Deep Reads: Who Vouches for the Answer
Somebody forwards you a finding this week. A number, a chart, a line in a deck, with a question attached. What do you make of it?
0:00 / 10:25
Deep Reads: Who Vouches for the Answer
This week's three reads were published between 13 and 16 October 2026, and none of them is about a better way to produce a finding. Each one is about what has to hold up around a finding once it leaves the room it was made in. A respondent who cannot report what moved them. A supplier who now has to say where its own work could be wrong. A company where nobody can see the whole of the research being done. Underneath all three sits one question: when nobody in the room saw the thing happen, who vouches for the answer?
What people can't report
William Leach writes on behavioural science, and before that he ran shopper marketing and a retail testing lab at Frito-Lay. In a column published on 16 October he takes three of the most studied influences in the field — social proof, scarcity and authority — and argues that most insights teams build studies that cannot detect any of them. 1
The reason is simple. People do not report these forces. Ask a shopper why they chose a brand and you will hear taste, price and ingredients, not that everyone else seemed to be buying it.
His example is a field experiment on household energy use in California. Households received doorhangers with different appeals: protect the environment, save money, be socially responsible. One simply said that most of their neighbours were already conserving. That message changed behaviour more than any other, and residents rated it last when asked what would motivate them. A companion survey of 810 Californians found beliefs about what neighbours were doing predicted conservation better than any other belief, and respondents ranked that belief least important. 1
Leach then does what most columns avoid: he goes looking for the replication that failed. The hotel towel studies found that telling guests most other guests reuse their towels lifted reuse to 44.1%, against 35.1% for the standard environmental appeal. Bohner and Schlüter repeated the work in two German hotels, and the norm message did no better than the standard one. Social proof is real; how far it moves people depends on context, culture, and how close the "other people" feel. 1
On scarcity he recovers the detail most retellings drop. In Worchel, Lee and Adewole's 1975 cookie study, cookies from a jar of two were rated more desirable than identical cookies from a jar of ten, and each participant saw only one jar. What mattered was why the jar was low: cookies that had dropped from ten to two beat cookies that were scarce all along, and scarcity created by other participants taking the cookies beat scarcity that happened by accident. 1
Authority works more slowly than researchers assume. In an fMRI study by Klucharev, Smidts and Fernández, a single pairing of a product with an expert produced a long-lasting positive effect on both memory for the product and attitude toward it — the expert changed what people stored, not just what they felt in the moment. A preference question asked immediately after exposure will therefore understate the effect. 1
Two things in the piece date it to 2026. AI shopping assistants now read ratings, review counts, expert badges and stock flags before a shopper sees the page, so the cue still works but often arrives secondhand. And the fake version is now a legal problem in the United States: the Federal Trade Commission's rule on consumer reviews, in force since October 2024, bans fabricated reviews including AI-generated ones and bans buying or selling fake social indicators. 1
The boundary to keep. This is a practitioner's argument column, not a study. The experiments it cites are real and reported carefully, including the one that failed to replicate. The design advice — test the cue as a stimulus in matched cells, measure behaviour or a close proxy, and never ask respondents whether a bias affects them — is the author's judgment built on that literature.
What must be declared
Philippe Guilbert sits on Esomar's Professional Standards Committee. His 14 October piece maps the three frameworks that actually govern a piece of research: law, technical standards, and codes of conduct. 2
The order is not decoration. The ICC/Esomar Code's first fundamental principle is that all research must be legal, and he is explicit that no technical or ethical standard justifies breaking a legal obligation. It is also the law of the territory rather than of your head office: European data protection and the AI Act, plus national rules such as bans on publishing polls during an election campaign, healthcare obligations, and prior notification of staff for mystery shopping.
The technical layer has just moved. ISO 20252, the international standard for market, opinion and social research, was published in September 2026. Certification means documenting procedures and submitting to an external audit, repeated after three years.
The ethical layer is his subject. The Code, revised in 2025, does not standardise how a study is built; it sets obligations to participants, clients, the public and the functioning of the market. Its fifth fundamental principle says researchers hold overall responsibility and oversight for the research, irrespective of method, technique and technology — which is the Code's answer to the claim that standards cannot keep up and therefore innovation gets to break them.
Then comes the obligation that matters most for anyone buying AI-assisted work. Article 8a requires suppliers to identify and disclose known, potential or suspected biases, and Guilbert is explicit that this covers generative AI, agentic AI and synthetic data. That is a very different sentence from a claim of rigour: it is an admission about where the work could be wrong. 2
The other half of the piece is addressed to buyers. Esomar publishes two question sets — 10 Questions Buyers Should Ask Suppliers of AI-Based Research Services and 10 Questions to Consider When Using AI in the Research Process — on the argument that expertise and reliability are hard to judge from a sales presentation, so you need something to compare answers against.
The boundary to keep. This is an insider describing his own institution's instruments. It is a reliable map of what exists and what is required. It cannot tell you whether anyone is following it.
Nobody is conducting
Nelson Whipple runs the GRIT survey for Greenbook. In a piece published on 15 October he works through the first measurement of a role the 2026 Insights Practice Report added to its list: insights, analytics or research operations. 3
More than 80% of brand-side participants called it a significant role, and 28% picked it as the most important role of any the survey tracks. Because it is the first measurement, nobody can say whether that is rising.
What distinguishes the people who picked it is scale and complexity rather than company size. A quarter of them work alongside a hundred or more insights professionals, against 14% of everyone else. They also buy differently: 55% weigh a supplier's use of technology heavily, against 28% of others, while general pricing (53% against 70%) and existing relationships (24% against 40%) matter considerably less. They are not pulling work in-house — they are more likely to say they rely too little on external suppliers. Technology maturity and outsourcing come out as complements, not substitutes. 3
The argument of the piece is about metaphor. The industry has started saying "insight orchestration", and Whipple points out what orchestration actually means: arranging music written for one instrument so that an orchestra reproduces it faithfully and repeatably. That is the opposite of an insight, which is understanding you did not have.
His replacement is an ensemble, and ensembles differ: a symphony with a conductor, a jazz combo improvising on a common theme, a festival with different stages playing to different audiences at once. The festival is the one that fits most organisations, because GRIT's data shows the work moving in both directions at once. Among brand-side research professionals, the share working inside an insights or research group rose from 53% in 2022 to 67% in 2026; among analytics professionals, the share inside an analytics or data science function rose from 27% in 2023 to 52%. In the same survey, the share of research professionals saying all insights professionals sit in a formal group fell from 65% to 52%, and about a third now work outside an insights group altogether, embedded in marketing, product, customer experience and elsewhere. 3
Earlier GRIT waves have recorded concern about the rigour of insights work done by non-specialists, though nobody can say how much of it there is. Hence the case for an operations role — and hence why it cannot be a conductor. Whipple's better image is an operating system: it does not decide what every application does, it supplies the infrastructure, protocols and rules that let many independent things run in one environment.
The boundary to keep. This is the industry surveying itself, on a term nobody has pinned down, measured once. Whipple calls the operating-system reading a hypothesis and says the next Business & Innovation study will go after the definition. The data shows who is worried about coordination; it cannot yet show what coordination is.
What to take into the week
Three reads, three places where the chain goes quiet. The respondent cannot say what moved them, so the design has to be the witness: two matched cells, one difference between them, behaviour instead of a rating. The client cannot see the method, so the disclosure has to be the witness: a named list of what could be wrong with this particular piece of work. The company cannot see its own scattered research, so somebody has to run the environment the work happens in, which is a different job from doing the work.
In none of the three does the answer arrive with a person attached who watched it being produced. So the test for the next finding that lands on you is two questions. Was this built to catch the thing that actually decides the outcome, or did it ask people to report it? And can anyone tell you, in writing, what is wrong with it — not whether the supplier is any good, but what could be wrong with this study, named before you act on it.
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.