SpeakableSpecification Schema: Voice AI Optimization for Alexa, Google Assistant & ChatGPT Voice

Speakable schema tells voice assistants which passages on a page are safe to read aloud, but Google’s own documentation labels it beta, limits it to news and article pages, and the markup has not expanded beyond that scope despite years of anticipation. What has changed is the broader voice AI landscape: Alexa+, Gemini Live, and ChatGPT Voice Mode now handle millions of conversational queries daily, and all of them need short, self-contained answer text. This guide covers what Speakable actually does today, where it works, how to implement it, and the voice-answer optimization work that matters even on pages where Speakable is irrelevant.

What SpeakableSpecification Schema Actually Does

SpeakableSpecification is a schema.org type that marks specific sections of an Article or webpage as best suited for text-to-speech playback. Per Google’s official Speakable documentation, the markup identifies sections of web content that are particularly appropriate for audio playback by digital assistants. Google introduced the specification in 2018 and it remains labeled beta today. Google currently restricts eligibility to news publishers and article pages.

That beta label is not cosmetic. Google has not added Speakable to its Rich Results gallery alongside FAQ, HowTo, or Product. There is no Speakable feature in Google AI Mode or AI Overviews documentation. The use case Google documents is narrow: a news article read aloud by Google Assistant on a smart speaker or display. For most B2B SaaS, ecommerce, and agency websites, Speakable is at best a marginal signal and at worst irrelevant.

That is the honest framing. Speakable is worth implementing correctly on eligible pages. It is not worth treating as a silver bullet for voice AI visibility in 2026. The bigger opportunity is optimizing the content itself to be answer-ready for any voice engine, regardless of schema.

Which Content Types Benefit from Speakable Markup

Google’s documentation specifies that Speakable applies to Article and NewsArticle schema types. In practice, the pages that can realistically benefit are a narrow set:

• News articles - The original intended use case. The headline and opening summary are natural speakable candidates. News publishers with Google News eligibility have the clearest path to Speakable recognition.
• How-to guides - Step summaries of one sentence each read cleanly aloud. Mark the overview sentence at the top of each step, not the full step body.
• FAQ pages - Each FAQ answer is already a near-perfect speakable unit if kept under 60 words.
• Definition and glossary pages - The opening definition sentence for each term is a clean voice unit.
• Product summaries - A single benefit-focused sentence about a product makes a direct voice answer for pricing or feature queries.

Long-form thought leadership, opinion essays, and deep technical tutorials rarely benefit. Those formats do not compress to a 30-word spoken passage, so marking sections as speakable gives voice assistants material that sounds awkward read aloud. Do not mark content speakable simply because it ranks well.

In our client work, the highest-leverage speakable targets on commercial sites are usually pricing and comparison pages. A sentence like ‘Acme Pro starts at $49 per month and includes full features for solo consultants’ is exactly what a voice user wants when they ask about pricing. Keep it under 60 words and update it whenever pricing changes.

Implementation: CSS Selector vs XPath

Speakable markup lives inside your page’s JSON-LD structured data. The specification supports two methods to identify which sections are speakable: a CSS selector or an XPath expression. Use CSS selector by default. Reach for XPath only when CSS cannot cleanly target the content.

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [".speakable-headline", ".speakable-summary"]
  }
}

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "speakable": {
    "@type": "SpeakableSpecification",
    "xpath": ["//h1", "//div[@class='article-summary']"]
  }
}

1. Use CSS selectors for new builds where you control the markup. Cleaner, more maintainable, works in all parsers.
2. Use XPath when adding speakable to legacy sites where you cannot easily add CSS classes. More flexible but more brittle.
3. Mark the headline plus one summary section per page. More than two speakable sections is overkill and can confuse assistants.
4. Keep each targeted section under 60 words. Longer sections get truncated mid-sentence by voice engines.
5. Validate with Google’s Rich Results Test after deployment. Speakable validation is part of the Article test suite.
6. Check Google Search Console for structured data errors after the page is crawled. Speakable errors appear under the Enhancement tab.

Speakable vs Voice-Answer Optimization: What to Prioritize

Speakable schema is one signal among several. The table below maps where Speakable applies versus where general voice-answer content optimization delivers results regardless of schema:

Voice AI Optimization Beyond Speakable Schema

ChatGPT Voice Mode, Alexa+, and Gemini Live all surface answers from the same retrieval pipelines that power their text-mode responses. Voice adds an output constraint: the answer must compress to one spoken response, typically 20 to 60 words. Voice engines do not prefer pages with Speakable markup on them over pages without it, outside of Google’s narrow news use case. They prefer pages whose content is already structured as short, standalone factual units.

Our May 2026 study of 153,425 citations found that cited sentences average 9.27 words and 74.9% of cited sentences are in the first half of the document. That positional and length bias applies to voice-mode responses just as it applies to text citations: engines pull from short sentences near the top of the page.

Three content changes that improve voice citation rates on any page type:

• Write atomic answer sentences of 6 to 15 words for every core claim. Each sentence should be extractable without context. In our practice, we treat this as a structural edit pass after the draft is complete.
• Front-load the answer in every section. BLUF (Bottom Line Up Front) formatting matches how voice engines extract spoken answers. The first sentence of each section should be a complete answer, not a scene-setting intro.
• Pair FAQ schema with short answer copy. FAQ and HowTo schema on page gives voice assistants structured question-answer pairs they can pull directly into spoken responses. Keep answers under 60 words per FAQ item. This is the single highest-leverage voice optimization we run across client sites.

How ChatGPT Voice Mode Selects Content

ChatGPT Voice Mode launched in late 2024 and processes voice queries through the same retrieval pipeline as text-mode search. OpenAI has not published documentation specifying that it reads Speakable markup. What we observe in client tracking is that Voice Mode answers tend to draw from the same short, factual sentences near the top of a document that Perplexity and Gemini citation data identify as preferred.

In our practice, we frame the 20 to 60 word section length guideline as a content optimization target, not as a Speakable requirement. Passages shorter than 20 words often lack enough context to be cited without awkward padding. Passages over 60 words risk truncation mid-sentence when read aloud. The same length range that works for FAQ schema answers works for voice citation targets.

FAQ schema combined with short answer copy produces the highest voice citation rates we measure across client work. Mark the question and the answer as separate, self-contained units. The semantic HTML structure underneath the schema also matters: voice retrieval systems favor clean article and section markup over div-heavy layouts where extractable text is harder to isolate.

Realistic Expectations for Voice Search Visibility

Voice citation is a small slice of total AI search engagement for most B2B brands today. The reason to invest in voice optimization is not raw volume in 2026 but compounding entity recognition over time. Voice attribution is a strong signal back to text-mode retrievers that your brand is a trusted reference, which tends to lift text citation rates across a 60 to 120 day window.

We track voice-mode citation rates as part of our broader AI citation tracking work using our open-source GEO/AEO Tracker. Run your top 20 voice-likely prompts through ChatGPT Voice Mode and Google Assistant weekly. Track verbatim quote rate as your primary voice KPI. Do not benchmark against generic industry percentages: voice citation rates vary by topic, domain authority, and content structure more than by schema.

Roll out Speakable in waves. Start with your 10 highest-traffic informational pages and the 5 highest-converting commercial pages that carry Article schema. Measure baseline citation rate for 30 days. Deploy Speakable, hold other variables constant, and measure again at 60 and 90 days.

Testing Voice Visibility: A Practical Workflow

Voice testing is harder than text testing because no API returns ‘here is what Alexa would say for this query.’ The workflow below captures 80% of the diagnostic value:

1. Manual device testing. Ask Google Assistant, Alexa, Siri, and ChatGPT Voice Mode the prompts that should surface your content. Note verbatim attribution, competitor substitutions, and answer length.
2. Google Rich Results Test. Validates Speakable syntax before deployment. A required pre-deploy step for any page with Article schema.
3. Google Search Console structured data report. Confirms Google parsed your speakable markup after crawling. Check the Enhancement tab post-deploy.
4. Weekly Voice Mode prompt log. Run your top 20 voice prompts through ChatGPT Voice Mode each week. Track verbatim quote rate as a proxy for voice citation health.
5. GEO/AEO Tracker monitoring. Our open-source tracker includes voice-style prompt templates that can be adapted for Alexa and Google Assistant query formats alongside text-mode tracking.

The Speakable and FAQ Schema Combination

The highest-performing pattern we have found across news-eligible pages is combining SpeakableSpecification with FAQPage schema. Mark the FAQ question as one speakable CSS selector and the FAQ answer as a second selector. The question signals to voice assistants what the passage answers. The answer provides the spoken text. Together they produce cleaner voice citation than either schema achieves alone.

The underlying internal linking and content architecture still matters. Pages where the FAQPage schema lives must be crawlable by GPTBot, ClaudeBot, and OAI-SearchBot. Any page blocking these bots in robots.txt will not appear in ChatGPT Voice Mode regardless of schema quality. Check your bot directives before optimizing schema.

Voice AI is still a developing surface. Brands that build the foundation now: clean Article schema with Speakable on eligible pages, short atomic sentences across all pages, and FAQ schema with tight answer copy, compound entity recognition faster than brands waiting for a definitive playbook. The original research and primary data we publish comes from running these patterns at scale and tracking what changes.