How are synthetic respondents built?
Built from real panelists: Each synthetic respondent is built from real PickFu panelists. It uses around 200 demographic and attitudinal profile traits, plus the choices people like them made on past PickFu surveys in that category. That includes what they rejected and why. A prediction, not a person: A synthetic respondent predicts what that type of person would choose. It has no opinions of its own, and it isn’t a copy of one specific person. Dozens of separate votes: When you run a test, PickFu draws synthetic respondents at random from the audience you picked. It asks each one separately which option it would choose and why. Then the votes are counted. A result is dozens of separate predictions counted into a vote, not one AI’s opinion. Past answers it draws on: Synthetic respondents use up to 50 past answers for text tests and up to 10 for image tests. They see past images as written descriptions, not the images themselves. That’s one reason image results are less reliable than text. Why it’s not just asking ChatGPT: Ask a general AI chatbot to act like a dog owner and it invents one. A synthetic respondent starts from the recorded choices of real dog owners on PickFu. PickFu can also check the method against past surveys where it already knows what real people chose.How do I run a synthetic test?
Where you can run it
- PickFu web app: Open the Synthetic Audiences page from Synthetic audiences in the sidebar, and start from one of the example prompts. You can also ask the PickFu Agent to run a synthetic test.
- Slack: Ask the PickFu Agent in Slack.
- AI tools like Claude: Use the PickFu MCP. Through the MCP, your own agent can generate options, test them, refine, and retest without leaving the tool.
Steps
- Describe what you want to test. For example, share a question and the options, or a concept you want reactions to.
- Let the agent run it. When you’ve said what to test, the agent drafts the test and runs it right away, without waiting for your approval. If your request is vague and the agent would have to guess the question, the options, or the audience, it asks you first. It tells you the synthetic audience, the question and options, and the sample size. To have the PickFu Agent ask first, choose Always ask or add a permission rule in Settings > Agent.
- Read the results in the chat. The agent’s summary is the result. There’s no separate results page yet.
- Decide on next steps. The agent suggests variations to retest and offers a human survey on your stronger options.
Two ways to run a test
- Comparison: Give it a question and 2 to 10 options, either text or images. You get the options grouped into a stronger half and a weaker half, the themes behind that split, and one-line reasons from the synthetic respondents. Two options is a head-to-head. Three or more are grouped into a stronger and weaker half rather than ranked one by one.
- Single concept: Share one idea and one open question. The agent plans the research, gathers open-ended reactions from the synthetic respondents, and returns a summary. The summary covers the themes, where segments disagreed, and follow-up questions worth asking.
What to expect from your results
- Speed: Results come back in the chat in a few minutes.
- What you get: For comparisons, you get the stronger and weaker half, the themes behind the split, and a one-line reason from each synthetic respondent. For single concepts, you get a summary, the themes, where segments disagreed, and follow-up questions worth asking.
- Clear labeling: Every result is labeled synthetic. Keep that label on it wherever you share it, and don’t present it as observed consumer behavior.
- Wide margins: Synthetic respondents agree with each other more than real people do. If real people would split 60/40, synthetic respondents might split 96/4. They usually favor the same side real people would, but the gap looks bigger than it is. That’s why results focus on direction, not exact percentages. An option with zero support isn’t necessarily a bad option.
- More detail on request: You can ask the agent for the vote split or the individual synthetic responses. Treat the split as direction, not measurement.
- Reruns vary: Running the same test twice can return different numbers. The stronger and weaker halves and the themes should stay about the same.
- The human survey step: After a result, the agent offers a ready-to-run PickFu survey with real people on your stronger options. It shows you the price and asks you to confirm before publishing anything. You can skip it, but we recommend it for any decision with real money behind it. Real respondents give you written reasons, actual margins, and results you can show to others.
- When there’s no matching audience: The agent tells you instead of substituting a different audience. It suggests a human survey with the targeting you need.
When should I use Synthetic Audiences?
Synthetic decides what to test. Real people decide what to do. Synthetic Audiences works best for:- Narrowing a long list: Screen eight names down to the two worth testing with real people. This is the best use.
- Text comparisons: Names, titles, taglines, claims, and price framing.
- Early iteration: Get feedback at the idea stage, including content, feature, clarity, and trust questions.
- Clearly different images: Image comparisons work when the options are clearly different from each other.
- Categories where coverage is deepest: Supplements, health and household, beauty and personal care, books, and grocery.
- Aesthetic calls: Taste is hard to predict from past behavior. Subtle visual differences belong with people who can see them.
- Emotionally loaded questions: Synthetic respondents under-read fear and over-weight efficiency and price.
- Anything genuinely new: Synthetic respondents predict from precedent, and a truly new concept has none.
- Expensive or hard-to-reverse decisions: Final packaging, launch creative, or a name you plan to trademark.
- Exact margins or individual responses: Synthetic results show direction, not the size of the gap.
- Evidence for others: Anything you’ll present to a stakeholder, investor, or retail buyer.
The synthetic + human workflow
- Generate your options: Create your options, such as names, titles, or images.
- Screen them with a synthetic audience: Test the whole list in a few minutes.
- Improve the ones that survive: Drop the weaker options, refine the rest, and retest. Run as many rounds as you want.
- Confirm the decision with real people: Run a PickFu survey with the human panel on your strongest options. You get a written answer from every respondent.
How accurate is it?
PickFu benchmarks Synthetic Audiences against hundreds of past PickFu surveys where we already know what real people chose. It picks the same side as real people most of the time on text comparisons, and less often on images. Synthetic Audiences is in beta, and we keep testing and adjusting the models. Keep these points in mind when you read a result:- Trust the direction, not the size of the gap. Synthetic margins run wider than human ones. That’s why results show a stronger half and a weaker half instead of a single winner.
- Accuracy for a specific audience isn’t validated. PickFu’s accuracy testing covers the method overall, not each audience individually.
Frequently asked questions
Is this replacing PickFu’s human panel?
No. Synthetic Audiences helps you decide what to test. Real people help you decide what to do. PickFu’s confidence in Synthetic Audiences comes from the human panel, because that’s what we score the synthetic respondents against. See Who are PickFu’s respondents? to learn about the human panel.Can synthetic responses end up in my human surveys?
No. Synthetic and human results never mix. Synthetic responses never appear in your PickFu surveys. You only get synthetic results when you ask for them through the PickFu Agent or the synthetic tools in the MCP and API. They’re always labeled as synthetic, and PickFu never pools them with human results. Every human response comes from PickFu’s respondent panel, which goes through multiple quality and identity checks. You get a written answer and demographic info from each respondent.How is this different from asking ChatGPT?
General AI tools invent an answer with no record of what real people chose and nothing to check it against. Synthetic Audiences predicts from real PickFu panelists’ recorded choices, counts dozens of separate predictions, and is benchmarked against past surveys with known results.How is this different from other synthetic research tools?
Most synthetic tools can’t check their output against a live human panel. PickFu can, and does. See How accurate is it?What does it cost?
Synthetic Audiences runs on platform credits — every plan includes a monthly allowance. See What do platform credits cost? for what each action costs. The human survey the agent recommends afterward is optional and priced like any other PickFu survey.How many synthetic respondents should I use?
Use 30 to 50 for most tests. The maximum is 75. Past about 50, each extra synthetic respondent adds very little because their answers converge. Each synthetic respondent represents a type of real respondent, so results show themes and patterns rather than individual variation. If you need a large sample or exact margins, run a human survey.Why did I get different numbers when I ran the same test twice?
Each run draws a fresh set of synthetic respondents at random, and the model isn’t deterministic. The stronger and weaker groupings and the themes should hold, but the exact split moves. Treat the split as direction, not measurement.Is there a results page?
Not yet. Results appear in the chat with the PickFu Agent or your AI tool. The agent’s summary is the result.Which audiences are available?
Synthetic Audiences offers over 300 synthetic audiences across 25 product categories. Coverage is deepest in supplements, health and household, beauty and personal care, books, and grocery. Each audience is a category sliced by a trait or demographic, such as dog owners in pet food or women who take supplements daily. Audiences describe what a specific type of person would choose, not the US population as a whole. Current limits:- Audiences are US-only, with no geographic slices within the US.
- A test runs against one audience. You can’t combine audiences.
- Custom audiences aren’t available.
Can I test my own customers or a non-US market?
Not yet. Synthetic audiences are pre-built from US PickFu panelists. To reach a different audience, run a PickFu survey with the targeting you need.How does PickFu handle my data?
Survey and respondent IDs are hashed before anything reaches the model, so a synthetic respondent’s history can’t be traced back to a specific customer’s survey. Your survey results are never shared with respondents or other customers. PickFu may use survey data to train our own products. See our Terms of Service and Privacy Policy, or contact support@pickfu.com with questions.How do I share feedback?
Ask the PickFu Agent to submit feedback while you’re chatting with it, or email support@pickfu.com. Customizable audiences and non-US markets are the top requests so far.Related Articles
