If I were building a podcast for GEO, I'd start before recording

If a business already has access to interesting people, a podcast in production and control of its YouTube publishing, GEO should not begin after the episode is finished. It should help decide what the organisation needs to learn while the guest is still in the room.

John Lewis is launching a YouTube chatshow, which would not be especially notable on its own. Brands have been making podcasts, vodcasts and celebrity interview formats for years. What caught my attention was that the retailer has explicitly connected the new show to changing search behaviour and a desire to improve its visibility in AI-driven discovery.

According to The Guardian, the six-part Gift List show will be hosted by Angela Scanlon and distributed through YouTube and social media. The outgoing John Lewis managing director, Peter Ruis, said the proportion of customers using large language models to search for products had risen from 0.3% a year earlier to 2.5%.

That does not prove that starting a YouTube show improves visibility in ChatGPT, Gemini or any other generative system. The mechanisms are too variable and opaque for that claim, and a commissioning decision is not an outcome study. What it does show is that AI-mediated discovery is beginning to influence what a major brand decides to make in the first place.

That is the part I find interesting, because most generative engine optimisation still arrives much later. The article exists, the video has been edited, the campaign page is built, and only then does somebody ask how to restructure the material so that an AI system can understand or retrieve it more easily.

I have already argued in Are you getting the GEO value from all that video? that this is too late. If a business is paying for access to customers, practitioners or experts, audience and AI-search intelligence can help decide which missing evidence is worth acquiring while those people are available.

The John Lewis story made me want to make that argument more operational. If I were building a B2B podcast today, already had access to YouTube and wanted the work to contribute to GEO, what would I actually change?

I would start with a question territory, not a YouTube checklist

I would not begin by changing the video titles, writing longer descriptions or researching YouTube tags. Those things can matter later, but they do not improve what the organisation knows. I would start by identifying a small set of questions the intended audience genuinely cares about within the territory the podcast is supposed to serve.

Those questions could come from search behaviour, generative-search prompts, customer conversations, sales teams, event questions, community discussions, site search, research or the questions competitors are currently answering better. I would probably begin with ten to twenty questions for a series, not hundreds. The purpose is to create a useful editorial map rather than another keyword database.

Then I would assess the organisation’s current answer to each question. Some will already be answered well, with credible evidence and useful pages that merely need to remain current. Some will be answered badly even though the underlying evidence exists somewhere in research, customer stories, webinars or internal expertise. A third group will expose something more important: the organisation cannot answer confidently because it does not yet have enough evidence.

That last category is where a podcast can become much more than a content format.

A weak answer is not always an optimisation problem. Sometimes the organisation has already done the hard work and simply needs to make its evidence easier to find, understand and attribute. But if the evidence itself is missing, no amount of restructuring will manufacture knowledge that has never been acquired.

The useful distinction is between a content gap and an evidence gap

Imagine a podcast aimed at professional services firms. The organisation behind it knows, perhaps from research, that many firms routinely perform work outside an agreed scope without charging for it. The published content can establish that the behaviour happens, but the organisation may still have a weak answer to the more interesting question: why do experienced professionals keep doing it?

A writer can generate plausible explanations very quickly. Perhaps clients ask for additional help in increments too small to feel billable; perhaps people worry that enforcing the boundary will damage the relationship; perhaps fixed fees obscure the true cost of the extra work; perhaps staff do not feel confident having the conversation. Generative AI can make those explanations sound even more convincing, which is precisely why plausibility is not enough.

If the next podcast guest runs a firm and has direct experience of the problem, the recording creates a reporting opportunity. Instead of asking a broad question such as “Do you struggle with scope creep?”, I would want the interviewer to ask for a specific episode: when was the last time a client asked for something outside scope and you chose not to charge for it, what happened, and what stopped you from drawing the line?

The guest’s first answer may still be vague. A good interviewer can then ask what the request actually was, whether the decision was conscious, what the work cost the firm, whether the client knew it was outside scope, and what changed afterwards. The objective is not to force a pre-written GEO answer into the conversation; it is to use a known evidence gap to make the reporting more precise.

That is a different relationship between GEO and editorial work. Search and generative-answer analysis identify a question worth investigating, while editorial judgement decides whether this guest is a credible source and how to ask the question naturally enough to produce useful evidence.

I would protect the podcast from GEO

There is an obvious way to ruin this idea. Give the producer a spreadsheet containing fifty prompts people use in ChatGPT, insist that the presenter asks them verbatim, and turn a potentially interesting conversation into an FAQ with microphones.

I would do almost the opposite. The existing editorial concept would remain the starting point, and most of the planned interview would be left alone. Before each recording I would compare the producer’s question set with the series question territory and ask whether this particular guest is unusually well placed to fill one or two important evidence gaps.

Sometimes no change would be needed. If the interview already covers the question well, there is no reason to interfere, and if the guest is not a credible source for any of the gaps, the podcast should not pretend otherwise. Where the opportunity is real, I would add or sharpen perhaps one or two questions and give the interviewer enough context to recognise a useful answer when it appears.

This matters because access is scarce. The organisation may have spent weeks securing the guest, arranging diaries, preparing a studio and paying a production team. Using five minutes of that access to investigate a question the organisation genuinely needs answered is a very different proposition from redesigning the entire programme around search.

I would treat YouTube as a knowledge surface, not a dumping ground

Once the episode exists, having control of YouTube becomes genuinely useful. I would publish the full conversation where possible, because a thirty- or forty-minute discussion contains more retrievable knowledge than a sequence of promotional clips can preserve, and I would take the metadata and navigation around it seriously.

The title should describe what the conversation is actually about rather than relying entirely on a series name and episode number. The description should explain the useful territory covered, link to relevant supporting material and identify the guest clearly. Captions should be reviewed for names, technical language and other errors that automatic transcription commonly introduces, because a transcript that consistently misstates the important terms is a poor knowledge record even before any GEO question arises.

I would also add meaningful chapters. YouTube’s own guidance allows creators to add timestamps and titles to the description so that viewers can move between sections, while Google Search documentation says creator-specified timestamps and labels in YouTube descriptions can be used for key moments in search.

That means a chapter called “18:42 Why firms don’t charge for out-of-scope work” does useful work for a person trying to navigate a long episode, and it also gives search systems a more explicit description of what happens at that point. I would not claim that the chapter heading guarantees an AI citation. It does something more basic and defensible: it makes the informational structure of the video clearer.

The same principle applies to short clips. I would still publish the emotional, surprising or entertaining moments that make good social content, but some clips should be selected because they contain unusually good answers to identifiable audience questions. “Why accountants struggle to charge for scope creep” has a different informational function from a generic quote card, even if both come from the same conversation.

I would not make YouTube carry the entire GEO job

This is the biggest practical mistake I would want to avoid. A YouTube upload should not become the sole home of an answer simply because the video contains it.

Google has a mature video-search layer and publishes detailed guidance on making video discoverable, indexable and eligible for features such as key moments. It also says videos can appear across standard search results, Video mode, Images and Discover, so there is a real search-discovery case for publishing video carefully. Google’s video SEO guidance is useful here precisely because it is specific about what it does and does not guarantee.

Generative search is less predictable. OpenAI’s current publisher guidance says public webpages can appear in ChatGPT search and advises publishers not to block OAI-SearchBot if they want their content to be discoverable, surfaced and cited. Its separate search guidance also makes clear that placement is not guaranteed.

So if a podcast produces a genuinely stronger answer to an important question, I would create or improve an owned, crawlable written expression of that knowledge as well. That does not mean publishing an “episode recap” by default. It means asking which existing page deserves to become better because of what was learned, or whether the recording has earned a new answer that did not previously exist.

A useful page might combine a direct answer, the organisation’s existing research, the practitioner’s explanation, an attributable quote, relevant examples, caveats, links to supporting evidence and the embedded or timestamped video. The video supplies voice, context and first-hand testimony; the webpage can make the argument concise, sourceable and easy to connect to the rest of the organisation’s knowledge.

This is why I increasingly think of video as a capture layer rather than the whole GEO answer. The value is not that “LLMs like video”. The value is that a good recording can create proprietary evidence, while the publishing system gives that evidence several durable and retrievable forms.

I would stop measuring reuse by how many assets one recording produces

Content marketing has spent years celebrating the multiplication diagram: one podcast becomes one article, six clips, four quote cards, a newsletter and ten social posts. There is nothing inherently wrong with that, but the diagram measures production efficiency rather than what the organisation learned.

I would want a different post-recording review. What did the guest tell us that we could not previously substantiate? Which important question received a materially better answer? Which useful phrase or distinction came from the practitioner’s own language rather than the category’s marketing vocabulary? Which claim now needs verification, and which question became more interesting because the interview did not resolve it?

A recording that produces one piece of genuinely reusable first-hand evidence may be more valuable than a recording that produces twenty derivative assets. That evidence can improve the next article, a sales conversation, a customer story, an executive argument or another interview without requiring the organisation to rediscover it from scratch.

The shift is from asset reuse to knowledge reuse. It sounds semantic until you change what gets captured at the end of the project.

Every episode should leave the next brief smarter

This is where the model connects to something I have been developing in The last video should make the next video smarter. A production creates a hidden second output alongside the finished film: learning about the questions that worked, the evidence that was missing, the language people actually used and the gaps that became visible too late.

After each podcast episode I would therefore keep a very small editorial record. It would note the important questions the episode genuinely helped answer, the strongest new evidence, useful language, unresolved claims, new questions and any interviewing lesson worth carrying into the next recording. Most observations would remain observations rather than becoming permanent rules.

That last qualification matters. One guest’s experience is not a market truth, and one good interview technique does not automatically become a standard. AI can help compare transcripts and surface candidate patterns, but a human editor still has to decide what is credible, reusable and worth preserving.

Over several episodes the question territory should change. Some gaps will close, others will prove unimportant, and new ones will appear. The podcast therefore becomes a learning loop rather than a sequence of isolated productions: audience question, evidence gap, interview, first-hand answer, publication, response, new question, better next interview.

That is also how this connects to GEO-informed thought leadership. GEO is most useful upstream when it helps an organisation decide what is worth learning or proving, not when it merely gives an already-finished argument a different heading structure.

The pilot I would actually run

I would resist turning this into a large “GEO podcast strategy” before any evidence exists that the extra work is useful. If a business already had a four-episode series in production and access to its YouTube channel, that would be enough for a bounded test.

Before the series, I would identify roughly ten important audience questions and baseline the current answer to each. The test is not whether those exact prompts rank somewhere; it is whether the exercise reveals two or three meaningful evidence gaps that the planned guests could credibly help close.

Before each recording, I would review the existing interview guide and choose no more than one or two gaps worth investigating. The producer would keep control of the conversation, while the extra briefing would explain why the question matters and what is currently unknown. During the interview, the priority would remain listening well enough to follow the useful answer rather than mechanically completing the list.

After recording, I would work from the full transcript rather than the final edit alone. I would extract the strongest new evidence, flag anything that needs verification, identify what remains unanswered and note any language or example worth carrying into future work.

Publication would then happen in two connected places. YouTube would receive the full episode with clear metadata, reviewed captions, useful chapters and links; the owned site would receive a stronger answer wherever the new evidence genuinely improves one. Social clips would distribute selected moments, but they would not be treated as the evidence repository.

Before the next episode, I would update the question and evidence record. That closes the loop while the learning is still fresh and gives the next guest a chance to advance the territory rather than restart it.

I would measure whether the knowledge improved before claiming GEO success

There are familiar metrics worth watching: views, watch time, retention, search traffic, video indexing, referrals and the performance of individual clips. I would keep them, but they are not enough to tell me whether this particular intervention worked.

The first measure would be evidence quality. Did the additional question produce something specific, attributable and useful that the organisation did not previously have? The second would be answer quality: did that evidence materially improve an owned page or create a credible new answer rather than simply another summary?

Then I would watch discoverability over time. Does the improved page appear more often for the relevant conventional searches? Does the video surface for useful queries or key moments? Does the organisation’s evidence begin to appear in generative answers or produce referral traffic? Those signals are worth tracking, but I would treat them as observations across repeated checks rather than pretending that one AI result is a stable ranking.

Finally, I would measure whether the production itself learned. Did an unanswered question from episode one change the brief for episode two? Did a weak interviewing pattern get corrected? Did a guest’s language reveal a better way to frame the problem? If none of that happens, the series may still be a perfectly good podcast, but the proposed intelligence loop is not doing much.

The strongest counterargument is that none of this may improve AI visibility

That possibility should stay inside the model rather than being treated as an awkward footnote. John Lewis commissioning a YouTube show partly in response to AI discovery is interesting evidence about changing brand behaviour; it is not evidence that the show will gain generative-search visibility because it is on YouTube.

Google’s relationship with video is comparatively well documented, while generative systems use different retrieval, ranking and citation processes that can change quickly. A business could implement everything I have described, produce an excellent series and see no measurable improvement in AI citations.

I would still regard the pilot as useful if the organisation acquired better first-hand evidence, created stronger owned answers and stopped losing the learning from each interview. That is an important test of whether a proposed GEO practice is actually good editorial practice underneath. If the workflow only makes sense when an opaque model rewards it, it is too fragile to build much around.

The more interesting version creates value even if the AI-search effect turns out to be small. It improves the questions, the evidence, the publishing and the organisational memory first; discoverability is an additional outcome to test rather than the sole reason for doing the work.

GEO gets more interesting when it changes the brief

The John Lewis story matters to me because it suggests the commissioning layer is beginning to move. A major retailer is not only asking how to optimise the pages it already has; changing discovery behaviour is helping shape a new media investment.

For B2B organisations that already make podcasts, webinars and customer videos, the opportunity does not require copying John Lewis or launching another show. Much of the expensive infrastructure already exists: the access, the guests, the producer, the recording time and the channel. The relatively cheap intervention is to become more deliberate about what the organisation needs to learn while all of that is available.

The operating model I would test is therefore straightforward: use audience and generative-search questions to identify important evidence gaps; add only the questions a particular guest can credibly help answer; preserve the full first-hand response; make the video navigable on YouTube; turn the strongest knowledge into an owned, crawlable answer; and feed the unresolved questions and production learning into the next episode.

That is not a trick for making a chatbot cite a podcast. It is a way of connecting audience questions → evidence acquisition → useful publishing → discoverability → learning without creating a separate GEO content machine.

If it works, the podcast does more than generate episodes. It becomes part of the organisation’s evidence infrastructure, and every recording leaves behind something the next one can use.