A concept test answers one question before you spend on tooling, print or a production run: what do the people you want to sell to make of this idea? An AI agent can now run that study end to end. You describe the decision, attach an image of each concept, approve the plan, and get back what participants understood, what they believed, which concept they preferred and why.
This guide walks through that workflow for packaging and product concepts. It covers the brief, the images, the exposure design, what the moderator probes, the checks to make before you approve, and how to read and iterate on the results. If your stimulus is a line of copy rather than a design, the tagline testing workflow covers that case.
When is an agent-run concept test the right tool?
Use one when you have a visual concept and need to know how a real audience reads it before you commit. The common consumer cases:
- Packaging redesigns. A refreshed pack for an established product, where the risk is that loyal shoppers stop recognizing it on the shelf.
- New product concepts. A new flavor, format or line extension shown as a render or concept board, before formulation or tooling.
- On-pack claims and cues. A sustainability badge, a protein callout or a premium finish, tested inside the full pack rather than as a sentence on its own.
- Ad and display creative. A key visual or end-aisle display concept, where you need the first read before the explanation.
- Product concepts for an app or service. A feature sketch or onboarding screen, judged on whether people understand the value.
The agent is the right operator when a concept will go through more than one round. It runs identical probes across rounds, so round two compares directly with round one, and every finding comes back cited to the quote behind it. It also keeps pace with the design work, since most panel studies fill within 24 hours.
It is the wrong tool when the question is usability. If you need to watch people click through a live site or a working flow, run a prototype test with a link instead of a concept test with an image.
How do you write the concept test brief?
The brief is the single biggest lever on the quality of the result. Give the agent four things.
- The decision. Name the choice the research will inform: “Pick one of three pack designs for the spring reset” or “Decide whether the oat-based recipe concept is worth formulating.” A decision gives the moderator something to probe toward, and it gives the report a question to answer.
- The audience. Describe who buys the category and where: “US shoppers who bought refrigerated plant-based yogurt in the past month, mix of current brand buyers and competitor buyers.” Name the country, because panel recruiting runs by country, and say whether you want the panel or your own customers.
- The stimuli. List each concept with its label and one sentence on what makes it different: “Pack A keeps the current color system; Pack B moves to a white background with a larger product photo; Pack C adds a front-of-pack protein callout.”
- What you need back. Say what the report must contain: a preference count by concept, the reasons behind each pick, which concept current buyers recognize fastest, and any objection that would block purchase.
A strong brief also says what you already believe, so the study can test it rather than confirm it: “We think Pack B looks more premium but may lose shelf recognition with current buyers.” The planner turns that into a probe, and the report tells you whether participants agreed.
How do you prepare concept images?
The moderator shows each image inline during the interview and refers to it by its label, so the images themselves are part of the research design.
- One concept per image. A board with three packs side by side invites a comparison before you have captured a first read. Export each concept as its own file.
- Consistent framing. Same angle, same lighting, same background and same size across concepts. If one pack is a polished render and another is a sketch, participants react to the finish rather than the idea.
- Matched information. Keep the brand name, price point and required claims at the same level across concepts. Change only what you are testing.
- A participant-facing label. Labels like “Pack A” or “Recipe concept B” work. Internal names like “final_v3_legal” or “the Sarah option” leak into the conversation, because the moderator reads the label aloud.
Attach the images in the chat with your agent, or share a public link to each one. Files can be PNG, JPEG, GIF or WebP, up to 10 MB each, and three to five concepts is plenty for one study, since more than that leaves less interview time for each. The agent uploads each image with its label, and the planner checks the assets against the plan. Uploading a new image under an existing label replaces it, which makes it easy to swap a corrected render before launch. Concepts appear in the order you upload them, and the moderator decides when to surface each one during the interview, checking that the participant can see it before asking detailed questions.
Monadic or sequential exposure?
The exposure design decides what kind of evidence you get, so choose it in the brief.
A monadic read gives each participant one concept. Every reaction is uncontaminated by the others, which is the closest match to how a shopper meets a single pack on a shelf or a single product in a feed. Run one study per concept with the same plan and audience, so a three-concept monadic test is three studies whose results compare directly.
A sequential read gives each participant every concept. People can explain why one concept beats another, which is where the sharpest language about differences tends to come from. Attach every concept to one study and ask the moderator to capture a first reaction to each before any comparison. Concepts appear in upload order, so upload first the concept you most need a clean first read on.
Tell the agent which read you want, and the planning step builds the plan and the screeners around it.
What does the moderator probe on each concept?
The AI moderator runs 20–30 minute interviews and works through the same sequence for every concept, laddering 5–7 levels deep on the answers that matter.
- First read. What is this, and who is it for? Asked before any explanation, so you hear what the concept communicates on its own.
- Comprehension. What does it promise, in the participant’s own words? A pack that half the participants describe differently has a clarity problem, however much they like it.
- Appeal. What draws them in, and what puts them off? The moderator asks which element did the work: the color, the photo, the claim or the shape.
- Believability. Would the product deliver what the pack says? A “high protein” callout invites a look at the nutrition panel, and the moderator asks what the participant would need to see before trusting it.
- Fit with the brand. Does this look like the brand they know? A minimalist redesign can read as premium to new shoppers and as a different company to loyal ones.
- Purchase intent and why. Would they buy it, instead of what, and at what point in the trip? The “why” matters more than the yes or no.
- What would change their mind. What would turn a no into a yes, or a yes into a no? This is where the fixable objections surface.
Laddering is what turns a thin answer into a usable one. “It looks fresher” becomes which part looks fresher, what fresher means for a yogurt, whether it signals a recipe change, and whether a recipe change would worry them. Six questions in, you know whether the redesign is a win or a risk with your current buyers.
What should you check before you approve a concept test?
Any MCP client can run this workflow. Connect it in Claude, ChatGPT, Codex or Cursor in a few minutes (setup), then paste the brief and attach or link the images. The agent creates a concept-test study, sends the brief and each labelled image to the planner, relays the planner’s questions and returns the complete saved plan. The agent shows you the saved plan and a cost estimate before anything is spent; in Claude you approve with one click, elsewhere you confirm in the conversation.
A concept test has more moving parts than a plain interview study, so read the saved plan for these five things before you approve:
- Every concept, correctly labelled. Each image appears under the label participants will hear, in the order you uploaded it.
- The exposure design you asked for. A monadic test is one study per concept with the same plan and audience; a sequential test puts every concept in one study.
- First read before description. The first-read question comes before the moderator says anything about what the concept is.
- The right shoppers. The screener admits the category buyers named in the brief, in the country you named, with the mix of current and competitor buyers if you asked for one.
- Every probe, for every concept. Comprehension, appeal, believability, brand fit and purchase intent each appear for each concept, not only the first.
Any change to the plan or the images produces a fresh plan to review, so the version you approve is the version that fields. In Claude Code, the plugin’s run-a-concept-test skill walks through these steps; the Claude setup guide covers it.
Over the API, the same steps run in code, and your code launches against a saved estimate. The research API and MCP server for AI agents covers the product, and the developer integration guide shows the calls.
How do you read concept test results?
Read in this order, and ask the agent for each layer explicitly.
Prevalence by concept. Start with the counts: how many participants preferred each concept, how many recognized the brand at first glance, how many paraphrased the main promise correctly. Every count should carry its denominator, written as n of N, so a small sample is visible as a small sample. If you asked the same preference question in every interview, the report tallies the explicit picks and lists who chose none.
The reasons behind preference. A preference without a reason is not actionable. Ask the agent to group the reasons for each concept and show the quotes behind each group. Two concepts can win for very different reasons, and the reason tells you what to keep in the next round.
Minority objections. The objection raised by a handful of participants is often the one that matters at launch: the pack that looks like a competitor’s, the claim that reads as a price increase, the color that signals a flavor the product does not have. Ask the agent to list every objection that appeared, with who raised it and the quote.
Quotes in context. Ask the agent to open two or three full interviews for each concept. Each quote links to the source interview and, for voice and video, to the moment in the recording, so the team can hear the reaction rather than read a paraphrase. Concept tests also capture the screen, so you can see which image was on screen when the participant spoke.
Here is the reading pattern that works best for a packaging decision. Start with the first-read and comprehension findings, because a concept that participants misread cannot be rescued by appeal. Then check believability and brand fit, since a concept people like but do not trust will underperform once it meets the real shelf. Only then look at the preference count, and read it alongside the reasons, the segment that held each view, and the objections that would block purchase. A concept that wins on preference among competitor buyers but loses recognition among current buyers is a strategic choice about which audience the redesign serves, not a research failure. Ask the agent to put that trade-off in one table: concept, who preferred it, why, what would block purchase, and the quote behind each entry. That table is usually the artifact the team decides from, and every entry traces back to an interview someone can open.
How do you iterate to round two?
A first round finds what each concept communicates and what blocks it. A second round tests whether the fixes worked.
Change one thing at a time. Keep the strongest concept as a control, then add variants that answer the specific problems round one found: a larger product photo where recognition was low, a proof point where the claim was doubted, a different color where the pack read as the wrong flavor. Ask the agent to write the round-two brief from the round-one findings, quoting the objection each variant is meant to answer, and to keep the probes identical so the two rounds compare directly.
The agent can then answer the question that matters: did the fix move the reaction, or did a new objection replace the old one? Because every study stays in your research library, the agent can cite interviews from both rounds in the same answer. When the shortlist comes down to a close call that needs a measured gap, the refined concepts are ready to hand to a quantitative test.
How does User Intuition run concept tests from an agent?
User Intuition is a research platform that agents operate through its MCP research API. For a concept test, the agent creates a study with the concept-test type, sends your brief and each image with its participant-facing label to the planner, relays the planner’s questions back to you, and returns the complete saved plan for your review. A concept test with images runs in voice, chat or video; choose video when the participant’s expression is part of what you want to see.
Participants come from a vetted 4M+ consumer panel (B2B audiences recruited to order per study), or your own participants, invited by email or link. Interviews run in 80+ languages, so the same pack can be tested in several markets from one brief. Every panelist call is reviewed by the User Intuition team, panelists pass a video identity check, and only interviews that pass quality scoring are billed; see how quality works. User Intuition earns 98% participant satisfaction and is rated 4.9/5 on G2 and 5/5 on Capterra.
A concept test costs $30 per voice interview, and signup includes 3 free interviews with your own participants; panel recruiting is billed separately. To see the shape of a finished study, explore an example study. For research design guidance beyond the agent workflow, see the concept testing solution.
Copy-paste prompts
Each prompt names the decision, the audience, the country, the stimuli and what to return. Attach the images before you send.
1. Packaging redesign, monadic.
Run a concept test on our yogurt packaging redesign. Decision: choose one of three packs for the spring reset. Audience: US shoppers who bought refrigerated yogurt in the past month, half current buyers of our brand, half competitor buyers. Use the panel in the United States. Design: monadic, so create three studies with the same plan and audience, one pack in each. Attached: Pack A (current colors), Pack B (white background, larger product photo), Pack C (front-of-pack protein callout). Probe first read, comprehension, appeal, believability, brand fit, purchase intent and what would change their mind. Show me the saved plans and estimates before launch. Return recognition, appeal and purchase-intent counts by pack, reasons, and every objection with quotes.2. New product concept, sequential.
Test two snack bar concepts before we formulate. Decision: which concept to develop. Audience: UK adults who bought a protein or cereal bar in the past two weeks. Use the panel in the United Kingdom. Design: sequential, both concepts in one study, with a first reaction to each before any comparison. Attached: Concept A (oat and date, lunchbox format), Concept B (nut butter, single-serve premium). Return which concept participants preferred and why, who it is for in their words, what they expected it to cost, and what would stop them buying it.3. On-pack sustainability claim, own customers.
Create a concept test with our own customers. I will invite them by email. Decision: whether to add a recyclable-pack badge to the front of our coffee bags. Attached: Bag A (current), Bag B (with badge). Design: sequential, Bag A uploaded first, with a first reaction to each bag before comparing them. Probe whether the badge is noticed unprompted, what participants think it means, whether they believe it, and whether it changes how they see the brand. Return counts with denominators and the quotes behind each finding.4. One pack, two markets.
Run the same pet food pack concept test in the United States in English and in Mexico in Spanish. Audience: dog owners who buy dry food at a grocery or mass retailer. Use the panel in each country. Attached: Pack A (current), Pack B (new photography and grain-free callout). Design: sequential, one study per country with the same plan, and a first reaction to each pack before comparing. Return results by country: preference, comprehension of the grain-free callout, believability, and objections, with quotes translated and linked to the source interview.5. Round two from round one.
Read the findings from our yogurt pack concept test and write a round-two brief. Keep Pack B as the control. Add two variants I will attach: Pack B2 (larger brand logo, to fix recognition) and Pack B3 (protein proof point added). Quote the round-one objection each variant answers. Keep the probes identical to round one. Show me the plan and estimate, then compare both rounds when interviews complete.6. App feature concept.
Test a feature concept for our budgeting app with our own users. I will share the study link in-app. Attached: Concept A (weekly spending summary screen), Concept B (savings goal tracker screen). Design: sequential, with a first reaction to each screen before comparing. Probe first read, whether participants understand the value, how they would use it in a normal week, and what would make them ignore it. Return preference with reasons, any confusion about what each screen does, and quotes.