When ChatGPT Recommends Your Show

00:00 Hello and welcome back to How To Get Discovered. I'm Maya. And I'm Tom. This is episode 3. Last week we argued about where your transcripts should live and Tom admitted that seeing what people actually search to find your show was quite interesting. I said curious, not converted. There's a difference. Noted. This week is when ChatGPT recommends your show. It's about what happens to podcast discovery when people stop typing into search boxes and start asking an assistant instead. And I'll be arguing that nobody really knows what that means yet, so we ought to be careful what we build for it.

00:41 Which is irritatingly reasonable. Someone has to be. You've brought a whiteboard. You promised last week that you'd walk in here and say, AI search is going to look quaint in three years. Still my opening statement. I agree with it. That's not how this works. Hear me out. The way ChatGPT and Perplexity and Gemini answer questions today is going to look primitive before long. Different interfaces, different models, and we'll all wince at the words we use to describe them. Lovely. Short episode then. But the habit underneath isn't going back.

01:21 People who used to type a question into a search box are opening an assistant and asking it instead. The tools will change. That shift is permanent. Go on then. What exactly are you claiming? I'm not agreeing to anything vague. Remember Sarah from the first episode? New Hive, slightly terrified of it. A couple of years ago, she'd have typed beekeeping podcast for beginners into Google and got 10 blue links. Now there's a decent chance she asks an assistant, what's a good podcast for someone who's just got their first hive? And it doesn't hand her links, it hands her an answer. Three shows, a sentence on each, and why she'd like them.

02:01 And you want to know whether yours is one of the three. Yes, and every podcaster listening should too. Most of them have never thought about it once, and I don't blame them. Two years ago, it wasn't a question worth asking. Right. Two objections. First, most search is still search. Most people still use Google, and most podcast discovery still happens inside podcast apps, through friends, through guest spots on other shows. Chatbots are a slice. Agreed. And it's a slice that's loudest among people who work in tech, who happen to be the same people writing about podcasting. So it sounds far bigger than it is. I'll give you that one entirely.

02:45 I'm not saying assistants have replaced search. I'm saying that slice is growing faster than anything else, and the thing that's loud and niche today has a habit of becoming the default. Sometimes. I remember being told everyone would be listening to podcasts on their smart speakers by now. Fair. Some of these things fizzle. But I'd watch what ordinary people do rather than what the industry says. And the people I know who aren't in tech at all, my mum included, now ask an assistant things they'd never have bothered to search for. She asks a chatbot for podcast recommendations? She asked one for a recipe that used up half a bag of spinach. Same muscle. Once you're used to asking, you ask about everything.

03:30 Fine. A growing slice. Which brings me to the objection I actually care about. Ask the same assistant the same question twice and you can get different shows. Different day, different model version, slightly different wording, and you get a different answer. So any plan that boils down to get ChatGPT to recommend us is building on sand. Yes. Sorry? You're right. If you chase being the answer to one particular question, you'll go mad. The recommendation moves every time the model sneezes. I wasn't expecting to win this quickly. You haven't. Being the answer is unstable. Being one of the things an assistant could plausibly pick from is much steadier.

04:16 Whichever model it is, however Sarah words it, it's drawing on roughly the same pool of material. So don't try to be the answer. Try to be in the pool. That's more defensible than I'd hoped. I'll have it printed on a mug. So what's in the pool? Hand wave it for me. I'm not a computer. Broadly, two things. Whatever the model learned from in training, and, for most of them now, whatever it can pull off the live web when you ask. Articles, reviews, Reddit threads, show notes, episode pages, all of it, text. Not audio. Correct. If there's one sentence I'd like people to take away today, it's that these assistants read.

05:00 They don't sit down with your MP3. Mostly, they know what your show's about from what's been written about it. So a show nobody's bothered to describe. Then, as far as the assistant's concerned, you don't exist. Doesn't matter how good the episode is. The best beekeeping show on earth. And nobody's told the robot. Same point I made about Google in the first episode. It mostly goes by the words too. Which is my problem with this. Isn't it just the same argument? Get good text onto the open web. That's the transcripts argument with a scarier word stuck on the front. Functionally, yes. What gets you found by a search engine is very close to what gets you found by an assistant.

05:44 Readable, structured text on a page anyone can reach. So why are we spending a whole episode on it? Because the stakes are higher. On Google, if you're on page two, a few stubborn people still scroll to page two. An assistant doesn't have a page two. It names three shows, and Sarah doesn't reply, and three more, please. She picks one and gets on with her day. So being invisible used to cost you a bit, and now it costs you the lot. Worse, it's a locked door. Fine. I'll grant you that changes the maths. But half the text you listed isn't yours. Reviews, Reddit threads, somebody's blog?

06:26 You can't write those. No, and you shouldn't try. The minute podcasters start seeding Reddit with fake recommendations, everyone can smell it. And I'd quite like the assistants to be able to smell it too. Optimistic of you? Maybe. Which is why I'd spend the effort on the bit you do control. Every episode you've made is a pile of genuinely useful words. Right now, they're locked in an audio file. Get them out, onto a page that's properly yours, and you've added more honest text about your show than a hundred Reddit posts would. Assuming the words are any good. If they're not, we've got a different problem, and it's not one an assistant can fix.

07:07 But if they are, it's your guests saying what they said. You can't get more honest than that. Then let me take it the other way. If text is the whole game, the quality of the text matters. And I've watched podcasters dump a raw machine transcript under the player and call it done. I know someone who did that. We covered my transcripts in episode one. Thank you. I wasn't talking about you, actually. I did it too. Years ago, I pasted a machine transcript onto a page with no paragraphs, no headings, and timestamps that didn't do anything. One long grey slab. How long did it stay like that?

07:48 Two years, give or take. Mortifying. It did nothing, and I decided transcripts were a waste of time. It took me ages to work out that transcripts hadn't failed. That particular page had. Makes sense. Raw machine output is a wall of words. It gets most of them right, fair enough. But it doesn't know which bits matter, where one topic stops and the next starts, or what the episode's actually about. And an assistant deciding what to quote is far happier with something that looks like a document. Headings, sections, a clear sense of what each part covers. So what's the alternative? Paying someone to rewrite every transcript by hand? Not anymore.

08:32 Transcriptions got much better at the basics, like who's speaking, punctuation, where a paragraph should break. And then you can build on it. Split the episode into chapters, pull out the topics, tie each one to a timestamp so someone can land on the exact moment a thing was said. Now the page reads like an article with a player attached. Go on, say it. Hmm? The name of the company this show's archive lives on. Fine. On every podherd tier, your back catalogue is transcribed and each episode gets a searchable public page with AI chapters and topics.

09:09 Other tools do versions of this too, and the point stands whichever you use. If you're going to do transcripts at all, do them properly. A half-done one is almost worse than nothing because you think the job's finished. That might be the thing we agree on most all episode. Though I'd add one thing. A structured page is only as good as what's on it. If the conversation rambled, you get beautifully organised rambling. Agreed. Chapters won't rescue a dull episode. They just make it easier to find out that it's dull. There's another way to look at all this, though. If people like asking a question and getting an answer, why should that only happen on ChatGPT?

09:51 Meaning? Your own archive could do the answering. PodHerd has a feature called Ask the Archive on its Grow the Herd and Lead the Herd tiers. A listener types a question and gets an answer drawn from your episodes with citations, and each citation plays the exact moment in the episode where it was said. And who's actually going to use that? Nobody goes to a podcast's website to ask it things. Most people won't. But think about who lands on an episode page with a question already in their head. Sarah, three weeks in, wondering whether her bees are about to swarm. She asks, and instead of a chatbot summary, she gets your guest explaining it, in their own voice.

10:34 Or she ignores the box completely, like I ignore every chat bubble that pops up on every website I've ever visited. Some will. I'm not pretending it's a front door. It's more like a really good member of staff for the people who are already inside and slightly lost. Which at least sends her into the episode rather than away from it. That's the bit I like. And a good answer is the sort of thing you'd pass on to a friend, but that's more your territory next week. I'll reserve judgment. My real worry's somewhere else, though. The assistant becomes the middleman. Sarah asks, it reads three transcripts, summarises them, and she never hears a word of anyone's show.

11:17 You've been found and skipped in the same breath. That one does bother me, honestly, and I haven't got a clean answer. Part of it is that the shows that survive are the ones you can't summarise, where the value's in the voice and the way a conversation moves. Which is most of the good ones. Hopefully. Nobody listens to a two-hour conversation for the bullet points. A summary tells Sarah when to check for swarm cells. It doesn't give her the beekeeper who's been doing it for 40 years and still sounds slightly nervous about it. That's the bit I'd actually want to hear. Me too. The other part is that you don't have to take any of this on faith. If the theory's right, it leaves fingerprints. Such as?

12:02 Referrers. Decent analytics tell you where each visitor came from, and PodHerd's analytics show that on every tier. So if assistants really are sending people your way, you'd expect to see visits from chatgpt.com or perplexity.ai landing on particular episode pages. And if they never turn up? Then the assistants probably aren't sending you much, whatever the hype says. And you've learned something useful before spending a single extra evening chasing it. That's the test. Not whether ChatGPT mentions your show when you ask it about yourself. Which everyone does. But whether actual people click through and land on a page.

12:45 That's half an answer to your middleman worry, too. A referrer is somebody who didn't stop at the summary. They came to hear it. And the ordinary Google side of things? From Grow the Herd up there are Search Insights and the Google Search Console integration, which shows your clicks and impressions from Google. The thing you were curious about last week. So you'd see the search side as well as the assistant side, and you'd know which door people were actually using. I'll be honest, I like this bit more than the rest of the episode. Because I might be wrong? Partly. Mostly because it's checkable. I can argue with a theory all day.

13:23 A number in a dashboard is much harder to shout at. So be findable, get into the pool, and then go and check whether anyone actually walks through the door. That's a more measured place to land than I expected from you. I have my moments. Next week. Finally, mine. The loyalty trap. Tom gets the floor for the whole episode. And I'm going to argue that the listeners you've already got matter more than any number of strangers, and that chasing new ones is how decent shows quietly get worse. And I'll be sitting there taking notes. That's what worries me. Thanks for listening to How To Get Discovered. See you next week. Bye.