Researching a podcast guest usually means reading their bio, skimming their social accounts, and listening to one earlier episode. That preparation gets you the podcast interview questions every other host already asked. Podcast host Jeremy Enns described the result this way in his guide to becoming a better interviewer: "Listen to 10 interviews a given guest has given on 10 different shows and it's almost guaranteed that 90% of the questions will be near-identical."
Guests notice. Someone who has answered "how did you get started?" on twelve shows gives the thirteenth host a polished stock answer, delivered on autopilot. Your listeners have often heard that answer already, because the shows in your niche share an audience.
The fix is to research the guest's past interviews, not only the guest. This post walks through a five-step method, gives you a one-page dossier template, and shows how to hand the tedious parts to an AI model without letting it invent things. A reasonable budget for your first guest is an hour. It gets faster once you have a ledger format you like.
- Build a ledger of every question a guest has been asked across 5 to 10 past interviews.
- Mark each row Skip, Go deeper or Ask, then look for contradictions, unexplained numbers and throwaway lines.
- Collapse it into a one-page dossier, and send the guest themes, not the questions.
- If AI helps, give it the transcripts and check every quote it returns.
What the usual advice leaves out
Most guides tell you to read the guest's blog, check their social profiles, watch their videos, and listen to a few other podcasts they have appeared on. Some go further. Alex Birkett's guide to podcast interviews says to listen to all of a guest's previous interviews, because "you don't want to come in and ask ALL the basic questions that everyone else has asked your guest." That is sound advice, and it leaves the practical part open. Listening to eight episodes gives you an impression. It does not tell you that the same question was asked in six of them.
The method below turns those appearances into a count. Once you can see which questions are worn out, the interesting ones are easy to find.
Step 1: List every appearance
Start with a plain list. Search for the guest's name plus "podcast interview" and "interview" on YouTube. Then use a podcast search engine that indexes people and transcripts: Listen Notes lets you search by person, and Podchaser's transcript search finds episodes where a name is spoken, not only listed in the description.
Aim for five to ten appearances: everything from the last twelve months plus the two or three most-listened older ones. Prioritize shows in your own niche. Your audience overlaps with theirs, so those are the answers your listeners have most likely heard.
Step 2: Get the text
You are hunting for questions, not enjoying the conversation, so reading beats listening. Check each episode page for a published transcript. If there is none, YouTube's auto-generated captions are usually good enough for this job. For audio-only episodes, run the file through a local transcription tool. With OpenAI's open-source Whisper command line tool it looks like this:
whisper guest-interview.mp3 --model small --output_format txt --output_dir transcripts
Auto transcripts mangle names and jargon. That is fine here. You only need enough accuracy to see what the host asked.
Step 3: Build the question ledger
Open a spreadsheet, or a table in your notes, and log every substantive question the hosts asked. Paraphrase them so that near-duplicates land in the same row. Then add three columns: how many of the interviews asked it, the guest's stock answer in a few words, and your verdict.
Here is an example for an invented guest, Maya Lindqvist, who founded a bike repair company. She and her interviews are made up for this post.
| Question (paraphrased) | Asked in | Stock answer | Verdict |
|---|---|---|---|
| How did you start the company? | 7 of 8 | Fixed bikes in her garage, a neighbor asked for help | Skip. One line in your intro covers it. |
| Advice for first-time founders? | 6 of 8 | "Talk to customers before you build" | Skip. |
| What was the hardest year? | 2 of 8 | Supply chain trouble in 2020 | Go deeper. Nobody asked what she changed afterward. |
| Why did you close the second shop? | 0 of 8 | Only mentioned once, in a newsletter | Ask. It is unexplored and she has already raised it herself. |
Three verdicts are enough. Skip means the answer is worn smooth and one line of context covers it. Go deeper means the topic came up but nobody followed the thread. Ask means nobody has touched it.
Step 4: Find what nobody has asked
The ledger shows you what is covered. To find what is missing, look in five places.
- Contradictions. Compare what the guest said two years ago with what they said last month. A changed opinion is a real story, and you can ask about it without setting a trap.
- Numbers with no explanation. If a guest says "we tripled revenue" and no host asks how, that is your question.
- Throwaway lines. Guests drop asides that hosts skip because the conversation moved on. Search the guest's transcripts for phrases like "long story," "we can get into that," and "that's a whole other episode."
- Recent changes. Anything that happened after their last interview: a product launch, a failure, a job change.
- Adjacent ground. Topics next to their work where they clearly have opinions, such as a book they love or an industry habit they think is wrong.
Birkett also recommends a "record skip" question, something that surprises the guest a little. Your list from this step is a good source for those.
Step 5: Check your own archive
Before you finalize the questions, search your own past episodes for the guest's name and their main topic. If you covered the subject with someone else last spring, you can reference it. If the guest has appeared on your show before, your audience will notice a repeat. This takes five minutes and prevents an awkward moment mid-recording.
The one-page dossier
Everything above should collapse into a single page you can glance at while recording. Longer than a page and you will read it instead of listening to the guest. Here is a template:
# Guest: [name] | Recording: [date]
## Context (max 5 lines)
Who they are, what they are known for, why they are on this show now.
## Skip list
Questions they have answered many times, with a one-line answer I can quote.
## Go-deeper list
Topics that came up but were never followed.
## Fresh questions (5 to 8)
Numbered, ordered by where I want the conversation to go.
## Record-skip question
The one that changes the tone.
## Facts to verify before I say them aloud
Names, dates, numbers, with the source for each.
## My archive
Episodes of mine that touch the same ground.
Send the guest themes, not the questions themselves. Most guides recommend this because a guest who has your exact wording in advance tends to script an answer. Something like "your early years, the second shop, and what you'd tell a mechanic thinking of going independent" gives them room to prepare without rehearsing.
Using AI for the legwork
The ledger is tedious to build by hand, and a language model can do most of it, with one condition: give it the transcripts. A model that only has the guest's name can only guess what they have and haven't been asked. The example prompt in School of Podcasting's ChatGPT interview questions post asks for questions a named guest "hasn't answered" and supplies little more than the guest's name and show. Given transcripts, a model can count instead of guess.
A prompt that works for step 3:
Below are transcripts of 8 interviews with [guest name]. For each one I have
marked the host's questions with "Q:".
1. List every substantive question the hosts asked, grouping near-duplicates.
2. For each group, say how many of the 8 interviews asked it and give the
guest's answer in one sentence.
3. Quote the exact words of the answer and name the transcript it came from.
4. Then list topics the guest mentioned but no host followed up on.
Use only the transcripts below. If something is not in them, say so.
Then check its work. Models sometimes fabricate quotes and can cite an interview that does not exist. Spot-check every quote you plan to use by finding it in the transcript. Treat the output as a draft ledger that you correct, not as a finished one.
Mistakes that waste the research
- Reading the dossier aloud. It is a map for you, not a script.
- Using an old quote as a gotcha. Asking "you said X in 2023, what changed?" is fine. Asking it to catch them out is not, and the guest will stop giving you candid answers.
- Treating a stock answer as a dead end. Sometimes a familiar question, asked with a twist, gets the best material. The ledger tells you where a twist is needed.
- Researching until the day of recording. Do the appearance list and the ledger a few days early, then let it sit. New questions tend to surface when you are not looking.
Checklist
- 5 to 10 appearances listed, most recent first
- Transcripts collected for each
- Question ledger built, with a verdict on every row
- Contradictions, unexplained numbers, and throwaway lines noted
- Own archive searched for the guest and the topic
- One-page dossier written
- Facts for on-air use checked against a source
- Themes, not questions, sent to the guest
We are building Episodemate, an assistant for a podcaster's own show, and guest research is the first workflow we built. The method above works without it, and a spreadsheet and an hour will get you most of the way.