Deep Reads: When the Evidence Lives in Two Places
Here's an ordinary scene from survey work. A respondent trips a flag in the middle of an interview — the device looks wrong, or the answers arrive too fast to be real.
0:00 / 9:47
Deep Reads: When the Evidence Lives in Two Places
This week's three reads share one shape. Two things that decide an answer are normally kept apart — the signal that flags a respondent and the review that decides whether to keep them, the analysis and the person who has to stand behind it, the preference someone states and the choice they actually make. None of the three is an independent study. Two are arguments from inside companies that sell to this industry, and the third is vendor research with its design laid out. Each is worth reading for the join it names.
The signal never reaches the review
Karine Pepin's Are We Cleaning Too Much? The Missing Signals in Post-Survey Review ran in Greenbook on September 25, 2026. Pepin co-founded the sample-quality firm The Research Heads, and the piece describes an internal tool her team built, so it is a practitioner's argument rather than an independent evaluation. 1
Her starting point is that the industry has become good at catching problems at the door and forgetful about what it caught. The signals fall into three kinds: technical signals about the device and the connection, such as a virtual machine or a duplicated respondent; behavioural signals about what happens inside the interview; and historical signals about what a respondent has done across the rest of the research ecosystem. 1
Only a strong signal blocks a respondent before the survey opens. Most flags sit below that threshold, so the person enters the study carrying context that the later review never sees. A researcher then has to decide, from an open-ended answer alone, whether the respondent is fraudulent or merely having a bad day — without knowing the case was flagged for a virtual machine.
The cost is the part worth carrying into your own process. Over-cleaning shrinks the pool of good participants, tilts the sample, and stretches fieldwork. 1 The piece's proposal is to bring the signals back into the review, score the answer alongside the device and the history, and let software point a researcher at the cases that need human attention. It frames AI as an orchestrator here rather than a judge.
The transferable test is about measurement. A quality process that counts removals rewards the wrong behaviour; a better question is how well supported each keep-or-drop call actually was.
The moat is accountability
Jordan Harper's How To Build Your Research Team of the Future is Greenbook partner content published September 21, 2026. Harper works on AI at Qualtrics, and the piece pulls together conversations he held with research leaders at enterprise organizations. No method or sample is described, so it reads as a vendor's synthesis of where the function is heading. 2
His first shift is the automation of execution: survey programming, data scrubbing, coding open-ends and recap notes are being absorbed, and the "velocity gap" between a research team's weeks and a board's days becomes fatal. What remains, he argues, is taste — the judgment to tell what is credible inside a flood of generated output — and accountability, the willingness to stand behind a recommendation and share the weight of a costly decision. 2
The risk he names is about the pipeline rather than the tools. The entry-level tasks that used to build a researcher's instinct are exactly the tasks AI now absorbs, so a junior may never learn to sense that something does not smell right. His remedy is not to preserve the old chores but to redesign the apprenticeship around critique, having juniors audit and red-team model output.
He also wants two bodies of evidence joined that usually sit in separate departments: the sentiment data and the behavioural and operational data. The most actionable insight, in his framing, now lives at the intersection of the two, which is an argument for breaking down the walls between research, customer experience and data science.
One sample, two kinds of evidence
Meera Sawkar's What Actually Sells a Sneaker Collab is the week's only piece with a study behind it. Sawkar leads research impact at Conveo, which sells an AI-moderated interview platform and ran the research; Greenbook published it on September 25, 2026. 3
The design is the interesting part. One hundred US sneaker enthusiasts, each owning at least ten pairs, went through a single 24-minute interview that combined an ethnographic show-and-tell, a MaxDiff exercise on what drives a collaboration purchase, a concept test of three real collaborations, and a Gabor-Granger price ladder for the concept each person chose. 3 Stated preference, revealed choice and willingness to pay therefore came from the same people in the same sitting, rather than from three separate studies reconciled after the fact.
The attribute ranking puts the collaborator near the bottom. Design and look take 31% of preference share, comfort and everyday wearability 24%, sneaker brand 16%, and everything else — colourway, price, story, the collaborator — 5% or below. The concept test agreed with the ranking: the Nike collaboration won on wearability, and the Adidas collaboration, which carried the strongest cultural name of the three, finished last. 3
Two sub-groups pull in different directions. The three quarters of the sample who buy for everyday design and comfort are won on whether the shoe will actually get worn; the one in four serious collectors weight brand and story more heavily and are judging whether a premium has been earned. Price acceptance followed the same order, so the product set the ceiling and the collaboration justified the number rather than setting it.
One limit matters when you weigh the finding. The price ladder was answered by 50 people on Nike, 37 on New Balance and 13 on Adidas, and the authors label the Adidas result directional rather than conclusive. 3
Where the join has to happen
Three reads, three pairs. A quality signal is worth nothing if it never reaches the review that removes a respondent. An analysis is worth nothing until somebody is willing to put their name on it, and the apprenticeship that builds that willingness is the first thing automation is eating. A preference is worth more when the people who state it are the same people who choose.
All three pieces come from companies selling into this market — a sample-quality firm, a survey platform, an AI research company. That does not make them wrong, and it does mean each argument is also the case for buying what its author sells, which is why the study design and the sample sizes are worth reading closely.
Before you use a number this week, ask what it had to be joined to before it meant anything — and whether that join is happening, or the other half is still sitting in another system.
References
- 1
- 2
- 3What Actually Sells a Sneaker Collab — Greenbook
greenbook.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.