Putting One Executive on Ten Podcasts: The Study Design

What the experiment would actually test

Whether ten podcast appearances by one executive, over a defined window, change six measurable things about how that person appears online.

The outcomes are fixed in advance and deliberately span the range from easily measured to barely measurable. Indexed URLs mentioning the subject. Branded search impressions for their name. Citation in AI assistant answers. Entity association between person and company. Referral traffic to properties they control. And durability of all of the above at twelve months.

The ordering matters, because the reliability of each outcome falls sharply as the list goes on. Counting indexed pages is nearly mechanical. Measuring whether an assistant cites someone requires repeated sampling of a non-deterministic system whose operators publish no selection criteria.

A study reporting only the last outcome would be the weakest possible version of this work. A study reporting the first would be modest, honest and genuinely useful — which is why both belong in the same design.

One design decision is worth defending because it looks like a weakness. Using a single executive rather than a cohort makes the study unable to generalise, and it makes it executable. A cohort study with adequate statistical power would need dozens of subjects and a budget no agency would fund, and the version that never runs answers nothing at all. A single well-instrumented subject with controls is a case study, reported as one.

The case-study framing is the honest label for it.

The outcome that can actually be measured

Indexed URLs, because it depends on things this series has already shown to be checkable.

For each appearance, the protocol records whether an episode page exists, how much server-rendered text it carries, whether a transcript is published, whether the host's robots.txt permits crawling, and whether the page carries PodcastEpisode markup. Our own measurement found those variables ranging from 448 to 53,552 words across comparable shows, which is a hundredfold spread that dwarfs anything else in this design.

Then, at each window, a count of distinct indexed URLs that mention the subject by name. This is measured through site-restricted searches against each host domain rather than a single vanity query, because a general name search returns a ranked selection rather than a count.

The prediction is specific and falsifiable: appearances on shows that publish transcripts on crawlable domains will produce indexed URLs, and appearances on shows that publish a title and a player will not, regardless of audience size. If ten appearances on thin-page shows produce no indexed text, that is a finding worth more to a buyer than any citation result.

Recording this at baseline also produces something useful regardless of the study's outcome: a documented inventory of which shows publish readable pages, which is exactly the information a buyer needs when choosing where to spend and which nobody currently compiles.

What cannot be used, and why

Two instruments that appear in almost every proposal of this kind have to be excluded.

Google Trends cannot measure brand lift for an individual. It reports relative popularity scaled 0 to 100 against a term's own peak, drawn from a sample of searches, and Google states it is "not a scientific poll". A percentage change in a normalised index is not a percentage change in searches, and for a person with modest search volume the sampling noise will exceed any effect.

Screenshots of assistant answers cannot be used as evidence either. Responses vary by user, session, phrasing and model version, so a captured favourable answer demonstrates that one response existed once. The protocol substitutes repeated sampling with full response capture, which is weaker than people expect and stronger than what is currently offered.

What replaces both is first-party data: Search Console for branded impressions on properties the subject controls, and Bing Webmaster Tools for per-page citation counts and the grounding queries that retrieved them.

Both exclusions cost the study its most quotable numbers, which is rather the point of deciding them in advance.

Selecting the ten shows

Deliberately mixed, because a design that books ten similar shows cannot separate audience from infrastructure.

Five shows selected for publishing practice — crawlable episode pages with substantial text and published transcripts — and five selected for audience size regardless of publishing practice. If the outcome tracks publishing practice rather than audience, that is the most commercially useful finding available from this study.

Each show's characteristics are recorded before booking: episode page word count from a recent episode, transcript presence, robots.txt rules, structured data, and whether the feed is connected to YouTube.

This is the part of the design a placement business is uniquely placed to execute, because selecting shows on measurable publishing characteristics rather than on a media kit is something we can do and a researcher without booking access cannot.

Show characteristics must be captured before booking rather than reconstructed afterwards, because they change. A host can add transcripts, alter robots.txt or migrate platforms mid-study, and a record taken at the start is the only way to know what the conditions actually were when the episode published.

Booking access is the scarce input here, not analytical skill.

What would count as a result

Thresholds are set before the first booking, so that a disappointing outcome cannot be reframed afterwards.

A positive result on indexed URLs requires a materially larger count at 90 days than baseline, attributable to identifiable episode pages. A positive result on branded search requires a sustained rise in Search Console impressions beyond normal variation, not a spike in publication week. A positive result on AI citation requires the subject or an associated page appearing in Bing's AI Performance data where it did not before.

Anything less is reported as no detected effect. The design does not include a category for encouraging signs.

The reason for this severity is that the alternative is what the market currently produces. As documented across this series, vendor claims in this area arrive without method, sample or threshold, and are therefore compatible with any outcome.

Pre-committing to thresholds also protects against the subtler failure, which is not fabrication but drift. Faced with flat numbers, an interested party will reach for a softer measure — mentions rather than citations, impressions rather than clicks, any movement rather than a stated threshold — and each substitution is individually defensible. Fixing the measure in advance removes the option.

Why a null result is likely, and worth publishing

Because the independent evidence available suggests these distributions are flat and slow-moving.

Research examined earlier in this series found English Wikipedia the single most-cited source globally across six commercial assistants, with the paper publishing a ranking rather than shares, and regional figures in low single digits. A distribution where the leading source holds a few percent of citations is not one where ten podcast appearances move a named individual into it within 90 days.

Meanwhile Google's own condition for AI feature eligibility is that a page be indexed and snippet-eligible — which ten appearances can satisfy for the pages, without any guarantee about selection among eligible pages, because no operator publishes how that selection works.

So the honest expectation is a measurable increase in indexed URLs, an unclear effect on branded search, and probably nothing detectable in assistant citation. Publishing that would be more useful to a buyer than any of the confident figures currently circulating, and we have committed in advance to publishing it either way. Media Morpheus sells podcast placement, which is exactly why the thresholds are fixed now rather than after the data arrives.

Publishing it also sets a precedent we would rather exist in this market than not.

Methodology and limitations

This report contains no findings. It is a study protocol published in advance of execution. The intended design places one executive on ten podcasts selected on mixed criteria — five for publishing practice, five for audience — over a defined window, with matched passive controls, measuring six pre-specified outcome families at baseline and at 30, 60, 90 and 365 days. Six limits are known in advance. First, a single subject with controls cannot support causal inference; the design detects large effects and is blind to modest ones. Second, assistant outputs are non-deterministic and change with model releases invisible to the researcher. Third, no operator publishes how sources are selected for citation, so mechanism cannot be inferred from any observed change. Fourth, Google Trends is excluded as an instrument because it reports a normalised sample rather than search volumes. Fifth, indexed URL counts depend on search operators whose completeness Google does not guarantee. Sixth, the authors sell podcast placement and have a commercial interest in a positive result, which is why outcomes, thresholds and a commitment to publish null findings are fixed before execution.

Conclusion

The flagship version of this question — put one executive on ten podcasts and measure what changes — is worth running and has not been run here. What we can publish now is the design: six outcomes ordered by how reliably each can be measured, thresholds fixed in advance, Google Trends excluded because it cannot measure what people use it for, first-party data from Search Console and Bing Webmaster Tools where measurement is genuinely possible, and shows selected on publishing characteristics rather than media kits. The likeliest result is more indexed pages, an ambiguous effect on branded search, and nothing detectable in AI citation at 90 days. We have committed to publishing that outcome, which is the part of this design that distinguishes it from the market it examines.

Frequently Asked Questions

Sources

1. official Google. Search Console, for first-party branded query impressions.
2. official Microsoft. AI Performance in Bing Webmaster Tools, for per-page citation data and grounding queries.
3. official Google. Trends Help, for the normalisation and sampling that make Trends unsuitable as a lift instrument.