Podcast show notes are the most underrated written asset in a creator's catalog. Done well, they pull listeners back to specific moments, rank in search, get shared on socials, and serve as the canonical link in the episode description. Done badly, they read like an AI summarized a transcript and gave up halfway through. Listeners can tell. Search engines can tell. The drop in click- through on episode pages is real and measurable.
The good news is that AI is genuinely useful for show notes, because the raw material is structured (a transcript) and the output format is repeatable. The bad news is that 90% of AI-drafted show notes have a specific, recognizable smell. This guide is how to remove it.
What "AI smell" looks like in show notes
AI-drafted show notes have a small set of patterns that stack up across an episode page. None of them are wrong individually. The combination is what gives the work away.
- Generic episode openers. "In this episode, we explore..." or "Join us as we dive into..." The pattern repeats from episode to episode.
- Bullet points that summarize without quoting. Real notes pull a specific phrase from the guest. AI notes paraphrase everything, which loses the voice that made the episode interesting.
- Timestamps that round to the nearest minute. "00:14:00" instead of "00:14:32." The rounding is a tell. Real listeners reference specific moments.
- Vocabulary mismatch. The guest spoke conversationally; the notes read like a McKinsey report. "Stakeholder engagement," "actionable insights," "value proposition" appear nowhere in the actual conversation.
- A wrap-up sentence at the end. "Tune in to hear more about..." This is the AI equivalent of a wave. Cut it.
The three formats of show notes
Before any AI step, decide which format the episode needs. Each format has different failure modes and different humanization needs.
Transcript-based notes
Full or near-full transcript with timestamps. Good for SEO and for listeners who want to search the episode. Long, dense, not heavily edited. The risk with AI here is mostly accuracy (mis-attributing quotes, mishearing names) rather than tone.
Highlight notes
A short summary plus 5 to 12 bullet points covering the key moments. This is the most common format and the one where AI smell is most damaging. Listeners scan these, decide whether to play the episode, and remember moments by the bullet description.
Description-only notes
Two to three short paragraphs, no bullets. Used on apps where bullet formatting does not render well (some podcast players strip markdown). This is the format that needs the most voice work because there is nothing to hide behind.
The workflow: transcript to clean notes
This is the workflow we recommend for highlight notes, which is what most creators publish. Adapt for the other formats.
1. Start with an actual transcript, not a summary
Many creators ask AI to summarize the episode from memory or from the title. The output is generic by definition because the AI does not know what was actually said. Always start from a real transcript (Descript, Otter, Riverside all export them). The difference between transcript-grounded notes and summary-only notes is enormous and obvious to readers.
2. Ask AI for moments, not summaries
The default AI behavior is to summarize. What you want is the opposite: specific, quotable moments. Prompt accordingly. A useful framing:
"From this transcript, extract 10 to 12 specific moments where the guest said something memorable, counterintuitive, or specific. For each moment, give me the timestamp, a one-sentence description of what they said, and a short pull quote (under 15 words) lifted directly from the transcript."
That prompt produces something you can actually edit. The default "summarize this episode" prompt produces something you have to rewrite from scratch.
3. Edit the descriptions, keep the quotes
The pull quotes are gold because they came directly from the guest. Do not edit them. The one-sentence descriptions, however, will be AI-flavored. Edit those by hand. Compare:
AI version: "The guest discusses the importance of building strong customer relationships in the early stages of a startup."
Edited version: "Why she fired 3 of her first 10 customers in her second month, and what it taught her."
The edited version is shorter, more specific, and gives the listener a reason to click. The AI version says nothing the title does not already say.
4. Use precise timestamps
Round timestamps are a tell because they could not have come from someone who actually listened. "00:14:00" is suspicious. "00:14:32" is credible. Round only when you genuinely want to point to a topical section ("00:30:00 Discussion of pricing models"), and use precise timestamps for specific quotes ("00:32:18 'The first time we tried this it bankrupted us.'").
5. Write the intro paragraph yourself
The 50 to 80 word intro paragraph at the top of the notes is where AI smell is strongest and where it does the most damage. This is the first thing listeners read. Write it by hand.
Good intro paragraphs do three things in three sentences:
- Sentence 1: Name the guest and their angle ("Maya Chen built her first company by accident").
- Sentence 2: Give the specific reason the episode is worth playing ("We talked about why she fires customers, what her early pricing got wrong, and the 11pm voice memo that changed her company").
- Sentence 3: A turn or callout ("Spoiler: the customers she fired are now her biggest advocates").
Three sentences, written by you, take five minutes. The rest of the notes can be AI-extracted-then-edited. The intro paragraph does most of the voice work.
6. Add links and resources by hand
AI tends to hallucinate links. It will invent book titles, mis-cite URLs, or refer to "tools mentioned" that were never mentioned. Do not let AI write your links section. Pull links from your guest's own website, from your show prep, or from the actual transcript by searching for "https" or specific brand names.
7. One humanizer pass on the long-form description
If you publish a longer description-style notes section (200+ words), run it through a humanizer once after editing. This catches residual AI cadence in the connective sentences between bullets. Metric37 handles this in a single pass without flattening the specific moments you wrote in.
Episode summaries: the social-post variant
Most podcasters also post a 1 to 2 sentence episode summary on social media. These short forms are where AI smell is hardest to hide because there is no length to dilute it. A bad social summary reads as AI within 8 words.
The fix is to write social summaries as a specific quote plus a specific reason, not a generic teaser:
Generic AI version: "This week, we dive into the fascinating journey of an entrepreneur who built a successful company against the odds."
Specific version: "Maya Chen fired 3 of her first 10 customers in month two. They are now her biggest advocates. We talk about why."
Notice that the specific version uses the guest's name, gives a number, and previews a tension. None of these require AI to write. All of them require you to know what was actually in the episode.
Intro and outro scripts: keep the spoken cadence
Many podcasters use AI to draft intro and outro scripts. This is fine, but the AI default is reading-cadence, not speaking-cadence. Spoken scripts have to survive the test of being read out loud at normal pace without your mouth getting tangled.
Adjustments that make AI scripts work for voice:
- Cut every comma you can. Spoken sentences have fewer pauses than written ones. If a sentence has three commas, it has three places you will sound robotic.
- Replace formal connectives with conversational ones. "However" becomes "but." "Therefore" becomes "so." "In addition" becomes "also."
- Read every sentence out loud as you edit. If you stumble, shorten. If you sound like a press release, swap a word.
- Add stage directions in brackets. "[laugh]," "[pause]," "[ad-lib]" tell future-you where to relax the read.
For a 60-second intro, the script should be 130 to 160 words. AI defaults to 200+. Cut accordingly.
The SEO angle for show notes
Show notes are one of the few places where the podcast can rank in search. Listeners type in a guest name, a topic, a specific quote, and land on the episode page. The thinner the notes, the less surface area for search.
That said, do not let SEO push you back into AI-flavored writing. The thing that ranks now is genuine, specific content, not keyword-stuffed summaries. The workflow above (extract specific moments, write a real intro paragraph, use precise quotes) is already SEO-optimal because it produces content that nobody else has, anchored to a real conversation.
Common mistakes to fix
The "transformative conversation" opener
AI loves describing conversations as "transformative," "fascinating," "thought-provoking," or "eye-opening." Real listeners do not use these words. Replace with what was actually said.
The bullet point that previews instead of tells
"How to build a stronger brand" is a preview. "Why we killed our biggest product line in 2023" is a tell. Listeners click bullets that have already given them the gist, not bullets that promise to tell them later.
The summary paragraph at the end
AI always wants to end with a summary. Cut it. End with the guest's strongest line or your next-episode tease, not a recap of what the reader just read.
Listing every topic in the description
AI tries to mention every topic from the episode in the intro paragraph. The result is a list disguised as prose. Pick the two or three strongest threads and ignore the rest. The bullets cover the rest.
The 20-minute workflow for a weekly show
For a 45 to 60 minute episode with highlight-style notes, the full workflow runs about 20 minutes once you have done it three or four times. Breakdown:
- Export transcript from your editor: 1 minute.
- Prompt AI for 10 specific moments with quotes and timestamps: 2 minutes (AI runs in the background while you do the next step).
- Write the 3-sentence intro paragraph by hand: 5 minutes.
- Edit each bullet description (keep quotes verbatim): 7 minutes.
- Verify timestamps and add links: 3 minutes.
- One humanizer pass on the longer description if any: 2 minutes.
Compare to the alternative of either (a) writing notes from scratch (60+ minutes) or (b) pasting unedited AI output (5 minutes but readers and search both notice). The 20-minute version is the sweet spot.
The real goal
The job of show notes is not to summarize the episode. It is to give the listener a reason to play it, the search engine something to rank, and the social channel something to share. Generic AI summaries do none of those jobs. Specific, transcript-grounded notes do all three.
Use AI for what it is good at: extraction at scale. Do the voice work yourself for the parts where voice matters: the opener, the bullet descriptions, the social summary. Run a single clean pass through a humanizer for the connective tissue. Twenty minutes per episode. Notes that read like a person made them. Listeners who come back.