Can synthetic respondents estimate market size?

Agree what the market number is supposed to measure

Before commissioning research, finish this sentence: we need to estimate the number of ___ in ___ who could buy ___ during ___. Specify people, households, accounts or locations; geography; eligibility; the offer and price; and the time horizon. Decide whether the deliverable describes a broad addressable population, a reachable audience, stated interest or expected purchases.

The U.S. Small Business Administration's market-research guidance treats demand, market size, location, competition and pricing as separate questions. Our buying recommendation is to keep those questions separate in the model too. A plausible answer to one does not fill in the others.

For example, the number of households eligible for a service is a different quantity from the number interested in its description. Neither automatically gives the number that will subscribe in the next year. Ask the supplier to name the quantity behind every headline.

Audit the population count before multiplying it

To convert a share into a count, you need a defined population base. Trace that base to a dated source rather than asking generated profiles to establish how many real customers exist. Public statistics, a documented customer register or other appropriately sourced records may supply different parts of that count.

Units matter. The Census Bureau's Statistics of U.S. Businesses glossary distinguishes a physical business establishment from a company or enterprise that can contain several establishments. It also says SUSB excludes nonemployer businesses. Those definitions can materially change the relevant base for a product sold per location, per company or to sole proprietors.

Ask for a short denominator audit with six fields:

• Unit: what is counted, and what counts as one customer?
• Coverage: which geography, industry and population are included or excluded?
• Reference period: what date does the source describe?
• Eligibility: which product requirements can the source actually identify?
• Overlap: can the same household or account appear in more than one segment?
• Remaining gaps: which filters or counts are assumptions rather than measured inputs?

Do not multiply separate marginal percentages as though they identify the same people. Knowing how many households rent and how many have a particular income does not, on its own, establish how many meet both conditions. Obtain a suitable cross-tabulation or disclose and investigate the assumption used to combine them.

Keep simulated interest separate from purchase evidence

Ask what the synthetic question measures. Interest in a description, intention to try at a specified price and a purchase within a year are different outcomes. A market model needs an explicit justification for any step from one to another. Do not silently rename the share choosing an appealing answer as the share that will become customers.

For a consequential purchase forecast, request relevant human or behavioral evidence for that translation. Check the offer, audience, price, availability and period. Existing conversion data may help, but a free-trial result or a different product is not automatically an appropriate conversion rate for the proposed offer.

The willingness-to-pay guide covers the separate price question. The synthetic-panel validation guide explains how to protect evaluation outcomes and test the intended use. Agreement on unrelated benchmark questions does not establish the purchase rate for this market.

Worked example: 10,000 interested households is a scenario

Consider a fictional service. Assume, solely for this example, that a documented source and eligibility review identify 40,000 households in its service area. A synthetic exercise returns 25% interest in the described offer. These are invented figures, not a Lewsearch result or a real market estimate.

Multiplying 40,000 by 25% gives 10,000 households under the modeled-interest assumption. It does not establish 10,000 paying subscribers. The population count and the generated interest share need separate evidence; the step from interest to purchase remains untested.

To examine dependence on the interest assumption, a buyer could also calculate 15% and 35% scenarios: 6,000 and 14,000 households. These are deliberately chosen sensitivity cases, not a confidence interval and not evidence that the true value falls between them.

If the business needs 8,000 subscribers to proceed, these calculations do not resolve the decision. Even the middle interest scenario is not a subscriber forecast. Identify the purchase behavior that needs testing before committing to a launch; do not choose the scenario that makes the plan work.

Commission a model whose inputs can be challenged

Ask for a reproducible calculation with each input labeled as a sourced count, human measurement, observed behavior, synthetic output or editorial assumption. Keep units visible throughout. If revenue is requested, state the purchase frequency and revenue per transaction or subscription, using the same period and customer unit. Avoid counting overlapping segments twice.

For each uncertain input, ask what evidence supports it, what plausible alternatives change the decision and what additional research would reduce the important uncertainty. More generated responses may help investigate run-to-run variability; they do not by themselves establish population coverage or remove a wrong assumption.

The AAPOR Code of Professional Ethics and Practices distinguishes AI-generated cases from human research participants and calls for disclosure of how responses were generated. It also limits population and precision claims to what the methods warrant. Applying that transparency principle here means labeling the simulated inputs and their limits, rather than presenting the model as an observed customer count. AAPOR does not certify this market-sizing procedure.

Do not attach a conventional survey margin of error to the whole calculation just because software returns a large number of synthetic cases. Any uncertainty statement should identify what it covers, how it was calculated and which sources of error remain outside it.

Choose the next study from the weakest important input

If the population base is unclear, resolve definitions and coverage first. If the base is sound but the need or interest share is unsupported, commission an appropriate human comparison. If interest is supported but purchase conversion is unknown, investigate actual behavior with a suitable test before calling the result a sales forecast.

A synthetic exercise can be scoped as an exploratory input, with its assumptions exposed and its next evidence requirement written down. Keep the final deliverable at that level when stronger evidence is missing.

For a Lewsearch demo, bring the population definition, source counts, offer and business decision. Discuss which part of the question is suitable for a synthetic screen and what further evidence would be needed. This guide does not establish that Lewsearch has validated market-size or revenue forecasting for any particular use.

FAQ

Can I multiply a synthetic interest percentage by a census population?

You can show the arithmetic as a labeled scenario. Check that the population unit and eligibility match the question, and do not describe the result as measured demand without suitable evidence for the generated share and its intended interpretation.

Does a larger synthetic sample make a market estimate representative?

Not by itself. The population source, coverage, eligibility and model assumptions still need evaluation. More generated cases do not establish how many real customers exist.

Can simulated purchase intent be reported as expected revenue?

Not without an evidence-supported model connecting intent to purchases, timing and transaction value. Show those assumptions separately, test the ones that drive the decision and label unsupported outputs as scenarios.