Speakable Schema and AI Citation: What Actually Gets You Cited in 2026

Every guide to AI search this year gives the same advice: add speakable schema to your pages, and the AI engines will start quoting you. I kept seeing it, so I went and read the actual specification and the studies behind it.

What I found is not what those guides promise.

Speakable is a genuinely good idea. But in 2026 it is still a Google beta that never grew up, it only works for a sliver of the web, and the AI engines everyone is chasing, ChatGPT, Perplexity and Claude, do not read your schema at all. So here is the honest version: what speakable schema is, why it will not get you cited on its own, and the one thing about it that is still worth your time.

What is speakable schema, in plain English?

Speakable schema is a small piece of structured data that tells a machine which sentences on your page are worth reading aloud. You point it at your headline and a two or three sentence summary, and a voice assistant or an AI system can lift that passage as the spoken answer. That is the whole idea: mark the most quotable part of the page.

It lives inside your Article or WebPage JSON-LD as a SpeakableSpecification. You address the text one of two ways, a cssSelector (a CSS class, the easy way) or an xPath.

{
  "@context": "https://schema.org/",
  "@type": "WebPage",
  "name": "How image CDNs work",
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [".lead", "article h2"]
  },
  "url": "https://example.com/image-cdns"
}

Google's own guidance is to keep each marked section to "around 20-30 seconds of content, or roughly two to three sentences", and to leave out datelines, photo captions and source lines that sound odd read out. Hold on to that two-to-three-sentence rule. It is the only part of this that still matters, and I will come back to it.

The part the guides skip: speakable is still a beta

Here is what the tutorials leave out. Speakable launched in 2018. Eight years later, Google's documentation still carries a BETA label and the line "this feature is in beta and subject to change". It never went to general release.

The eligibility is the real shock. Google's spec says speakable works only for "users in the U.S. that have Google Home devices set to English", the publisher must publish in English, and it is built for "topical news queries". Google then adds that it hopes "to launch in other countries and languages as soon as a sufficient number of publishers have implemented speakable".

That was the plan in 2018. Eight years of hoping later, nothing changed.

So if you are not a United States news publisher whose readers ask Google Home for the day's headlines, speakable does nothing for your voice results. That is almost all of us.

One thing to be fair about: it is not dead. Google retired seven structured-data types in January 2026, and John Mueller said plainly that Google is "not killing schema". Speakable survived that cull. It just never grew up.

Do ChatGPT, Perplexity and Claude read your schema?

No. This is the fact that quietly undoes most "schema for AI citations" advice.

When ChatGPT, Perplexity, Claude or Gemini fetch your page to answer a question, they read the visible HTML, the words a human sees. Your JSON-LD, sitting in a script tag, is skipped. One experiment tested five major AI systems on direct retrieval and found every one of them pulled visible HTML and ignored the JSON-LD, hidden microdata and RDFa entirely.

"But most AI-cited pages have schema," the counter goes. True, and it is a trap. It is a correlation. Authoritative, well-run sites tend to have both good schema and the authority that earns citations. The schema rides along. It removes friction; it is not the cause. Marking a passage as "speakable" does not make ChatGPT more likely to read it, because ChatGPT never sees the mark.

So does speakable schema help AI citation at all?

Yes, but only sideways, in two honest ways.

The first is the discipline. To add speakable properly, you have to decide which two or three sentences are your actual answer. That forces you to write a tight, self-contained passage near the top of the page, an answer capsule. And that passage, in the visible HTML, is exactly what every AI engine lifts. The schema gets thrown away; the writing discipline it forced on you is the win. It is why the TL;DR on this site is a real, load-bearing thing and not decoration.

The second is Google's own AI. Google and Microsoft both confirmed in March 2025 that they use schema inside their generative features. Google's AI Mode and AI Overviews are fed structured data during query fan-out. So on Google's first-party AI surfaces, your schema, speakable included, does get read. It is only the third-party LLMs that ignore it.

So speakable is not useless. It is just not the lever it gets sold as.

What actually gets you cited by AI?

Since the markup is not the lever, here is what is, measured rather than guessed.

On the page, add real statistics and named-expert quotations to the passage you want quoted. The Princeton GEO study measured expert quotations at about a 41% visibility lift, specific statistics at about 33%, and inline citations to authoritative sources at about 28%. Keyword stuffing was the one tactic that went backwards, down 8 to 10%.

Off your own site, being talked about beats being linked. Ahrefs studied 75,000 brands and found unlinked brand mentions correlate with AI visibility about three times more strongly than backlinks, 0.66 against 0.22. That is the doctrine shift, and it is measurable now.

And get your entity straight. A single, consistent Organization and author identity, wired with a clean sameAs ring, tells the engines who you are before they dare to quote you. That is entity SEO applied to AI search, and it does more for citations than any single markup block.

Do all of this and speakable becomes a footnote, because you will already have written the liftable passage it was only ever pointing at. If you want the full field guide, I wrote one on how to get cited in ChatGPT and Perplexity. And no, an llms.txt file is not the trick either; it is the same story as speakable, a tidy idea that the AI-search bots do not actually use as a ranking signal.

How to add speakable schema without fooling yourself

If you still want it, and it costs nothing to add, here is the honest way to do it.

Is speakable schema worth adding in 2026?

If it is automatic or a five-minute job, yes, as hygiene, not as a strategy. It validates, it is not deprecated, it feeds Google's own AI, and it makes you write a better answer passage.

But if a consultant is selling you "speakable schema for AI citations" as the plan, that is your cue to walk away. They are charging you for a beta that ChatGPT cannot read. The citation comes from the words on the page and the mentions off it, and treating a markup block as the growth lever is exactly how brands waste a year optimising the wrong layer.

Final take

Speakable is a good idea Google half-finished and the AEO crowd oversold.

Add it if it is free. Write the two-sentence answer it makes you write. Then put your real effort where the citations actually come from: liftable, sourced, first-hand content on the page, and the third-party mentions that make AI trust you off it.

Old fundamentals, new surface. Again.