No One Said the Ferrari Luce Felt Like a Ferrari
Synthetic consumer testing changes the cost of asking a basic question: what will people think?
Before a team builds prototypes, it can screen ten product concepts. Before a marketing team buys media, it can compare campaign angles. Before a growth team ships an A/B test, it can see which homepage claim feels credible, confusing, premium, cheap, risky, or exciting. The revolutionary part is not that synthetic respondents replace real customers. They do not. The revolutionary part is that early research becomes cheap and fast enough to run while the work is still easy to change.
This post shows that loop with a deliberately hard example: Ferrari’s electric Luce. What happens when a Ferrari becomes quiet, electric, and practical enough to seat five people? That is a brand question before it is a car question. Ferrari can make an electric GT fast. The harder problem is whether people still hear Ferrari in the product when the product has no engine note, no two-seat fantasy, and no obvious visual debt to the cars they grew up wanting.
I used AudienceKit to run 850 synthetic consumer interviews on a public-use sample designed to approximate the U.S. adult population. The result was not a clean rejection. It was more uncomfortable than that: the Luce scored closer to Ferrari references than to luxury EV rivals, but not one respondent gave it an outright “yes” on whether it felt like a real Ferrari.
The research loop
A synthetic audience test is a structured role-play, but with two important constraints. The “person” is not invented from vibes, and the interview is not an open chat. Each row in a respondent dataset becomes a persona card, then the model receives that card, a stimulus, and a fixed questionnaire. The output is a table: one row per respondent, one column per question.
That table is what makes the method practical. You can rank concepts, compare positioning frames, debug a landing page, test objections, segment responses, and read the open-ended language behind the numbers without waiting weeks for a traditional study. The right use is directional: sharpen the options, find the failure modes, and decide what deserves expensive human validation.
Unlike prompt-only persona demos, this approach starts with a sampling frame. The reason to use the General Social Survey here is not that GSS is special for car research. It is that the GSS is a public-use survey designed to represent U.S. adults, so it gives an inspectable way to build a broad synthetic audience from real population structure rather than invented persona lists. AudienceKit’s default GSS adapter loads the bundled 2024 public-use file, prepares respondent rows, keeps the survey weights, and renders each respondent with the library’s default persona template.
The useful pieces are deliberately plain:
- Rows, not invented personas. Each respondent starts as a row from a public-use U.S. survey, so income, education, region, occupation, household structure, politics, religion, and self-reported wellbeing stay in a real joint distribution.
- A visible persona renderer.
ak.build_persona(row)turns that row into a short identity card. Missing values render as “not reported” instead of being hallucinated into place. - Segment rules you can inspect. The affluent segment uses AudienceKit’s built-in luxury proxy: an income or upper-class anchor plus a second signal such as degree, high occupational prestige, or above-average finances.
- A fixed study contract. Each respondent sees one stimulus and returns the same JSON fields, so the analysis is ordinary pandas rather than transcript archaeology.
The methodology is easy to reproduce:
uv pip install "git+https://github.com/lfiaschi/audiencekit.git"import audiencekit as ak
pool = ak.load_panel() # public-use GSS 2024 file
broad = ak.sample_panel(pool, n=600, segment="broad", seed=20260602)
affluent = ak.sample_panel(pool, n=250, segment="luxury", seed=20260603)
persona = ak.build_persona(broad.iloc[0].to_dict())
print(persona.splitlines()[0])
# You are a 65 year old Male adult living in the South region of the United States.The persona card is only the first half of the interview. In this experiment, I used AudienceKit for the panel and persona renderer, then wrapped each persona in a Ferrari-specific study prompt and passed the stimulus image as a separate Gemini image input:
persona = ak.build_persona(respondent_row)
prompt = build_ferrari_luce_prompt(persona, cell)
image = image_path_for_cell(cell)
response = gemini.generate_content(
contents=[prompt, image],
response_schema=RESPONSE_SCHEMA,
)The actual text prompt was longer than the excerpt below, but its shape was:
# Role
You are a consumer participating in a product research interview.
Answer naturally from your own point of view and stay consistent with this profile.
# Who you are
You are a 65 year old Male adult living in the South region of the United States.
You describe your race or ethnicity as White; you were born in the United States, and you were raised in a small city or town (under 50,000).
You are Married, have 2 children, and your household has 2 adults.
Your reported family income last year before taxes was $110,000 to $129,999, from all family sources, not just salary.
You describe your social class as Middle Class.
Your labor-force status is retired. Your current or most recent occupation area is management, business, science, and arts occupations (occupational prestige: High).
Compared with other households, your financial situation is Average; financial satisfaction: Satisfied.
Politically you identify as Independent and consider yourself Moderate.
Your religious preference is Catholic; you describe yourself as moderately religious, and you attend services Once A Month.
You describe yourself as Pretty Happy overall and your health as Good.
# Product stimulus
Vehicle: Ferrari Luce
Description: Ferrari has launched the Ferrari Luce, its first fully electric GT. The car is a five-seat, four-door Ferrari with a new exterior and interior design.
Image context: You are looking at visual material for Ferrari Luce shared launch montage.
# Interview task
Evaluate the vehicle from your own point of view. There are no right answers.
Focus on the car's design, Ferrari-ness or brand fit, emotional pull, visual
recognizability, innovation or performance credibility, status signal, expected
controversy, and any loss of traditional Ferrari ritual. Be direct. You can like
some aspects while disliking others. Do not pretend to be a Ferrari expert unless
the profile gives you a reason.
Return JSON only. Do not include markdown, commentary, or extra text.
Required fields:
- "brand_fit": Rate how well this vehicle fits what Ferrari means as a brand, even when the vehicle itself is a benchmark from another marque.
- "ferrari_distinctiveness": Rate how unmistakably Ferrari this vehicle feels in design, product identity, and emotional character.
- "emotional_pull": Rate how much excitement or desire this vehicle creates for you.
- "visual_recognizability": Rate how easily someone would recognize this as Ferrari or as close to Ferrari's visual world.
- "innovation_credibility": Rate how credible this vehicle feels as advanced performance or luxury innovation.
- "ritual_loss": Rate how much the absence of traditional Ferrari sound, engine feel, and mechanical ritual would diminish this vehicle.
- "status_signal": Rate how strong a statement of status and success owning this vehicle would make.
- "controversy_expectation": Rate how much controversy this vehicle would create among Ferrari enthusiasts and loyalists.
- "ferrari_feel_verdict": Choose whether, bottom line, this feels like a real Ferrari: yes, mixed, or no.
- "first_reaction": Give your honest first reaction in one or two natural sentences.
- "best_argument_for": Give the strongest argument for this vehicle from your point of view.
- "best_argument_against": Give the strongest argument against this vehicle from your point of view.
- "quote": Write one concise sentence that captures your bottom-line reaction.
Use integers from 1 to 5 for the rating fields:
1 = very low / strongly disagree
2 = low / disagree
3 = neutral or mixed
4 = high / agree
5 = very high / strongly agree
Use exactly one of these strings for ferrari_feel_verdict: yes, mixed, no.That is the basic mechanism. The model is not being asked to “be a Ferrari buyer” in the abstract. It is being asked to answer one survey as one sampled respondent, while also seeing the assigned vehicle image. Every respondent in a cell gets the same text stimulus, the same image, and the same response schema.
AudienceKit also includes a generic build_survey_prompt helper for ordinary structured studies. For this study I used AudienceKit for a public-use, GSS-based U.S. sampling frame and default persona renderer, then ran a small schema-constrained Gemini 2.5 Flash survey runner around the Ferrari-specific prompt.
The study
A raw Luce score would not mean much. If a respondent gives the car a 3.1 on “Ferrari identity”, is that weak, middling, or better than we should expect for a four-door EV? The answer depends on where real Ferraris and non-Ferrari luxury EVs land on the same instrument.
I used ten between-subjects cells:
| Group | Cells | Why they are there |
|---|---|---|
| Ferrari references | Purosangue, Roma, 296 GTB | Estimate the Ferrari zone |
| Rival references | Porsche Taycan, Lotus Eletre, Rolls-Royce Spectre | Estimate the premium EV and luxury rival zone |
| Luce framings | Neutral, performance, heritage, LoveFrom design | Test launch stories with the same image |
Every respondent saw one image stimulus and one text description. The four Luce cells shared the same image montage, so movement between those cells comes from the text framing rather than the visual.
The main dependent variable is a Ferrari identity index: brand fit, Ferrari distinctiveness, emotional pull, visual recognizability, innovation credibility, status signal, and reverse-coded ritual loss. Higher is more Ferrari-like. I also asked a direct yes/mixed/no question: “Bottom line, does this feel like a real Ferrari?”
Cell sizes were 60 per cell in the broad panel and 25 per cell in the affluent-GSS proxy panel, for 850 total interviews.
The result
The benchmark map gives the cleanest read, but the marks matter. The x-axis is the Ferrari identity index: higher means the vehicle felt more Ferrari-like across brand fit, distinctiveness, emotional pull, recognition, innovation credibility, status, and preserved ritual. The y-axis is expected controversy. Red circles are individual Ferrari reference cells, gray squares are rival reference cells, and blue diamonds are Luce framing cells. The vertical guide lines are pooled identity-index averages inside each segment.
In the broad panel, Ferrari references averaged 3.66 on the identity index, with a 95% bootstrap interval from 3.61 to 3.71. Rival references averaged 2.57, with an interval from 2.51 to 2.64. The neutral Luce scored 2.99, with an interval from 2.96 to 3.03.
That puts the neutral Luce 0.66 points below the Ferrari references and 0.42 points above the pooled rivals. In the affluent-GSS proxy panel, Ferrari references averaged 3.72, with an interval from 3.64 to 3.80. Rivals averaged 2.57, with an interval from 2.45 to 2.69. Neutral Luce scored 3.02, with an interval from 2.93 to 3.13. The affluent neutral gap was 0.70 points to Ferrari and 0.45 points above rivals. Luce did not collapse into the rival set, but the neutral read was not close to the Ferrari reference zone either. It sat in the middle.
The direct verdict is harsher. Across all 340 Luce responses, 336 respondents said “mixed”, 4 said “no”, and zero said “yes”. The performance and heritage stories moved scale scores, but they did not flip a single respondent into a clean Ferrari verdict.
The component gaps explain why. In the next figure, each row is one part of the identity index. The gray dot is the neutral Luce gap to the pooled Ferrari reference average. The colored dot is the best-performing Luce framing in that segment. Positive values mean Luce is still behind Ferrari references on that identity signal.
Ritual continuity is the largest gap by a wide margin. In the broad panel, the neutral Luce was 1.24 points behind Ferrari references on preserved ritual. In the affluent panel, it was 1.17 points behind. That item is the reverse of “how much would losing the traditional Ferrari engine sound and mechanical feel diminish this car?” Respondents did not merely dislike the missing sound. They treated it as the central identity problem.
The next gaps were Ferrari distinctiveness, visual recognition, status signal, brand fit, and emotional pull. The Luce reads as expensive and modern, but the panel keeps asking the same question in different words: where is the Ferrari theater?
The open-ended responses made the numeric result less abstract:
“It’s a Ferrari, but without the engine noise, it’s missing something big.”
“It looks like a nice car, but if you took the Ferrari badge off, I’m not sure I’d recognize it.”
“It’s a step into the future for Ferrari, but it loses some of its soul along the way.”
“Four doors and electric? That ain’t the Ferrari I grew up dreaming about.”
“Where’s the roar? A Ferrari without the engine sound is like a burger without the meat.”
“It’s a beautiful car, but it feels like Ferrari is trying to be something it’s not with the four doors and electric engine.”
“Where’s the rumble? It just looks like another electric car, not a Ferrari.”
The concern map summarizes the open-ended responses rather than the ratings. Each row is a deterministic theme; the open circle is the broad panel share and the red dot is the affluent-GSS proxy share. The classifier is intentionally simple: keyword rules over first reactions, arguments against, and respondent quotes. It is not a second model call. Even with that caveat, the concentration is hard to ignore. Lost sensory ritual appeared in 86% of broad-panel Luce responses and 87% of affluent-panel Luce responses.
I would not report that as a human population percentage. I would use it as a research planning result: separate the next survey into sound, mechanical feel, silhouette, badge recognizability, and four-door practicality. Do not hide all of that inside one generic “brand fit” item.
What the launch stories changed
The framing test is where the story gets useful for marketing decisions. The figure below starts each row at the neutral Luce score and ends at the score under that launch story. The percent label shows how much of the neutral gap to Ferrari references the story closed. The text-only launch frame changed the identity index, but not the bottom-line verdict.
In the broad panel, heritage moved the index by +0.25 points versus neutral, closing 37% of the neutral Ferrari gap. Performance moved it by +0.21 points, closing 31% of the gap. LoveFrom design moved slightly backward at -0.02 points.
In the affluent-GSS proxy panel, heritage moved the index by +0.38 points, closing 55% of the gap. Performance moved it by +0.29 points, closing 42%. The LoveFrom design story again moved backward, this time by -0.10.
This is the most actionable result in the study. A Ferrari EV can borrow some identity from performance and lineage. A celebrity-design collaboration does not solve the Ferrari-ness problem, and in this run it made the car feel less Ferrari-like. The audience did not want an explanation of authorship. It wanted a reason to believe the sensory ritual had been replaced by something equally specific.
What this means
For product teams, the result is a design brief: do not treat the EV transition as a drivetrain substitution. The identity work sits in sound, tactility, silhouette, and ceremony. The panel is not asking for fake exhaust noise. It is saying that Ferrari needs a new ritual with the same emotional weight as the old one.
For marketers, the result says heritage and performance language are worth testing with real people. They moved the index by roughly a third to half of the neutral gap. That is not trivial. But the zero-yes verdict means the launch story cannot carry the product alone.
For researchers, the useful output is not the synthetic point estimate. It is the sharpened human study. I would now run a real concept test with three hard probes:
- Can performance framing beat heritage when respondents see richer product evidence?
- Is the LoveFrom penalty real, or is it an artifact of synthetic respondents reacting against outside authorship?
- Which replacement ritual, sound design, haptics, cockpit ceremony, charging ritual, launch sequence, or design signature, actually moves the Ferrari verdict from “mixed” to “yes”?
That is a better human survey than the one I would have written cold.
What I do not trust
The intervals are conditional on the model. They quantify variation across sampled personas and sampled model responses under one prompt and one model. They do not turn Gemini into a human panel.
The variance is also compressed. Ratings cluster more tightly than human Likert data usually does. That makes small differences look cleaner than they deserve. I trust the direction of the large gaps more than I trust the exact decimal places.
The affluent panel is not a Ferrari-buyer panel. It is an inspectable affluent proxy built from public-use U.S. survey rows. That is useful for a brand-identity stress test. It would be the wrong tool for pricing, purchase intent, or dealer allocation.
The theme classifier is blunt. It catches “engine”, “sound”, “roar”, “ritual”, and related language, so the sensory theme is broad by construction. That is acceptable for planning a follow-up instrument. It is not acceptable as a final estimate of public opinion.
The strongest limitation is validation. This study has no matched human sample. Without that, the right claim is modest: this is a structured, cheap, fast way to generate hypotheses before fieldwork. It is not fieldwork.
Closing
AudienceKit made the most important part of the experiment boring: load a public-use representative U.S. sampling frame, sample weighted rows, render default personas, run a fixed study, and analyze a table. In this post, that sampling frame is GSS 2024. In another study, it could be a different open survey, a customer panel, or a proprietary respondent file. That boring part matters because the Ferrari Luce result is subtle. The car did not fail outright. It beat the rival EV set on Ferrari identity and responded to heritage and performance framing. It also failed to earn a single clean “yes” on feeling like a real Ferrari.
That is the kind of result I want from a synthetic audience run. Not a verdict to ship, but a sharper argument for what to test next. For the Luce, the next test should stop asking whether electric can be Ferrari and start asking what new ritual could make electric feel inevitable inside Ferrari.