Skip to content
Subconscious

Concept Testing Before You Build

On this page

Test a product concept as a choice, not as a rating. Put the complete concept in front of buyers next to the alternatives they have today, including doing nothing, and measure which one they pick. A rating tells you whether people like the idea in isolation. A choice tells you whether the concept wins against the thing the buyer already uses. When the attributes are varied across the choice sets, it also tells you which attribute does the winning. That is the question a build decision turns on.

Subconscious runs that choice experiment on a simulated market before you recruit anyone. It randomizes the concept's attributes across simulated buyers, estimates the effect of each attribute level on modeled choice with a confidence interval, and uses those estimates to rank the concepts you could ship. The public evidence grades the attribute estimates, not whole-concept rankings. The July 2026 working paper, not peer reviewed and summarized on fidelity.subconscious.ai, reports a mean Spearman rank correlation of estimated choice parameters of 0.55 across roughly 300 replicated published choice studies and 0.73 across the 43 studies that pass design filters. No public comparison yet grades a ranking of complete concepts against human choices. Treat the simulated ranking as the set of candidates to carry forward, and confirm direction, order and size with human research for the options you ship. The validation page lists every public comparison, misses included.

This page covers product concepts: the offer, its features and its price. For taglines, value propositions and claims, read how to test messaging before launch. For price alone, read how to test pricing before launch.

Which concept-testing method measures what

MethodWhat the buyer seesWhat it measuresWhat it cannot tell you
Monadic concept ratingOne concept, then a purchase-intent or appeal scaleStated appeal of that concept on its ownHow it fares against the alternative the buyer has today, or which attribute drove the score
Focus groupSeveral concepts, discussed in a roomWords: first reactions, objections, confusions, the language buyers useA measured choice share for any concept; it is qualitative research
Concept-choice experiment (choice-based conjoint)Sets of complete concepts with varied attributes, plus a "none of these" optionEstimated effect of each attribute level on choice; preference share for whole concepts against the outside optionRealized sales; a conjoint is still a stated choice
Simulated concept-choice experiment (Subconscious)The same choice sets, shown to simulated buyersThe same estimates, before anyone is recruited; a ranked set of concept candidatesHow large each effect will be in the market
Fake-door live testA real listing, button or landing page for a concept that does not exist yetObserved clicks or sign-ups from real trafficWhy people clicked, and anything about concepts you did not build a page for

The methods stack rather than compete: the focus group gives the attributes and the words, the choice experiment gives the ranking and the effect of each attribute, the fake-door test gives observed responses from real traffic to the concepts you built a page for. The mistake is to read a ranking of appeal scores as a choice share against the buyer's current option.

Worked example: a moisturizer conjoint graded against two human waves

Mintel, a market research firm, had already run a choice-based conjoint on a 50ml facial moisturizer the standard way: one wave of 1,000 human respondents, then a second blind wave of 1,000 as a validation cohort. Subconscious ran the same attribute set on simulated buyers built from demographics only, with no human preference or choice data as input (Mintel case).

Modeled attribute distributions reached 0.79 to 0.88 Jensen-Shannon similarity against the human waves. That is distribution agreement, not a percentage of correct predictions.

The model also missed something. Human demand dropped sharply between $34 and $48, and the model assumed a smoother decline, so the two diverged at the top of the price range. For a concept test that is the limit to design around: a simulation can rank the concepts and still miss a cliff in one attribute that only a human wave showed.

Worked example: compostability wording and a price point for an existing product

New Age Floral sells an eco-friendly replacement for floral foam whose sales had gone flat. The company asked which features and price points would drive sales and whether "home compostable" or "commercially compostable" would resonate with buyers. Subconscious ran five simulated experiments on those questions (New Age Floral case, source document dated 2026-05-15).

"Home compostable" produced higher modeled purchase likelihood than "commercially compostable". Price was the strongest driver in the simulated choices; the simulator favored $60 for maintaining modeled share and estimated declining demand above $100. These are modeled findings, not observed sales thresholds. The case does not document a post-study sales result, so the $60 price and the wording comparison stand as hypotheses, not as outcomes.

How to set up a concept test

  1. Define each concept as a complete offer. Name the product, the features that come with it, the price, the terms and what is excluded. A roadmap with ten possible features is not a concept; a buyer cannot choose a list.
  2. Define the outside option. The buyer's current product, a competitor, or nothing at all belongs in every choice set, shown complete and held constant. Without it, each share is a share among the shown concepts only, and it says nothing about whether buyers would leave what they have.
  3. Name the attributes that vary and their levels. Record the combinations you cannot build so the design never shows one.
  4. Set the smallest useful difference before the run: the gap in modeled choice share, at the cost of building the feature, that would change what you ship.
  5. Randomize the attributes across buyers so each attribute effect comes from a designed contrast rather than from one favorite concept against one straw man.
  6. Read the simulated result as a ranked set of candidates. Then run a human choice study that varies the attributes of the leading concepts, sized from the smallest useful difference in step 4, and read direction, order and size from it. Where a cliff like Mintel's price drop would change the decision, put the human sample's effort there.
  7. Keep the record: every tested concept, the outside option, the design, the population settings and the run identity, so a later reader can tell a measured difference from a cost constraint and from an idea nobody tested.

Steps 1 to 5 apply to a human conjoint as much as to a simulated one. The simulated run changes the order: the ranking comes first and sets the priorities for human research. Keep a few lower-ranked concepts in the human study, because the public evidence does not show that the simulated ranking of complete concepts matches human choice.

What the public evidence supports

The public study registry grades each replicated study on its own page: whether the direction of the effects held, whether their rank order held, and how the simulated effect sizes compared with the human ones. Results differ from study to study, and the registry lists the misses next to the matches.

For a concept test that reads as follows. Which attribute moves choice, and in which direction, is supported as a screen, in settings like the registry studies where direction held. How the attributes rank by strength is supported in settings like those where rank held. How large the shift will be, in share or in revenue, is not validated: the evidence page says no result has yet shown simulated estimates landing inside human confidence intervals experiment by experiment. A registry study that resembles yours does not validate your population, your design or the model version that runs it.

What a focus group gives that a choice experiment does not

A choice experiment produces numbers: effects, ranks and shares. It does not produce the sentence a buyer says when she rejects your concept, the feature she assumed was included, or the word she uses for the problem you solve. A focus group does, and those words become the attributes, levels and stimulus wording of the choice experiment that follows. Run the group while the concept is still loose, and the choice test once the concepts are fixed enough to compare.

The same distinction applies to chat-based synthetic research. Nielsen Norman Group tried a sample concept on a ChatGPT-generated user and, in an article dated June 21, 2024, reported that such users tend to find every idea good and to value every feature, which makes them a poor guide to feature priority (NN/g, Synthetic Users). A choice design forces a trade-off, so a simulated buyer cannot approve every concept at once, but the design alone does not make the result right.

To scope a concept test, book a decision review and bring the concepts you could build, the buyer's current alternative and the build deadline.

How this page was made

By: Avi Yashchin, CEO and Co-Founder. Drafted with AI assistance on 2026-11-07 from the public Mintel and New Age Floral case pages, the July 2026 working paper as summarized on fidelity.subconscious.ai, the public study registry (read 2026-11-07, rechecked 2026-11-12), and the Nielsen Norman Group article dated June 21, 2024. No private run data was used.