Summary
- Fast. Six days in field, and every ad live the whole time. No recruitment, no panel, no questionnaire.
- Low-cost. $300 of media per ad, $9,760 spent in total, for 4,428 observed responses from the target population itself.
- Reaches the unreachable. Market Experiments run on ad infrastructure, so the audience can be defined by any characteristics targetable by digital ads, including hard-to-reach groups such as those with income above $1M, assets above $50M, home owners, and executives at named companies. These are the people who sit on no panel, which is why survey-based segmentation has never reached them.
- Captures the emergent properties of the marketplace. Survey segmentation clusters answers. This clusters behaviors. One undifferentiated audience saw thirty-four value propositions, and the segments are whatever structure the response had. A crowd nobody had named appeared, and an assumed one dissolved.
- Actionable on arrival. Segments are profiled on the ad platform's own audience labels. Each one arrives as a targeting definition the media team can load as written. Insight and campaign share a vocabulary.
Segmentation without a panel
Classic segmentation asks a panel a battery of attitude statements, clusters the answers, and hands back personas. A media team must approximate those personas with the targeting on hand. No matter how good the execution, three structural problems remain:
- Panels reach people willing to take surveys for a reward, and for the most part the affluent decline these offers.
- What people say diverges from what they do. The selection that decides who clicks an ad is the selection that decides who buys.
- Every firm in a category buys the same panels, so panel insight is shared insight, not “competitive advantage.”
For a luxury brand, or any brand whose customers are hard to reach, the first problem is fatal: the customer who matters most is the one a panel cannot seat.
Market Experiments run the same study on behavior. Each attitude statement becomes an unbranded ad. The ads run on live ad networks against an audience defined with ad targeting, and the response arrives as clicks. The sample is whoever an advertiser can reach, and an advertiser can reach the wealthy. Think of it as a blinded, in-market version of the classic study, with the segments read off who responded.
Who a Market Experiment can reach
A sample of audience filters available on the platform. The sneaker study fielded on a broad interest audience: US adults 18 to 54 with fashion and sneaker signals.
Income
Wealth
Home ownership
Professional
Education
Life stage and geography
Available filters can vary by geography. Any combination defines the population an experiment fields against, and the same labels return in the segment profiles.
The question
The retailer believed its market split into three shopper mindsets: personal expression, hype culture, and status. Before committing a season of media to that model, it wanted to know whether those mindsets exist as distinct audiences in the market, and what each one responds to. A pool design, testing messages inside three pre-built audiences, could rank those three beliefs and nothing more. If the market divided along a line nobody had drawn, a pool design would miss it.
How it ran
1
Statement
Seventeen paired attitude statements across three mindsets.
2
Ad
Each side of each pair becomes its own headline. 34 ads, nothing else varies.
3
Guide
Every ad opens the same unbranded category guide. Nothing is sold.
4
Response
One audience, equal budgets, six days. 4,428 clicks observed.
5
Profile
The platform attaches age, gender, region and 13 audience labels to each response.
6
Segment
Ads whose responders look alike cluster. The clusters are the segments.
The spec
Population
US adults 18 to 54 with fashion and sneaker interest signals, on Google Display placements across sites and apps. One audience, no pools.
Technique
Response-clustering segmentation. Three mindset dimensions, seventeen statement pairs, both sides of every pair fielded as ads.
Profile read
Age, gender and state from demographic reporting. Eight interest and five purchase-intent segments attached in observation mode, so they report without steering delivery.
Cells
34 ad versions, both sides of all seventeen pairs, no audience splits.
Budget
Equal lifetime budget per version, $300 each. $9,760 spent in total.
Field period
6 days, every version live the whole time.
Output
Response-defined segments with profiles, the ads each one owns, and pairwise trait tests. The experiment produces the segments; nothing feeds them in.
Destination
One unbranded editorial guide, “Sneakers in 2026: culture, comfort, and the chase”. No brand named, none sold.
The design matrix
Abbreviated: 8 of 34 rows, the statement as a survey would ask it and the headline it became.
| Version | Mindset | The statement | The ad it became |
|---|---|---|---|
| b1A | Personal expression · A | “What I wear says a lot about who I am” | What I wear says who I am |
| b1B | Personal expression · B | “What I wear is mostly about what I need for the day” | It just needs to work today |
| f3A | Hype culture · A | “I'm willing to wait in line or go out of my way to get something when it drops” | Worth the line when it drops |
| f3B | Hype culture · B | “I'm fine getting things later, even if they've been out for a while” | It'll still be there next month |
| f5A | Hype culture · A | “Sneakers are meant to be preserved” | Sneakers are meant to be preserved |
| f5B | Hype culture · B | “Sneakers are meant to be worn” | Sneakers are meant to be worn |
| i5A | Status · A | “Luxury brands define status” | Luxury defines status |
| i5B | Status · B | “Being connected to culture or scene defines status” | The scene defines status |
Every version carries the same $300 lifetime budget and the same six days in field. Delivery optimizes toward clicks inside that budget, so impressions vary; pull rescales each ad's response rate to the delivery-weighted field average at 100.
The clustering reads no ad copy and no source statement. Two ads land together when the same kinds of people respond to both. Each count is modeled as a binomial draw with a flat prior, so every share, index and pull carries a posterior and a 95% credible interval. A trait joins a segment's signature when the posterior probability that its share sits above the field's reaches 97.5% and the gap at the median is five points or more.
What came back
The thirty-four ads drew 4,428 responses. Clustering the responder profiles yields four crowds, three with a targeting profile that clears the bar.
The drop chase
44%
Men (62%), 18 to 24. Over-index: basketball fans, luxury shoppers, rap and hip hop fans; in-market for luxury goods and athletic shoes.
“Worth the line when it drops” Exclusivity, visible status, the release calendar. The chase is the product.
Style authors
32%
Women (55%). Over-index: art and theater aficionados, fashionistas. No intent segment separates.
“Getting dressed is self-expression” Expression first, subtle status, culture over logos. Rejects hype mechanics.
Comfort first
16%
Women (58%), 35 to 54, South-leaning. Over-index: running enthusiasts, health and fitness buffs, value shoppers; in-market for fitness.
“It just needs to work today” Function, availability, ease. The crowd hype messaging misses.
On their own terms
9%
Separates from the field only by what it lacks; no trait clears the bar.
“You are who you are” Unmoved by scarcity and unbothered by the room. The loosest cluster.
Shares carry 95% credible intervals and no two neighboring intervals overlap, so the order holds.
Exhibit 1, abbreviated. Interests and purchase intent separate all four crowds; age, gender and region separate comfort first, and little else. Each cell indexes a segment's share on one trait against the field average at 100; the small figure is the share itself. Orange cells reach 97.5% posterior probability of sitting above or below the field, with a gap of five points or more. Read down a column for the segment's signature. Fifteen of the twenty-two traits are shown.
| Trait | Field | The drop chase | Style authors | Comfort first | On their own terms |
|---|---|---|---|---|---|
| Who they are | |||||
| Women | 48% | 8038% | 11655% | 12158% | 10651% |
| 18 to 24 | 33% | 12241% | 10636% | 4314% | 6923% |
| 35 to 44 | 22% | 8519% | 9321% | 14332% | 12127% |
| 45 to 54 | 10% | 727% | 889% | 19019% | 11912% |
| South | 27% | 9425% | 9024% | 14138% | 9024% |
| What they are into | |||||
| Fashionistas | 38% | 10741% | 12046% | 5722% | 7328% |
| Luxury shoppers | 21% | 13829% | 8819% | 4810% | 459% |
| Value shoppers | 18% | 7614% | 9016% | 17732% | 11521% |
| Running enthusiasts | 12% | 779% | 739% | 21426% | 10412% |
| Basketball fans | 23% | 14133% | 5613% | 8319% | 8319% |
| Rap & hip hop fans | 24% | 13031% | 9523% | 4411% | 6816% |
| Art & theater aficionados | 14% | 639% | 17425% | 619% | 9213% |
| What they are shopping for | |||||
| Athletic shoes | 31% | 12338% | 7423% | 10533% | 6922% |
| Luxury goods | 15% | 14422% | 8212% | 365% | 579% |
| Fitness products & services | 12% | 9111% | 708% | 20324% | 708% |
27 of 52 interest and intent cells clear the bar, against 11 of 36 demographic cells. Age and gender shares leave out the 31% of responders the platform reports as unknown. One person can sit in several interest or intent segments, so those shares do not add to 100.
One ranking, four different audiences inside it
If the market were one audience, the top of this chart would be one color. It is not. The two best ads belong to different crowds, and the top nine are split between them.
Exhibit 2. Thirty-four ads ranked by pull, colored by the segment that owns each.
Worth the line when it drops
136
Getting dressed is self-expression
136
They notice how it comes together
135
Trending or not, I wear what I like
132
Buy the styles you like
129
Hard to replicate is the point
126
What I wear says who I am
124
Drawn to the hard-to-get
122
The scene defines status
121
You are who you are
117
Luxury defines status
115
It just needs to work today
114
Style that works for the room too
114
Logos signal status
112
You are what you have
109
Limited drops make it mean more
109
Comfort defines style
108
Apps, raffles, stores. Whatever it takes
106
Low-key pieces signal status
106
Style choices are mine alone
104
If it's gone, I move on
100
I keep an eye on what's trending
95
Keep it simple, keep it easy
92
Availability doesn't change the shoe
90
Great style is worth a little discomfort
89
Great pairs, easy to find
88
The great outfit is worth the effort
88
They know exactly what I'm wearing
81
Buy what it could be worth later
79
Sneakers are meant to be worn
77
Hard to get is the point
74
Sneakers are meant to be preserved
71
Function first, every morning
70
It'll still be there next month
70
Pull rescales each ad's response rate to the delivery-weighted field average at 100 (the vertical rule), as a posterior median with a 95% credible interval (the whisker). Color names the segment whose responders the ad drew. Delivery optimizes toward clicks within each ad's equal budget, so impressions vary; pull removes that.
What emerged that a persona study would have missed
- The three assumed mindsets were not the map. Status split in two. “Luxury defines status” drew the drop chase; “The scene defines status” and “Low-key pieces signal status” drew the style authors. A persona study would have delivered a status persona because it was asked for one.
- A fourth crowd nobody named. Comfort first is 16% of response: older, running and fitness-led, the audience hype messaging misses. The ads for function and availability found it. Nobody had hypothesized it.
- The two sides of one belief land in different crowds. “Sneakers are meant to be preserved” clusters with the drop chase; “Sneakers are meant to be worn” clusters with the style authors. Same product, opposite mindsets. The clustering saw neither headline.
- The overall winner covers less than half the market. “Worth the line when it drops” leads the whole field, but its responders are the drop chase. A campaign built on the top-ranked ad would speak to 44% of response and leave the rest unclaimed.
Built on targeting rails
The profiles are made of the platform's own affinity and in-market labels, so each segment is handed over as a targeting definition the media team can load without translation. Spend, flighting, and creative remain theirs. The recommendation: three campaigns and one re-field.
The drop chase: “Worth the line when it drops”
Affinity Basketball fans, Luxury shoppers, Rap & hip hop fans; in-market Luxury goods, Athletic shoes; men 18 to 24.
Style authors: “Getting dressed is self-expression”
Affinity Art & theater aficionados, Fashionistas; women.
Comfort first: “It just needs to work today”
Affinity Running enthusiasts, Health & fitness buffs, Value shoppers; in-market Fitness; women 35 to 54, weighted to the South.
On their own terms: re-field before targeting it
Four ads, no trait clearing the bar. Treat it as a lead.
Evidence
Exhibit 3. Do two segments differ on a trait? Share of women, pairwise.
| Pair | Share A | Share B | Difference | P(differ) |
|---|---|---|---|---|
| The drop chase vs. Style authors | 38% | 55% | −17 pts | 1.00 |
| The drop chase vs. Comfort first | 38% | 58% | −19 pts | 1.00 |
| The drop chase vs. On their own terms | 38% | 51% | −12 pts | 1.00 |
| Style authors vs. Comfort first | 55% | 58% | −2 pts | 0.87 |
| Style authors vs. On their own terms | 55% | 51% | +4 pts | 0.87 |
| Comfort first vs. On their own terms | 58% | 51% | +7 pts | 0.97 |
Beta-Binomial posteriors on each segment's known-gender base, 380 to 1,950 responses. Two segments differ when the probability in one direction reaches 97.5%. Comfort first and On their own terms sit at 0.97: probable, and short of the bar.
Evidence table, abbreviated. Every fielded ad carries spend, delivery, response, and pull.
| Ad version | Segment | Spend | Impressions | Resp. | Pull (95% CrI) |
|---|---|---|---|---|---|
| f3A · Worth the line when it drops | The drop chase | $280.62 | 58,350 | 186 | 136 (118–157) |
| b2A · Getting dressed is self-expression | Style authors | $280.68 | 34,570 | 110 | 136 (113–164) |
| i2B · They notice how it comes together | Style authors | $283.07 | 36,890 | 117 | 135 (113–162) |
| b5A · Trending or not, I wear what I like | Style authors | $294.91 | 32,560 | 101 | 132 (109–161) |
| f6B · Buy the styles you like | Style authors | $290.08 | 46,090 | 139 | 129 (109–152) |
| i1B · Hard to replicate is the point | Style authors | $285.97 | 53,700 | 158 | 126 (107–147) |
Bases run 81 to 186 responses per ad, all above the 25-response suppression line. Single-ad cells carry wide intervals; segment-level signatures pool 380 to 1,950 responses and are the quotable read.
A few notes about this field. One boundary between segments is a judgment call. The farthest pair inside one segment sits at 0.078 and the nearest pair across segments at 0.056, so the drop chase and style authors touch at one point. No lead message has separated from its runner-up. The segments stand; the winning line inside each is the next experiment. Behavior shows what pulls, and is triangulated with qualitative research and domain expertise.
Why this matters
- It reaches the people who do not take surveys. Define an audience by income, wealth, home ownership, seniority or employer, then study what it clicks. No panel stands in between. For luxury brands, and for anyone else whose customers are hard to reach, the people most worth segmenting are the ones no panel can seat.
- Luxury intent sits inside the hype crowd, and status still splits. Luxury intent indexes 144 in the drop chase and below field everywhere else. Yet “Luxury defines status” and “The scene defines status” pulled different crowds. Which status story a brand tells is a segmentation decision, and this method measures it.
- It works globally. The same design fields on any platform with audience reporting, in any market, and hands back segments that are comparable because they are built from the same labels.
- It costs a fraction of the alternative. This study bought 4,428 observed responses for $9,760 of media. A panel study of affluent consumers pays a premium for each completed interview, seats few of the people it wants, and still returns statements in place of behavior.
- Confidence is computed, then stated. Power is set before fielding, so each study says in advance which claims its design can support. Flashpoint.AI establishes confidence with Bayesian inference.
- It is fast. Six days from first impression to targetable segments. No recruitment wave, no questionnaire to program, no fieldwork report to wait on. A question raised on a Monday can have a segment map before the next planning meeting.
- It can measure purchase. Add affiliate tracking links to the destination and the study reads what people buy, not only what they click.