Path to Purchase Research: Where the Decision Happens

Path to purchase research asks at home about a choice made at the dealership or the shelf. What the place changes, and what simulated buyers can carry.

A path to purchase study tries to rebuild how a buyer got from "we need another car" to a signature at the dealership. Most of the time it does that with a questionnaire answered on a phone, at home, weeks after the purchase, by someone remembering a Saturday spent in a showroom. Path to purchase research is good at recording what people planned and what they remember doing. The decision itself happened somewhere else, with a salesperson in the room, a trade-in offer on the desk and a model on the floor that was never on the shortlist.

Key Takeaways

• Path to purchase research maps the steps between a need and a purchase. Most studies rebuild those steps after the fact, away from the place where the choice was made.
• In Cox Automotive's study of 2,300 recent vehicle buyers, fielded in the fall of 2025, just 29 percent were certain of the vehicle when they started shopping. And 53 percent completed every step in person at the dealership.
• POPAI's 2012 study interviewed 2,400 shoppers before and after the trip and classed 76 percent of supermarket purchase decisions as made in the store.
• Agents built from two-hour interviews with 1,052 Americans reached 83 percent of the participants' own two-week test-retest consistency on held-out survey items. The benchmark is made of questionnaire answers, none collected at a counter.
• A simulation of buyers inherits the place its material came from, so its job sits between fieldwork waves, choosing which point of sale questions deserve real buyers.

What path to purchase research measures

Path to purchase research is the family of studies that maps the route from the moment a need appears to the moment the buyer commits. It covers what set the need off, which sources the buyer read, which brands they weighed, where they went and why they chose what they chose.

Two more terms carry the weight in any debrief on the subject. A consideration set is the short list of brands or models a buyer seriously weighs before choosing. The point of sale is the place where the purchase is closed, whether a dealership floor, a supermarket aisle or the counter of a store in a mall.

Cox Automotive has run its car buyer study every year for sixteen years. The latest edition, published in January 2026, surveyed 2,300 people who had bought a new or used vehicle in the previous 12 months, with fieldwork in the fall of 2025. It is a familiar design. The respondent is a recent buyer, the instrument is a questionnaire, and the decision is reconstructed from memory.

How a path to purchase study is run today

PERSONAL EXPERIENCE In the studies I have run and reviewed over twenty years, the sequence barely changes. The brand defines the category and the purchase window. The agency writes a questionnaire that walks the buyer backwards from the purchase. When did you start looking? Which sites did you visit, which brands did you consider, how many dealers did you go to? What made you decide? The panel recruits recent buyers, and they answer on their phones, at home.

If the budget allows, a qualitative module adds a handful of accompanied shops or interviews at the dealership exit. Those usually end up as quotes in the appendix. The deck draws a funnel, each stage gets a percentage, and the stage at the point of sale is described with the same instrument as the stage on the sofa.

Most categories still run on that recall. It is cheap to field, easy to repeat every year, and it produces a clean chart. What it cannot produce is the moment when the plan met the floor.

Why the place changes the answer

Because a large share of the decision is not made before the buyer arrives. In the same Cox Automotive study, just 29 percent of buyers were certain of the vehicle they wanted when they started shopping, down from 37 percent in 2020. The other 71 percent went in with an open mind. And 53 percent completed every step of the purchase in person at the dealership, against 7 percent who bought entirely online.

The supermarket has a longer record, because there the gap was measured directly. The in-store decision rate is the share of purchases that the shopper did not name by brand before walking in. It adds up items planned only as a category, items bought on impulse, and items swapped for another brand at the shelf. POPAI, the trade association for point of purchase advertising, ran its 2012 Shopper Engagement Study with pre and post shopping interviews. It reached 2,400 shoppers across the four US census regions, between the fall of 2011 and March 2012. The in-store decision rate came out at 76 percent, up from 70 percent in 1995.

The choice was still open at the door

Two panels from two published studies, each with its own unit. Panel one, vehicle buyers, in percent of buyers: 37 percent were certain of the vehicle they wanted when they started shopping in 2020, and 29 percent in 2025; in 2025, 53 percent completed every step of the purchase in person at the dealership and 7 percent bought entirely online. Source: Cox Automotive, 16th annual car buyer study, published January 2026, fieldwork in the fall of 2025 with 2,300 people who bought a new or used vehicle in the previous 12 months. Panel two, supermarket trips, in percent of purchases decided in store: the in-store decision rate was 70 percent in 1995 and 76 percent in 2012. Source: POPAI, 2012 Shopper Engagement Study, pre and post shopping interviews with 2,400 shoppers from the fall of 2011 to March 2012; POPAI is a trade association for point of purchase advertising. A closing line says a questionnaire answered at home records the choice after it closed. Every number belongs to Cox Automotive or POPAI and none of them is a Cassi.ai measurement.

Most vehicle buyers had not settled on a model when they started, and most still closed the deal on the floor. The questionnaire meets them afterwards, at home.

Two cautions. POPAI speaks for the people who sell in-store displays, so it has an interest in a high number. And a weekly grocery trip is not a car purchase. What carries over is the method. The gap only became visible because someone asked at the door and checked again at the exit.

The figures behind the rows come from Cox Automotive and POPAI. The split between the two columns is our own reading.

Who answers for the gap

The brand team decides the year from one wave, and the insight lead answers for that decision. A path to purchase study is expensive enough to run once a year. The category plan, the dealer training and the launch sequence all lean on a single snapshot taken from the sofa.

PERSONAL EXPERIENCE The questions that hurt arrive in between. Will the buyer who came for the gas version listen if the salesperson leads with the hybrid? What happens when the model in the ad is out of stock that weekend? Does a large screen on the dashboard change the conversation on the floor? None of these justifies a new wave on its own. So the company still answers them the way it always has, with the opinion of whoever is senior in the meeting. It finds out at the next wave whether the bet was right.

The cost of a wrong bet at the point of sale is not visible in the research budget. It appears as a launch that underperforms in the showroom while every chart in the deck looked fine.

What simulated buyers can carry, and what they cannot

A synthetic persona is a simulated respondent that a language model generates from data about real people, and that answers questions in their place. A generative agent is a synthetic persona that also keeps a memory of its background and of the conversation so far. Its answers stay consistent from one question to the next.

The research behind them is serious and recent. Algorithmic fidelity is the degree to which a language model, conditioned on the backgrounds of real people, reproduces the answer patterns of the groups those people belong to. The term comes from Argyle and colleagues in Political Analysis, who built "silicon samples" by conditioning GPT-3 on thousands of backstories from participants in large US surveys.

The clearest result so far comes from Park and colleagues, who built agents for 1,052 Americans from two-hour semi-structured interviews, from structured surveys, or from both. On held-out General Social Survey items, the interview agents reached 83 percent of the participants' own two-week test-retest consistency, the survey agents 82 percent and the combined agents 86 percent. Agents built from demographics alone reached 74 percent.

UNIQUE INSIGHT Look at what the benchmark is made of. The yardstick is the same person answering the same questionnaire again two weeks later. It measures how well an agent reproduces what someone says about themselves when asked at home. That is a real achievement, and it is also the sofa again.

The marketing evidence points the same way. In a conjoint study run with GPT-3.5 Turbo, Brand, Israeli and Ngwe found that the model reproduced consumers' willingness to pay for features that already existed in a category. For new features it needed fine-tuning on earlier human results, and that fine-tuning may not carry to unrelated categories.

A simulation inherits the place where its material was collected. The same limit shows up when synthetic cases fill a thin subgroup, as in synthetic sample augmentation: the model can only return what its material contains. Built only from answers given at home, it answers from home. Built with what was observed at the point of sale, the accompanied shop, the exit interview, the salesperson's account of the objections, it can start to rehearse the point of sale.

Rehearse the hypothesis between two waves

The simulator belongs between fieldwork waves. Its job is to take the tactical questions that pile up after a study, rehearse them with personas placed where the decision happens, and pick the few that deserve real buyers. This is the sequence we recommend.

1. Collect point of sale material in the wave you already run: accompanied shops, exit interviews, notes from the sales floor.
2. Build the personas from that material and the questionnaire together, never from demographic fields alone.
3. Place them in a model of the setting where the choice is made, the showroom, the aisle, the store.
4. Rehearse each tactical question as a scenario, with the researcher asking and the personas answering on the spot.
5. Rank the hypotheses by how far the personas move and how much the brand would change if they were right.
6. Take the top few to field, with real buyers at the point of sale.
7. Feed the new wave back into the personas before the next round of questions.

Where the simulator sits: between two waves

A flow in five steps, with a note to the right of each. The simulation chooses the questions and the fieldwork answers them. Step one, a wave of fieldwork: the questionnaire at home plus material from the point of sale; the note says what to collect on the floor, accompanied shops, exit interviews and notes from the sales floor. Step two, build the personas from the studies and the point of sale material together; the note says to leave out personas built from demographic fields alone. Step three, rehearse in the setting: each tactical question becomes a scenario in the showroom or the aisle; the note says the researcher walks in and the personas answer on the spot. Step four, rank the hypotheses by how far the personas move and what the brand would change; the note says questions that move nothing do not go to field. Step five, the next wave of fieldwork: real buyers at the point of sale test the top few; the note says the results go into the personas before the next round. A closing band says personas cannot rehearse a place nobody observed, so step one decides what the simulation can say. The figure contains no measured numbers.

The simulator works in the gap between waves. What it rehearses is limited by what the last wave collected on the floor.

Step one decides the rest. Personas cannot rehearse a place nobody ever observed.

Where Cassi.ai comes in

@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights, founded by Cassiano Albuquerque after twenty years in market research. We automate research projects with or without language models.

Synthetic Vertical Simulator is a Cassi.ai product that rebuilds the client's consumers as synthetic personas from the client's own data and places them in a virtual copy of the environment where the decision happens.

The Synthetic Vertical Simulator rebuilds a client's consumers as synthetic personas, from the client's own data and studies. It puts them inside a virtual copy of the place where they decide: a mall, a supermarket, a plaza, a dealership, a beach, an office. The researcher walks in, stops any persona and asks. Clusters, settings and activities are built for each client, with the methods the study calls for.

It does not replace the wave. It decides what the next wave should ask. How a persona built from real people differs from a demographic sketch is covered in customer digital twins in market research. The evidence on screening concepts this way is in synthetic respondents in concept testing.

FAQ

What is path to purchase research?

A family of studies that maps how a buyer goes from a need to a purchase: triggers, sources, the brands considered, the visits and the final choice. Most of it is collected after the fact, as in the Cox Automotive study of 2,300 recent vehicle buyers.

Why do survey answers differ from what happens at the point of sale?

Because many decisions are made there. Cox Automotive found that just 29 percent of vehicle buyers were certain of the vehicle when they started shopping. POPAI classed 76 percent of supermarket purchase decisions as made in the store.

Can synthetic personas replace path to purchase fieldwork?

No. The best published agents reproduce what people say about themselves on a questionnaire, measured in Park and colleagues' study of 1,052 Americans. They help choose which questions go to field, and the field validates.

What should a simulation of buyers be built from?

From the studies the brand already has, plus material collected at the point of sale. Agents built from demographics alone performed worst in the published benchmark.

Before the next wave

UNIQUE INSIGHT A path to purchase study that only asks at home measures the plan. The decision belongs to the place, and the place has to be in the material, whether the buyer is real or simulated.

Open the last path to purchase deck your team approved and count the stages measured at the point of sale with something other than recall. If the count is zero, the next wave already has its first question. If you want to see the questions in between rehearsed in your own category, we build the personas from your studies and the setting where your buyer decides.

About the author

Cassiano Albuquerque is the founder of @Cassi.ai. He spent twenty years in market research operations before building software for research teams. To talk about simulating the point of sale in your category, contact him on LinkedIn.

Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/path-to-purchase-research-context.