Industries / Finance and business services / Advertising and marketing / Market research
Reporting
Reporting is telling clients and the public what a study found and how it was done, so that others can judge how far to trust it.
5
Milestones achieved, from 1977 to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1977
One code worldwide
ESOMAR, the world association for market research, has developed its code of conduct with the International Chamber of Commerce (ICC) since 1977, including articles on transparency. The code, "established over 70 years ago," is recognized by over 60 associations in more than 50 countries, and was revised in 2016 and again in 2025.
Credit
ESOMAR and the ICC
2004
Rules for published polls
The British Polling Council, launched on 1 November 2004, requires members' published polls to state the client, interview dates, method, population covered and sample size, and to post full data tables on the member's website within two working days. Its rules now also require a statement on uncertainty with vote-intention polls: "there is a 9 in 10 chance that the true value of a party's support lies within 4 points."
Credit
British Polling Council
2009
A pollster publicly censured
AAPOR found the firm Strategic Vision in violation of its code after trying "for more than one year" to get basic information about its 2008 New Hampshire and Wisconsin primary polls, such as response rates and weighting. AAPOR stressed that it had not judged the polls' quality or accuracy, only the failure to disclose.
LessonDisclosure rules work only when someone asks and says in public who did not answer.
Credit
AAPOR Standards Committee
2014
Disclosure by pledge
AAPOR opened its Transparency Initiative to all survey organizations in October 2014. Members pledge to disclose their basic methods for public inspection, and AAPOR "makes no judgment about the approach, quality or rigor" of the methods disclosed.
Credit
AAPOR
2026
AI answers must be labeled
The 2025 ICC/ESOMAR code requires that the use of a "synthetic persona," a digital stand-in made to mimic real people, in collecting data be told to the person at the start. AAPOR's code, as amended by its members in June 2026, says AI-generated responses "are not research 'participants'" and must be identified as made by AI, and requires reports to say whether a sample was AI-generated and how the responses were produced.
Credit
ESOMAR and the ICC; AAPOR members
Watch for
Reports written by machine
Research software firms now sell reports drafted by AI. Research Live describes Listen Labs' platform as creating "reports, themes, highlight reels and decks" from its AI interviews, and Brad Anderson, president of products at Qualtrics, said in March 2026 that its customers "get insights in hours instead of weeks."
Achieved whenan independent check of such reports against their underlying data publishes how often their numbers and quotes are right.
Claimed by
Listen Labs, January 2026; Qualtrics, March 2026
Open
01
Audit "survey says" press releases
A running audit of press releases that cite commercial surveys run for publicity, recording for each whether it names the sponsor, field dates, how the sample was found, sample size, question wording and weighting, and whether it quotes a margin of error for an opt-in sample.
The briefCredit · unclaimed02
Check numbers against the tables
A checker that reads a research report and its data tables, links each number in the text to the table cell it comes from, and flags numbers that do not match and differences described as real that the data do not support. Test it on reports with planted errors, then on AI-written reports.
The briefCredit · unclaimed
Open 01 · the brief
Audit "survey says" press releases
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A running audit of press releases that cite commercial surveys run for publicity, recording for each whether it names the sponsor, field dates, how the sample was found, sample size, question wording and weighting, and whether it quotes a margin of error for an opt-in sample.
Why it is still open
British Polling Council rules let polls "conducted to support the PR activities of a commercial client" skip posting full tables within two working days, and AAPOR's guide for journalists says error margins "should not be reported for non-probability samples." Searches found no published audit of how such releases disclose their methods.
Built with what exists
Press release feeds such as PR Newswire and Business Wire; an AI model to extract the items, with yes/no checks run through Gemini Decision 1 (see the Note "A thousand questions, one pass"); two human coders on a sample; AAPOR's disclosure standards as the checklist.
First step
Collect 100 releases from one month that contain a phrase such as "survey of 2,000" and code them by hand.
Done looks like
At least 500 releases scored on at least 10 disclosure items, with the extraction agreeing with two human coders at least 90% of the time, and results by sponsor industry and research firm published with the data.
Open 02 · the brief
Check numbers against the tables
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that reads a research report and its data tables, links each number in the text to the table cell it comes from, and flags numbers that do not match and differences described as real that the data do not support. Test it on reports with planted errors, then on AI-written reports.
Why it is still open
AI-drafted reports are now sold, and AAPOR's May 2026 task force report on AI covers its use "from questionnaire design to interviewing, coding, analysis, and reporting" and sets out a framework for evaluating it. Searches found no published test of number accuracy in market research reports.
Built with what exists
Public reports that publish their tables, such as Pew Research Center reports with their toplines; a PDF table reader such as pdfplumber; current AI models as reviewers (the Note "Seven more flagged" describes an AI reviewer that caught problems people missed in drug-manufacturing records); Python.
First step
Take 10 Pew reports with their toplines, plant five numeric errors in each, and test a first version.
Done looks like
On 50 reports with 10 planted errors each, the share of errors caught and the false alarms per report, then the same figures for 50 AI-written reports, published with the code.