Industries / Finance and business services / Advertising and marketing
Market research
Market research is the work of finding out what buyers want and do, by asking them in surveys, interviews and group discussions, and by recording what they buy.
7
Milestones achieved, from 1911 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1911
Research becomes a department
Curtis Publishing, the Philadelphia publisher of the Saturday Evening Post and Ladies' Home Journal, set up a Division of Commercial Research in its advertising department and put Charles Coolidge Parlin in charge. His staff interviewed consumers, wholesalers and shop owners, and studied the contents of pantries and household trash to learn which brands people bought. The University of Pennsylvania, which holds the company's records, calls it "the first marketing research operation in the United States."
LessonMarket research began inside an advertising department, as a way to find where advertising would work for a given product and market.
Credit
Charles Coolidge Parlin, Curtis Publishing Company
1936
A big sample goes wrong
The Literary Digest magazine mailed about 10 million ballots, drawn largely from telephone directories and car registration lists, and got almost 2.4 million back. It predicted that Alfred Landon would win 54% of the popular vote; Franklin Roosevelt won more than 60% and every state but Maine and Vermont. George Gallup's American Institute of Public Opinion, polling a much smaller sample chosen to match the voting population, predicted a Roosevelt win. Sources disagree on the cause: Gallup blamed the mailing list for over-representing the well-off, later studies (Peverill Squire in 1988 among them) also blamed the 24% response rate, and Sharon Lohr and J. Michael Brick found in 2017 that weighting the returned ballots by how people had voted in 1932 would have picked Roosevelt.
LessonA bigger sample does not fix a list that leaves out part of the population, or replies that come mostly from one side.
Credit
The Literary Digest; George Gallup
1941
The focused interview
In November 1941 the sociologist Paul Lazarsfeld took his new Columbia University colleague Robert K. Merton to watch a test of a government radio program. About a dozen listeners pressed buttons to record what they liked and disliked, on a machine Lazarsfeld had built with Frank Stanton for testing CBS radio programs, and Merton then questioned the group about why they had pressed. Merton and Patricia Kendall described the method in 1946 as "the focused interview", and the "focus group" of later marketing research takes its name from it, according to Liza Featherstone's history of the method.
LessonCounting reactions shows what people liked; asking them afterwards shows why.
Credit
Robert K. Merton and Paul F. Lazarsfeld, Columbia University
1980
Test towns with scanners
Information Resources Inc. (IRI), founded in 1977 by John Malec and Gerald Eskin, launched BehaviorScan in January 1980 in Marion, Indiana, and Pittsfield, Massachusetts. IRI gave 15 supermarkets free checkout scanners, enrolled 2,000 households in each town, recorded the product code of everything they bought, and fitted devices to their televisions so that it could show different commercials to different households over cable and compare what each group then bought.
LessonWhen Lodish and six co-authors analyzed 389 BehaviorScan tests in 1995, they found that raising ad budgets relative to competitors did not in general raise sales, and that the standard recall and persuasion scores used to judge ads showed no strong link to sales.
Credit
John Malec and Gerald Eskin, IRI
2024
Opt-in polls checked
Opt-in online panels recruit people through ads and social media instead of choosing them at random, and some members answer carelessly to collect rewards. In a February 2022 experiment by Pew Research Center, 12% of opt-in respondents under 30 said they were licensed to operate a nuclear submarine. When a December 2023 opt-in poll reported that 20% of US adults under 30 agreed that "The Holocaust is a myth", Pew asked the same question in January 2024 on its American Trends Panel, which is recruited by mail through random sampling, and found 3%.
LessonRare or alarming results from opt-in samples need checking against a randomly drawn sample before they are reported.
Credit
Pew Research Center
2025
An agent passes the checks
Sean J. Westwood of Dartmouth College built an AI agent that takes online surveys while keeping an assigned persona and a memory of its earlier answers. In his tests it passed 99.8% of 6,000 standard attention checks, the trick questions surveys use to catch careless or automated respondents, and he put the cost of one completed survey at about $0.05 with commercial AI models. He published the code for generating answers but withheld the scripts that submit them, to prevent misuse.
LessonA coherent set of answers is no longer evidence that a person gave them.
Credit
Sean J. Westwood, published in PNAS in November 2025
2026
Simulated samples fall short
In 2023 Lisa P. Argyle and five co-authors reported that the AI model GPT-3, given the backstories of real survey respondents, matched the answer patterns of many US groups, and called such stand-ins "silicon samples". In 2024 James Bisbee and four co-authors found that ChatGPT's answers to a US election survey varied less than real ones and shifted with small changes in wording. Pew Research Center then built an AI "digital twin" of each of thousands of members of its American Trends Panel and ran the twins on three survey waves fielded between January and April 2026. Across nearly 300 questions the AI estimates differed from the real ones by an average of 12 percentage points, and on nearly half of the questions at least one answer was chosen by no AI respondent.
LessonPew concluded that AI models "are not an adequate replacement for traditional polling on topics of broad public importance.".
api.crossref.orgapi.crossref.orgpewresearch.orgpewresearch.org
Credit
Argyle et al. and Bisbee et al., in the journal Political Analysis; Pew Research Center
Watch for
Synthetic shoppers rate products
Researchers at PyMC Labs and Colgate-Palmolive say AI models asked to answer as consumers reproduced the purchase intent recorded in 57 personal-care product surveys with 9,300 human responses. Their method has the model write a short answer in its own words and then places that answer on a five-point scale by comparing it with reference statements. By their count it reached 90% of test-retest reliability, which is how closely the human survey agrees with itself when repeated.
Achieved whenan independent team reproduces the result on product surveys the authors did not choose, or synthetic ratings made before a human survey match its results.
Claimed by
PyMC Labs and Colgate-Palmolive, October 2025
A detector catches every bot
Prolific, a firm that recruits paid participants for online studies, says its "bot authenticity checks", launched in February 2026, caught all 125 survey runs by five AI agents, including an in-house agent built to avoid detection, and flagged none of 125 human participants. In the same test, paired consistency questions caught 30% of agents.
Achieved whenan outside group tests the check against agents it builds itself and publishes both the catch rate and the false-alarm rate.
Claimed by
Prolific, February 2026
Open
01
Forecast trackers before release
A public, timestamped set of forecasts by synthetic respondents (AI models asked to answer as a described kind of person) for survey questions that are asked again on a schedule, filed before each new wave is published and scored once it is, so that a running record shows whether the forecasts improve.
The briefCredit · unclaimed02
Count survey bots both ways
One study that runs the two published methods for spotting AI agents in online samples side by side, on the same respondents from the same platforms, and reports where they agree and where they differ.
The briefCredit · unclaimed
Open 01 · the brief
Forecast trackers before release
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, timestamped set of forecasts by synthetic respondents (AI models asked to answer as a described kind of person) for survey questions that are asked again on a schedule, filed before each new wave is published and scored once it is, so that a running record shows whether the forecasts improve.
Why it is still open
Most published tests score synthetic respondents on results that already exist, which a model may have seen in training. Pew Research Center's 2026 test avoided that by running AI stand-ins on its own new survey waves, and found them off by 12 percentage points on average, but it was a one-time test of three waves run inside Pew. Ashwini Ashokkumar, Luke Hewitt, Isaias Ghezae and Robb Willer (Nature, July 2026) used an archive of 70 past survey experiments, including some not public before the model's training cutoff, and found that the predictions overstated effect sizes. Searches found no public, timestamped record of forecasts for a recurring tracker filed before each release.
Built with what exists
YouGov's public trackers, which publish the wording and results of its repeated surveys; an AI model that returns a probability for each listed answer, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); a public GitHub repository or the Open Science Framework (OSF) to timestamp each forecast; Python for scoring.
First step
Pick 100 YouGov tracker questions that update at least monthly, record their last three waves, and post the first round of forecasts before the next release.
Done looks like
Forecasts for at least 100 tracker questions, each filed before release, scored against the published result and against the simple forecast that the next wave will equal the last one, published with every prompt and the scoring code.
Open 02 · the brief
Count survey bots both ways
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
One study that runs the two published methods for spotting AI agents in online samples side by side, on the same respondents from the same platforms, and reports where they agree and where they differ.
Why it is still open
Two 2026 studies disagree on how common the agents are. Grace Zhang and eight co-authors (PNAS, October 2026) found that 6% to 41% of respondents failed their AI tests across seven platforms, against 2.4% of people tested in person. Andrew Gordon and six co-authors (a preprint, April 2026) recruited 5,200 respondents from 10 platforms and found agent detections at or below 1% everywhere except Amazon Mechanical Turk (11% to 16%). No study found has run both sets of tests on the same respondents.
Built with what exists
Qualtrics or LimeSurvey; the test items in each paper and its supplements; paid samples from five or more panel providers, bought at standard rates; for checking the tests in the lab, a browser agent such as Browser Use (see the Note "Deleting 80% to start over"), run only on the team's own unpaid test survey and never on a paid study.
First step
Write to both author teams for their test materials, and pre-register the comparison on OSF.
Done looks like
For each of at least five platforms and 400 respondents per platform, the share flagged by each method with 95% confidence intervals and the share flagged by one method but not the other, with data and code published.