What a thousand market research podcasts talk about
Between 2018 and August 2026, four market research podcasts published
This page is what came out. It argues for nothing. The numbers, the queries that produced them, and a list of everything I got wrong on the way are all here so you can check me.
All four numbers below are the core cohort only, the same population every chart on this page uses.
Episodes
Transcribed
Claims extracted
Guest appearances
The four shows above are the core cohort — the only one used for any trend on this page. A fifth show, Research Rockstars, was added later as a deliberate falsification test and is kept strictly separate; it appears once, near the bottom. Pooling shows that start in different years manufactures trends, which is why the cohort column exists at all.
Beat 1 · What the episodes are about
Each episode's transcript was read by a language model and tagged against a 26-topic controlled vocabulary. An episode can carry several topics, so these shares do not sum to 100.
Read this with the denominator. Episodes per year in the core cohort:
Beat 2 · The ChatGPT break
AI was already a topic in 2018 — this industry has been talking about "AI" since before it meant anything specific. Then it declined for four years, bottoming out in 2022. Then ChatGPT.
From
The check that makes this trustworthy
A pooled line like the one above is the easiest chart in the world to get wrong. This panel is badly unbalanced — Happy Market Research published 124 episodes in 2019 and zero in 2024; Greenbook did not exist before 2021. A jump in the pooled line could be nothing more than a change in which show dominated that year.
So here is the same measure computed inside each show separately. If the trend were a composition artifact, these four lines would disagree.
Every show jumps independently in 2023. Hollow points are years where that show
published fewer than
Beat 3 · What the AI conversation is about
Here the instrument changes. Everything above comes from a language model reading each transcript. Everything below is a regular expression run over the raw transcript text — no model, no interpretation, no controlled vocabulary that could file something in the wrong place. One hit per episode.
The two instruments have different denominators and they are labelled as such throughout. This section exists because the model-based instrument got this exact question wrong once already; that story is on the corrections page.
Talk about questionnaires roughly quadrupled. Talk about survey programming as an activity went up tenfold. Talk about the person who does it did not move, because it was already at zero.
Across the whole core cohort the survey programmer as a role appears in
This is a measure of language on podcasts. It is not a measure of what anyone does at work, what any company has built, or whether survey programmers are in demand. A thing can be enormously important and never come up on a podcast. The claim here is narrow and it is the only one the data supports: the conversation about building surveys grew fast, and the conversation about the people who build them did not.
The falsification test
The four core shows are industry-news shows. A fair objection to everything above is that execution work simply isn't what an industry-news show covers — you would not expect it to.
So a fifth show was added specifically to break the finding: Research Rockstars, a research-methods training podcast, 100 episodes running from December 2017. Its entire subject is how to actually do research. If the execution layer is discussed anywhere in podcasting, it is there.
It did not break the finding. Research Rockstars mentions questionnaires in 43% of its episodes — a higher rate than any of the four core shows, which run from 8% (Happy Market Research) to 29% (Data Gurus). The show that talks about questionnaires more than any other in the corpus never once mentions the survey programmer, across nine years.
Research Rockstars is reported here only as this side-by-side. It is never pooled into any trend on this page: it starts in 2017 and its coverage is lumpy (37 episodes in 2018, four in 2020), so merging it into a core time series would create movement that is purely an artifact of when it happened to publish.
Beat 4 · Who is in the room
Every guest's company was classified by type. This is an observation about booking, and it is placed next to the section above rather than inside it: the data supports the ratio, and nothing more. It does not establish that one causes the other.
Where to go next
- Methodology — the pipeline, the models, what it cost, and the four rules every chart here follows.
- What I got wrong — a substring bug that shipped a wrong figure into a draft, an extraction artifact that produced a confident and false finding, and an entire analysis that had to be thrown out.