guidance.

2. Will the exported records obey the original questionnaire?

Supply the codebook, routing, answer codes and missing-value rules. Ask for a small output example that shows how the provider handles screened-out respondents, questions not asked, multi-select answers and derived variables. Agree how mapping errors will be reported and corrected before delivery.

3. Does the boost improve the analysis you would otherwise use?

For a model that learns from respondent records, reserve evaluation cases before fitting or tuning. Compare the proposed boost with a pre-agreed human-only analysis, using the same permitted inputs and untouched reference. If weighting is your current approach, include it as a baseline.

Test the outputs that matter: answer shares, segment differences or a ranking. Matching the training data is insufficient evidence of improvement on unseen cases. The fuller blind-validation checklist covers scoring and separation of evaluation results.

4. How is uncertainty calculated after augmentation?

Ask for the original human count, generated-row count and an explanation of any claimed effective sample size. Do not assume ordinary survey formulas become valid by substituting the larger row count. Ipsos warns that treating synthetic observations as real ones in standard tests can substantially overstate confidence. Ipsos on calibrating synthetic confidence.

Illustrative example: a segment contains 60 human responses and receives 120 synthetic rows. The file now has 180 records, but still only 60 observed human responses. Whether modeling improves the estimate requires evidence; the arithmetic does not answer that question.

5. What exactly will you receive and pay for?

Specify the output format, synthetic labels, preserved human count, codebook, quality report and rules for reruns. Ask whether pricing covers one instrument, selected segments, revisions and evaluation. Confirm permitted use of uploaded data, retention and deletion terms before sharing respondent records.

When should you collect more human responses instead?

Prefer additional human research when the relevant audience is missing, the source questions do not cover the decision, or the proposed method cannot show improvement on a suitable reference. Keep modeled and observed results distinguishable in reporting, even when a blended analysis is useful.

Discuss a Lewsearch pilot

Lewsearch offers demo-led self-serve research starting at $595/month and custom Enterprise deployments. Discuss the scope of a sample-boosting project at the demo. Keep the human evaluation results hidden until the pilot’s predictions are locked, then compare the agreed outputs.

Book a demo or read the sample-boosting workflow.

FAQ

Do synthetic rows increase the independent human sample size?: No. A file with 60 human responses and 120 synthetic rows contains 180 records but still only 60 observed human responses. Any improvement in an estimate needs evidence and an appropriate account of uncertainty.
What should you compare a proposed boost against?: Use a pre-agreed human-only analysis and an untouched reference, with the same permitted inputs. Include weighting as a baseline if that is your current approach, and evaluate the answer shares, segment differences or rankings that matter to the decision.

Methodology and coverage Pricing Book a demo