Can ChatGPT Find You From a Podcast Interview? Here Is the Experiment.

Why publish a protocol instead of a result

Because we have not run the experiment, and saying otherwise would be the exact failure this series exists to document.

Doing it properly requires booking real podcast appearances for real people and waiting months, which is work a placement business can do and a research write-up cannot fake. Every vendor claim about AI visibility we have examined shares one property: the method is absent. Publishing the method first inverts that.

Pre-registration also removes the most common way studies of this kind mislead. If prompts are chosen after the appearances run, a researcher can keep trying phrasings until one produces a flattering answer, and report that one. Fixing them beforehand makes the result falsifiable.

So what follows is a design any competent team can execute and criticise. We will publish the results when we have them, including if they show nothing.

There is a second reason, less obvious and more practical. A protocol is useful to a reader immediately, whereas a result is useful only if they trust it. Anyone deciding whether to spend money on podcast appearances can take this design, run a cheaper version of it on their own name, and get an answer specific to their own situation rather than an average drawn from somebody else's subjects.

Subjects and controls

Three groups, minimum four people each, matched as closely as circumstances allow on seniority, industry and existing public footprint.

The treatment group receives a programme of podcast appearances over a defined window. The active control receives an equivalent budget spent on something else entirely — a contributed article programme, for instance — which tests whether any change is about podcasts or merely about publishing activity. The passive control receives nothing.

Matching on baseline footprint matters more than it sounds. Someone with an existing Wikipedia article, a knowledge panel or heavy press coverage starts from a position where retrieval systems already have material, and adding six podcasts to that is a different experiment from adding six to someone with almost no public record.

Subjects must consent to being named in published results, because an anonymised study of entity recognition cannot be verified by anyone.

Sequencing matters as much as grouping. Appearances should be spaced across the window rather than clustered, because six episodes published in one fortnight and six spread over three months produce very different retrieval conditions, and a design that clusters them cannot distinguish volume from recency.

Consent should be obtained in writing before baseline.

Fixed prompts, fixed in advance

Ten prompts per subject, written and sealed before the first appearance is booked, spanning four types.

Identity: "Who is [Name]?" and "What is [Name] known for?" Association: "Who is the founder of [Company]?" and "Who leads [function] at [Company]?" Expertise: "Who are experts in [specific niche]?" and "Who should I talk to about [specific problem]?" Provenance: "Has [Name] appeared on any podcasts?" and "Where can I read or hear [Name] discussing [topic]?"

Each prompt is issued to each assistant in a fresh session, five times, on the same day, recording every response in full along with any citations shown. Five repetitions exist because these systems are not deterministic, and a single response is an anecdote.

The scoring is decided beforehand too: whether the subject is named at all, whether the association to their company is correct, whether any podcast appearance is cited, and whether any claim made about them is false. That last category is the one most likely to produce something genuinely useful.

Two prompt types are deliberately excluded. Questions inviting the assistant to evaluate the subject — whether they are credible, or any good — produce text that reads like a finding and measures nothing, since these systems will generate plausible praise for almost any name. And prompts naming the podcast alongside the person test whether the model can read a question, not whether it knows anything.

Measurement windows and what gets recorded

Baseline before any appearance is published, then at 30, 60 and 90 days, with a 180-day tail if resources allow.

At each window: the ten prompts against each assistant, five repetitions each; branded search results for the subject's name, captured as a full results page; and Search Console impressions for the subject's name on any property they control.

Separately, and more usefully, Bing Webmaster Tools' AI Performance report records which of a verified site's pages were cited in Copilot answers and the grounding queries used to retrieve them. That is first-party citation data rather than inference, and it is the only place in this entire design where something is measured rather than observed.

We also record the mechanical facts our earlier measurements showed to matter: whether each episode produced a crawlable page, how much text it carried, whether a transcript was published, and whether the host's robots.txt permits crawling.

Recording the mechanical facts alongside the assistant responses is what makes a null result interpretable. If nothing changes and the appearances produced no crawlable text, that explains itself; if nothing changes despite six transcript-bearing pages being indexed, that is a far more interesting finding about selection.

What this design cannot establish

Causation at the level of an individual appearance, and probably not at all.

With a handful of subjects per group, any difference observed could easily be noise. The design is powered to detect a large effect and would miss a modest one entirely, which is worth stating rather than discovering at the analysis stage.

Assistant behaviour also changes underneath the experiment. Model versions ship, retrieval infrastructure changes and licensing deals are signed, none of which is visible to the researcher. A change at day 60 may reflect a model release rather than anything a subject did.

And as documented throughout this series, no operator publishes how sources are selected for citation. Even a clean positive result would show that something changed, not why — which is a real limit and still considerably more than any vendor claim in this category currently offers.

The honest expectation is worth stating too. Given how flat the citation distributions look in the independent research examined elsewhere in this series, the most likely outcome for a moderately known executive is that nothing measurable changes at 90 days. A null result is the probable result, which is exactly why it has to be committed to in advance rather than quietly abandoned if the numbers disappoint.

It also cannot separate the appearance from everything else a subject does in the same period. People who book podcasts tend to be simultaneously hiring, launching, posting and speaking, and a design that does not record those activities will attribute their effects to the podcasts.

How to run it yourself

The protocol is deliberately reproducible by anyone booking podcast appearances, and we would rather several people ran it than that we owned the answer.

Write the prompts first and store them somewhere timestamped. Recruit a passive control, which costs nothing and prevents the most common error in this field, which is mistaking the passage of time for the effect of a campaign. Record full responses rather than screenshots of favourable ones.

Connect Bing Webmaster Tools to any property the subject controls before the baseline, since its citation data is retrospective only from the point of connection.

Media Morpheus books podcast appearances, which makes us an interested party in the outcome. That is precisely why the prompts and scoring are fixed in advance, why a passive control is included, and why we will publish the results whether or not they favour the thing we sell.

A cheaper single-subject version of this is worth running before committing to the full design.

Methodology and limitations

This report contains no findings. It is a study protocol published before execution so that the design can be examined independently of its results. The intended execution retrieves ten pre-registered prompts per subject against multiple commercial AI assistants, five repetitions per prompt per window, at baseline and at 30, 60 and 90 days, alongside branded search capture, Search Console impressions and Bing Webmaster Tools AI Performance citation data. Five limits are known in advance. First, small subject groups mean the design can detect only a large effect. Second, assistant outputs are non-deterministic and vary by user, session and model version, which repeated sampling mitigates but does not remove. Third, model and infrastructure changes during the study period are invisible to the researcher and cannot be controlled for. Fourth, no operator publishes how sources are selected for citation, so any observed change cannot be attributed to a mechanism. Fifth, the authors book podcast appearances commercially and therefore have an interest in the outcome, which is the reason for pre-registration, matched controls and a commitment to publish null results.

Conclusion

We are not going to tell you we tested whether ChatGPT can find someone from a podcast appearance, because we have not. What we can offer is the design: matched subjects with a passive control, ten prompts sealed before the first booking, five repetitions per prompt, measurement at baseline and 30, 60 and 90 days, and first-party citation data from the one platform that publishes it. The design's limits are stated in advance — small samples, non-deterministic systems, invisible model changes, and no published selection criteria to attribute anything to. That is a weaker claim than the confident percentages circulating in this category, and it is the only kind of claim anyone in this market is currently in a position to make honestly.

Frequently Asked Questions

Sources

1. official Google Search Central. AI features in Google Search, for the indexing and snippet eligibility condition the protocol records.
2. official Microsoft. AI Performance in Bing Webmaster Tools, for per-page citation data and grounding queries.
3. official Google. Search Console, for first-party branded query impressions.