How to get your podcast cited by ChatGPT and Perplexity

The way people find things is splitting in two. Half still type into a search box and click a blue link. The other half ask ChatGPT, Perplexity, or Google's AI Overview a question and read the synthesized answer — often without clicking anything at all. For that second half, the game isn't ranking. It's being cited.

This is what people mean by GEO — generative engine optimization: the discipline of being the source an AI quotes when it answers a question in your domain. And podcasts, once again, start at a disadvantage. This post is the full playbook: why engines skip audio, how they actually choose citations, the structure that makes an episode quotable, and the loop for verifying and measuring it.

Why answer engines ignore audio

An answer engine builds its response from text it can read, retrieve, and trust, and it needs to point at a source — a sentence, a page, a citation. Your audio gives it nothing to point at. Even if your episode contains the single best explanation of a topic on the internet, a model can't cite a waveform — and it can't cite a video frame either, so a YouTube-only channel is in exactly the same position (the YouTube version of this playbook covers that side).

So it cites the blog post, the Reddit thread, the Wikipedia paragraph — anything textual — and your definitive take goes unmentioned.

The stakes: this is where discovery is moving

A few numbers make the shift concrete:

• ChatGPT drives the lion's share of AI search, at roughly 60% of AI-search usage, and it includes a citation in about 87% of answers (Digital Applied).
• Google AI Overviews now appear in about half of all searches, and pages cited in them earn roughly 35% more organic clicks than non-cited competitors (Memeburn).
• Coverage barely overlaps: only about 2% of cited URLs appear across AI Overviews, ChatGPT, and Perplexity at once — 91% show up in just one engine (Superlines). Being citable everywhere matters; betting on one engine is a coin flip.

How an engine actually picks its citation

Under every answer is the same pipeline: the engine retrieves passages — self-contained chunks its index scored as relevant — then composes an answer and attributes the load-bearing claims. Which means the competition isn't page vs. page; it's passage vs. passage. Three properties decide it:

• Self-containment. A chunk that carries its own meaning — claim, context, attribution — can be lifted into an answer. A sentence stranded mid-transcript can't.
• Answer-first placement. The first ~200 words of a page, and the first sentences of each section, do disproportionate retrieval work. Top-performing GEO content leads with the complete answer instead of building to it.
• Attributability. An engine quotes what it can safely credit. "A guest once said" is unquotable; "Sarah Chen, at 18:42" is a citation waiting to happen.

Hold a raw transcript up against those three and the problem is obvious — which brings us to structure.

What makes content citable

Answer engines favor sources that are:

• Readable — real text, not locked in a player, a PDF, or an image.
• Grounded — specific, attributable claims. A quotable sentence with a timestamp beats a paragraph of throat-clearing. Research from Princeton and collaborators found that adding statistics, quotations, and cited sources lifts AI-citation rates by 30–40% (LLM Pulse).
• Structured — clear topics, entities, and key moments, so retrieval can find the exact passage that answers a question.
• Answer-first — lead with the answer, in the page and in every section.

For a podcast, the atomic citable unit looks like this: a discrete claim + the speaker who made it + the timestamp + a stable URL to that second. Every layer of structure below exists to mass-produce that unit from your episodes.

From transcript to citable source

Getting cited isn't about gaming a model; it's about giving it something honest to quote. Concretely, each episode needs five layers:

1. The transcript, word-aligned to the audio — the raw text substrate, on the page as real text.
2. A grounded summary at the top — the answer-first block an engine can lift whole.
3. Key moments as discrete claims, each phrased to stand alone and each carrying its timestamp URL — the citable units themselves.
4. Disambiguated entities and speakers — who said it, about whom, connected across the catalog so retrieval finds the topic from a dozen phrasings.
5. Schema markup and stable URLs — the episode, its clips, its people as JSON-LD, on one domain where authority accrues.

(This is the same structure that makes you rank on Google — one fix, both surfaces.)

Go further: hand agents a front door

The most direct form of GEO is to stop hoping an AI stumbles onto your content and instead let it query you. Expose your whole archive as a live MCP endpoint — a tool agents connect to and answer from, with a timestamped citation back to the exact moment. That's the difference between being findable and being queryable: a crawler saw your pages weeks ago, maybe; an agent searches your catalog right now.

Alongside the endpoint, the discovery layer for crawlers: a sitemap, an llms.txt that indexes the published catalog, a Markdown twin per episode (append .md) for agents that prefer plain text, and a robots.txt that admits GPTBot, ClaudeBot, and PerplexityBot rather than blocking them by reflex.

The verification loop

Don't take eligibility on faith — test one representative episode end to end:

1. Open the episode page signed out — title, summary, key moments, transcript all present as real text.
2. Append .md — confirm the Markdown twin carries the substance and the moment URLs.
3. Check llms.txt — the episode and the MCP endpoint are listed.
4. Ask a connected agent a question the episode answers, and click the citation back to the timestamp.
5. Run the URL through Google's Rich Results Test — the schema layer parses.

Then watch the receipts: Access Analytics badges AI referrers — chatgpt.com, perplexity.ai — and records the questions agents ask your catalog verbatim. Recurring questions you haven't answered yet are, quite literally, market demand for your next episode. The get found & cited docs walk the full checklist.

Common mistakes to avoid

• The transcript dump. Crawlable but chunkless — retrieval has nothing self-contained to lift.
• Unstable or missing per-moment URLs. A citation needs somewhere precise to point; "the episode page, somewhere in hour two" loses to a blog paragraph with an anchor.
• Wrong speaker attribution. A claim credited to the wrong person is a claim no engine can safely quote — fix misattributions before promoting a topic cluster.
• Single-engine tunnel vision. The 2% overlap number cuts both ways: chasing one engine's quirks forfeits the other engines' citations. Build the neutral structure once.
• Blocking AI crawlers by default. A surprising number of sites still deny GPTBot and PerplexityBot sitewide — self-inflicted invisibility on the fastest-growing discovery surface.

When someone asks an AI a question your podcast already answered, the goal is simple: your episode should be the citation. That's what we build PodHood to do — see Get cited by AI.