Speakable schema: what it does and who reads it
What speakable schema does, who documents reading it, Google's beta limits for English news on Google Home in the U.S., and how to mark it up correctly.
By Crawlito · 7 min read
In short
- Speakable is a schema.org property that marks the sections of an Article or WebPage best suited to text-to-speech, using a SpeakableSpecification with CSS selectors or XPaths.
- Google's speakable feature is in beta and, at the time of writing, works only for U.S. users with Google Home devices set to English and for publishers of English news content.
- Google uses speakable to answer topical news queries on smart speakers, reading up to three articles from different sources, and recommends about 20-30 seconds, or two to three sentences, per section.
- Google says to use either cssSelector or xPath in a SpeakableSpecification, not both, and the Rich Results Test does not list speakable among its supported types.
- OpenAI, Anthropic and Perplexity do not mention speakable in their crawler documentation, and Google says its generative AI features need no special schema.org markup.
On this page
Speakable schema marks the parts of a page that are best suited to being read aloud by text-to-speech. The one documented feature built on it is Google's, and it is narrow: a beta feature that lets Google Assistant read news aloud to U.S. users of Google Home devices set to English. None of the AI search crawler documentation from OpenAI, Anthropic or Perplexity mentions it, so add it if you publish English news and want that smart speaker channel, not as a way into ChatGPT, Perplexity or AI Overviews.
This post covers what the property is, exactly who is eligible, how to write the markup and the text it points to, and how to test it. For the wider question of what structured data does for AI search, read structured data for AI search.
What speakable schema is
speakable is a schema.org property used on two types: Article and WebPage. Schema.org describes it as indicating the sections of a page that are "especially appropriate for text-to-speech conversion". Other parts of the page can still be spoken; speakable just points to the parts most likely to be useful when heard.
Schema.org allows three kinds of "content-locator" value:
| Value | How it locates content | Where it is documented |
|---|---|---|
| A URL | Refers to an element by its id, such as #summary | schema.org only |
A SpeakableSpecification with cssSelector | CSS selectors, usually class names | schema.org and Google |
A SpeakableSpecification with an XPath | XPath expressions over an XML view of the page | schema.org (as xpath) and Google (as xPath) |
The markup does not contain the text to be read. It points to text that is already on the page, so the quality of the result depends on the content of those sections, not on the JSON-LD itself.
At the time of writing, schema.org's page for the property shows usage on 100K-1M domains, based on Google's web index for August 2026. That is a measure of how often it is published, not of how often anything reads it.
Who reads speakable markup
Of the operators covered here, Google is the only one that documents a feature built on it. Its speakable documentation says Google Assistant uses speakable structured data to answer topical news queries on smart speakers. When someone asks for news on a topic, Assistant returns up to three articles from around the web and reads the speakable sections aloud. It names the source and sends the full article URL to the user's phone through the Google Assistant app.
Here is what the main AI search operators say, at the time of writing:
| Operator | Documents reading speakable? |
|---|---|
| Google (Assistant on smart speakers) | Yes, as a beta feature for English news in the U.S. |
| Google AI Overviews and AI Mode | Not documented. Google's AI optimization guide says structured data isn't required for generative AI search and there is no special schema.org markup to add |
| OpenAI (ChatGPT search) | No. Its crawler documentation does not mention speakable |
| Anthropic (Claude) | No. Its crawler help article does not mention it |
| Perplexity | No. Its crawlers page does not mention it |
So when an audit tool or plugin recommends speakable "for voice and AI search", the voice part is true for one Google product in one market. The AI search part has no documentation behind it from any of the operators above. The AI search myths post covers the broader "AI schema" claim; speakable is one specific case of it.
Who is eligible for Google's speakable feature
Google lists speakable in its search gallery as a way to identify news content to read aloud on Assistant-enabled devices. The speakable page itself is titled as a beta feature and says it is "in beta and subject to change". The eligibility conditions are tight:
| Condition | What Google's documentation says |
|---|---|
| Users | In the U.S., with Google Home devices set to English |
| Publishers | Content published in English |
| Content | News content, answering topical news queries |
| Types | Article or WebPage |
| Markup | Either cssSelector or xPath, not both |
| Guidelines | The speakable technical and content guidelines, Search Essentials and the general structured data guidelines |
Google says it hopes to launch in other countries and languages once enough publishers have implemented speakable. It gives no date. Selection is also not guaranteed: Google's troubleshooting section says ranking is algorithmic and that Assistant offers up to three articles from different news publications.
If you do not publish news in English, the markup has no documented effect. A product page, a SaaS feature page or a local business page gains nothing from it.
How to add speakable markup
Speakable sits inside the Article, NewsArticle or WebPage JSON-LD you already have. With CSS selectors, point to classes that exist on the page:
<article>
<h1 class="headline">City council approves new cycle lanes</h1>
<p class="summary">The city council voted on Tuesday to build 12 kilometres of protected cycle lanes. Work starts in March. The council expects it to finish by the end of next year.</p>
<p>Full article text...</p>
</article>
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "NewsArticle",
"headline": "City council approves new cycle lanes",
"datePublished": "2026-11-15T09:00:00+00:00",
"url": "https://www.example.com/news/cycle-lanes",
"speakable": {
"@type": "SpeakableSpecification",
"cssSelector": [".headline", ".summary"]
}
}
</script>Google's own example uses XPath instead, pointing to the page title and the meta description:
{
"@context": "https://schema.org/",
"@type": "WebPage",
"name": "Quick Brown Fox",
"speakable": {
"@type": "SpeakableSpecification",
"xPath": [
"/html/head/title",
"/html/head/meta[@name='description']/@content"
]
},
"url": "https://www.example.com/quick-brown-fox"
}Four details trip people up:
- Pick one locator. Google lists
cssSelectorandxPathas the required properties and says to use one or the other, not both. - Mind the spelling. Google's documentation writes
xPath; the schema.org property isxpath. Google is the only documented consumer, so its examples are the ones to follow for its feature. - Class selectors need a dot.
.summaryselectsclass="summary", and Google's owncssSelectorexample uses.headlineand.summary. The example on schema.org's own page uses"headline"and"summary"without dots, which as CSS would select elements named<headline>and<summary>. - The selectors must match real elements. If a redesign renames
.summaryto.standfirst, the markup still validates but points at nothing.
The rest of your article markup follows the usual rules, covered in Article and BlogPosting schema.
How to write the sections speakable points to
Because speakable only points at existing text, most of Google's technical and content guidelines are about writing. They say:
- Keep it short. Google recommends around 20-30 seconds of content per speakable section, roughly two to three sentences.
- Make it make sense when heard. Sections need concise headlines or summaries that give the listener comprehensible, useful information.
- One idea per sentence. If you mark up the top of the story, Google suggests rewriting it into individual sentences so it reads more clearly through text-to-speech.
- Skip what sounds odd aloud. Don't mark datelines (where the story was reported), photo captions or source attributions.
- Mark the key points, not the whole article. A full article would be cut off, losing details the listener needs.
A quick test: read the section aloud. If it starts with "LONDON, Nov 15 (Staff)", refers to "the chart below", or needs the headline to make sense, rewrite it or point the selector somewhere else.
The sections also have to be visible. Google's general structured data guidelines say not to mark up content that readers can't see, so don't create a hidden "audio summary" block just for speakable. See CSS-hidden content for why hidden text is a risk in general.
How to check your speakable markup
There is less tooling for speakable than for most Google features. Google's Rich Results Test does not list speakable among its supported rich result types at the time of writing, so a clean result there tells you nothing about speakable. Use three checks instead.
1. Check that the JSON-LD parses. The Schema Markup Validator shows whether the block is valid JSON-LD and how speakable was read. Syntax errors in the block break everything in it, not just speakable; common JSON-LD errors lists the usual causes.
2. Check that the selectors match. Open the article in a browser, open the developer console and run:
const scripts = document.querySelectorAll('script[type="application/ld+json"]');
for (const s of scripts) {
let data;
try {
data = JSON.parse(s.textContent);
} catch (e) {
console.log("Invalid JSON-LD block:", e.message);
continue;
}
const items = Array.isArray(data) ? data : (data["@graph"] || [data]);
for (const item of items) {
for (const spec of [].concat(item.speakable || [])) {
if (typeof spec !== "object" || spec.cssSelector === undefined) continue;
for (const sel of [].concat(spec.cssSelector)) {
const nodes = document.querySelectorAll(sel);
console.log(sel, nodes.length, nodes.length ? nodes[0].textContent.trim().slice(0, 200) : "NO MATCH");
}
}
}
}Each selector should match an element whose text is the short, spoken-friendly summary you intended. "NO MATCH" means the markup points at nothing. For XPath, document.evaluate() does the same job.
3. Check the raw HTML. Fetch the page with curl and confirm both the JSON-LD and the targeted elements are in the server response, not added later by JavaScript. Raw vs rendered HTML explains why this matters for crawlers.
Google also suggests a manual test on a device: ask Assistant "What's the latest news about" your topic, or "Play news about" it. If your article is not read, that may simply be ranking, which Google says is algorithmic.
Frequently asked questions
Does speakable schema help with ChatGPT, Perplexity or AI Overviews?
Not in any documented way. OpenAI, Anthropic and Perplexity do not mention speakable in their crawler documentation, and Google says no special schema.org markup is needed for its generative AI features.
Is speakable deprecated?
No. At the time of writing Google still documents it as a beta feature and lists it in its search gallery, limited to English news content for U.S. users of Google Home devices set to English.
Can I use speakable on pages that are not news?
You can publish it on any Article or WebPage, and schema.org allows that. Google only documents using it for topical news queries, so on other pages it has no documented effect.
Should I use cssSelector or xPath?
Either works for Google, but use only one per SpeakableSpecification. CSS class selectors are usually easier to maintain because they survive layout changes that would break an absolute XPath.
Check what crawlers actually receive
Speakable is a small, optional addition for English news publishers, and it only works if the article itself reaches crawlers in the HTML. Run the free check on an article to see whether the main AI and search crawlers can fetch it, with no account needed.
Filed under Structure and schema. Topics: structured-data, json-ld, google, ai-search