Speakable schema: make your content AI-citable

Speakable schema is the most underrated JSON-LD schema for AI search. It tells voice assistants and AI search engines which parts of your page summarize the content. They use this to decide what to read aloud or cite.

Adoption is low. The schema is documented but most sites miss it. This is an opportunity.

What is Speakable schema?

Speakable is a property on an Article schema. It uses CSS selectors or xpath to point to the parts of a page that summarize the content. Voice assistants and AI search engines use it to decide what to read or cite.

The schema lives inside the Article JSON-LD:

{ context schema.org, type Article, headline: title, speakable: { type SpeakableSpecification, xpath: [/html/head/title, /html/body//article//h1, /html/body//article//p[1]] } }

This tells the assistant: the H1 and the first paragraph summarize this article. Use those for the answer.

What to point to

The 3 to 5 most important parts of the page. Typically:

• The H1.
• The lede paragraph (the first paragraph after the H1).
• 1 to 3 key takeaways (the bullets under the lede).

Do not point to:

• The navigation. Voice assistants will read the menu.
• The footer. Voice assistants will read the legal copy.
• The entire article. This defeats the purpose.
• Sidebars and ads. These change frequently.

CSS selectors vs xpath

Both work. Pick one and stick with it.

CSS selectors: simpler, easier to read, limited to selectors.

Xpath: more powerful, more verbose, supports position predicates.

Google documentation uses xpath. Most implementations use xpath. We recommend xpath for consistency.

The four mistakes

### 1. Pointing to the entire article

If your selectors match the entire article, the assistant has to decide what to read. Speakable is meant to do that decision for them. Limit to 3 to 5 specific elements.

### 2. Pointing to the navigation

The navigation is the same on every page. If the assistant reads the navigation, they will not cite your content. Exclude the header and the footer from selectors.

### 3. Using class names that change

Some CMSes generate class names like css-1a2b3c that change on every build. Use semantic selectors like article or h1, not class names.

### 4. Skipping the lede

The lede is the first paragraph after the H1. It is the most important part for voice answers. If you skip the lede, the assistant may cite the wrong paragraph.

Validation

Use Google Rich Results test. Paste the URL or the HTML. It will report any schema errors.

Speakable errors do not prevent indexing. They prevent voice citations. If the test reports an error, the assistant will fall back to its own extraction.

Results we have seen

We added Speakable to the top 10 articles on a client site in week 1. By week 6, AI Overview citations on those articles doubled. Voice search traffic was small but high-quality.

The opportunity is the low adoption. Most sites do not have Speakable. If you add it, you have an edge that compounds.