Paste the cold open, the body, or the sponsor read. Get a 0-100 Authenticity Score in thirty seconds that predicts which lines will read fine on the page but trip the tongue on the mic. Podcasts are 100 percent audio. There is no image, no caption, no on-screen graphic to rescue a flat sentence. The voice is the entire product, and listeners drop off the moment a line stops sounding human. Sentence-level highlights flag the exact spans worth rewriting before you press record. Solo, interview, intro and outro, sponsor reads, transitions, and show notes all run through the same scorer.
Audio strips every escape hatch a reader gets on a page. Listeners cannot skim, cannot jump ahead, cannot pick their own pace. Every stiff sentence is delivered in your voice at the speed you record it, and the cost arrives the moment the line lands.
Listeners chose your show over a thousand others in the genre because of how you sound. The moment your voice carries a sentence nobody would actually say in conversation, you stop being a person and start being a feed. AI prose averages 22 to 30 words per sentence with no breath markers. Spoken English averages 8 to 14 words per clause with a real pause every 6 to 10 words. The cadence gap shows the second the mic turns on.
Spotify and Apple Podcasts both track first-minute drop-off as the single biggest signal for episode quality and downstream recommendation. An AI-drafted cold open is the most expensive mistake a host can make. The script feels productive to write because the words appear on the page. Only when you start reading does the cost arrive, and by then the recording session is already booked.
Even a casual podcast eats 90 minutes per finished hour once you count setup, breath retakes, and stumbles. A single stiff sentence buried in paragraph four costs four full re-reads, three of which the editor cannot save. Catching the line on the page before you record saves the session, not the edit.
Invisible on the page. Obvious the moment the mic turns on. The scorer flags every one of these, and the highlights show you where to cut.
Today on the show. Welcome back. In today's episode, we explore. Make sure to subscribe! Five openers carry 80 percent of AI-drafted podcasts, and listeners skip them in twelve seconds. The scorer drops the cold-open score below 40 on any of these openings. Open mid-confession, on a concrete number, or on the actual news instead.
AI loves a single sentence that runs through three commas and a conjunction. The page handles it. The mouth does not. The scorer flags every clause running past 18 spoken words. Cut each long sentence into two or three short ones, and the score lifts on the read-aloud pass without any further edit.
Utilise, leverage, navigate, robust, comprehensive, holistic. Words nobody says in conversation, dropped into a sentence you now have to read into a microphone. The audio reveals the source instantly. The highlights flag the eight or nine most common offenders. Swap for use, get, work through, strong, full, whole.
AI prose runs in identical paragraph lengths with no breath markers. Real scripts vary paragraph length wildly and mark the breath. The scorer reads uniform rhythm as a signal and lowers the band. Add ellipses where the pause belongs, let one paragraph be a single sentence, let another run long.
AI defaults to do not, cannot, it is, you are. Spoken English uses don't, can't, it's, you're roughly 95 percent of the time. Full forms in audio sound like reading a press release. Host voice also lives in throwaway asides: the mid-sentence opinion, the small story, the side note. AI strips both, and the scorer notices.
AI outros restate every point, thank the listener, then ask for a rating in the same flat cadence. Most listeners have already left. The scorer flags template close patterns. End on a sharp line, a question, or a teaser, and put the rating ask at the 30 percent mark of the episode, not the end.
Every block that ends up read into a mic, including the brand-supplied ones, sits inside the same scorer. Score them individually so the score tells you exactly which segment carries the AI risk.
Live or die on monologue energy. Score the cold open, the body, and the close as separate scans because a strong open can hide a weak body in a single blended average. Target 80 or higher across all three. Most solo hosts write scripts that feel under-prepared on the page but score in the top band because they were drafted by talking, not typing.
Pre-written questions that sound AI-generated are show killers. A great guest hears within 30 seconds whether the host wrote the questions or pasted them from a chatbot, and a disengaged guest gives flat, short answers. Score the question list before the call. Target 75 or higher and use the questions as conversation anchors, not scripts.
The first 60 seconds and last 30 seconds carry the episode. The cold open gets the most attention from the scorer because first-minute drop-off is a platform-level signal. Outro should end on a sharp line, a question, or a hand-off. Boilerplate closes get muted, and the scorer flags them every time.
Brand-supplied copy is the single biggest AI residue source in podcasts. Marketing teams generate talking points in ChatGPT and ship them to the host mostly unedited. Scan the sponsor read alone before the rest. If it lands below 60, rewrite into the host register while keeping the brand's required claims intact.
The throwaway lines between blocks are where AI rhythm leaks back in. They are short, so they get skipped in the rewrite pass. Listeners hear the seams. Scan and score every transition segment as one block, not as throwaway, and target 75 or higher.
Written text in support of the audio, but still part of the channel voice. Subscribers read show notes on the podcast app when deciding whether to play the episode. Templated descriptions read suspect and lower the click rate. Score the description and the chapter labels as one block before publishing, and treat the episode title as the single highest-leverage piece to score.
Free fits a weekly solo show on short scripts. Pro fits daily releases, interview prep, and sponsor reads. Business fits production studios and podcast networks. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
A loop that fits between draft and record. Most hosts run it on every priority episode. Total time under ten minutes for a 30-minute solo show after the first run.
Open the script in Google Docs, Notion, or wherever your show notes live. Paste the cold open into TextSight first, then the body, then the sponsor read and outro as separate scans. Score each segment alone so the band reflects the actual recording risk. A blended whole-script score hides a stiff cold open inside a strong body.
The scorer returns a 0-100 Authenticity Score for each segment plus a sentence-level colour map: green for spoken-natural, yellow for mixed, red for likely AI. The segment score tells you whether the block needs work. The highlights tell you exactly where. Two or three red sentences usually carry the entire score, and editing those three is faster than rewriting the whole segment.
Sort the segments by score and start with the lowest. Cold open and sponsor read carry the most weight on first-minute drop-off, so prioritise those even if the body scores lower. Read each red sentence out loud at recording pace. If the line trips your tongue, lengthens unnaturally, or feels like a different person, the score is right and the sentence needs to go.
Cut long clauses into two short ones. Add a contraction. Drop a filler where you would actually say one. Swap utilise for use, leverage for get. Read the cold open three more times and time it; if it runs over 60 seconds, cut. Paste the revised segment back into TextSight to verify. Target 80 or higher on every block before the recording session.
A high score on the page is necessary but not sufficient. The mic is the final judge, and the spoken-voice mismatch only reveals itself when you say the line at recording pace, in full voice, with no audience to perform for.
The scorer flags AI rhythm at the sentence level. The read-aloud test flags the host-voice mismatch at the line level. A sentence can score 92 on the page and still sound like someone else said it. Your favourite asides, your cadence, your specific opinions — none of that lives in the scorer. It lives in your mouth. Reading every paragraph out loud catches the words you would not say even when the scorer cannot.
Not in your head. Full voice, at the speed you would record. Anything you read silently you will not catch. Mark every stumble, slow spot, breathless line, and any place that did not sound like you. Each mark is a cue that the script is not yet speech, and each one gets edited by hand before the session.
Once the script is rewritten, run the full read-aloud pass one more time the morning of recording, after coffee, at your normal speaking volume. The cold open gets three reads minimum. If the cold open runs past 60 seconds in real time, cut until it fits. The mic does not save lines the script lost, and the cold open carries the entire episode.
A page lets the reader set the pace. A microphone takes that control away. Three rhythm patterns separate written prose from spoken prose, and AI defaults to the written shape every time.
Spoken English averages 8 to 14 words per clause with a real pause every 6 to 10 words. AI prose averages 22 to 30 words per sentence with no breaks. Break every long sentence into two or three short ones, and mark pauses with an ellipsis so the host knows where the breath belongs. The script becomes a delivery guide, not a wall of text.
Real hosts interrupt themselves mid-sentence. They shift gears, contradict the last line, add a quick story, then come back. AI never does this. Every sentence in an AI draft completes itself cleanly, which is the opposite of how people actually talk. One aside per minute is the cheapest way to add humanity back into the script, and the scorer rewards it every time.
Native speech contracts in roughly 95 percent of slots. Don't, can't, it's, you're, I'm. Real hosts also use small fillers like right, you know, and look, and address the listener as you. AI strips every one of these. Adding them back closes most of the gap between a scored draft and a script that sounds like a person made it.
If your script scored low. Three modes, sentence-level highlights, voice-preserving rewrites tuned for spoken cadence.
Open AI rewriter →Sister scorer for show notes, episode descriptions, and any written content that supports the audio.
See the workflow →The 0-100 score explained, with target bands for spoken vs written formats.
Read the explainer →The full detection tool. Same scorer, more controls, file and URL upload on paid tiers.
Open the detector →Free to try. No card. Score the cold open, body, sponsor read, and outro as separate segments, then run the read-aloud loop and protect listener retention before you press record. Your first scan in about six seconds.