Chapter 7: AI Authoring & Content-Generation Features

Feature 2 — Expansion, rewrite & improve

Beyond zero-to-draft, the most-used everyday feature is transforming text you already have: expand a stub, shorten, rewrite for clarity, fix grammar, change reading level.

• Jetpack AI offers one-click "improve" passes — readability improvement, flagging long sentences and complex jargon, plus rewrites (jetpack.com).
• Sanity's Transform action "modifies existing documents while preserving their structure," applying targeted changes to specific fields you designate via a target parameter with path arrays (e.g. ['title','body']) — so a rewrite touches only the fields you scope, leaving references and other fields intact (Sanity, 2025). This field-scoping is the structured-CMS answer to the "AI rewrote my whole entry" failure mode.
• Contentful ships grammar-correction and improvement templates among its pre-built AI Actions (Contentful Help Center, 2025).

The control concern here is blast radius: in a freeform editor, "rewrite" can silently overwrite intent; in a field-scoped model, the change is bounded and diffable.

Feature 3 — Summarization

Summarization powers TL;DRs, meta descriptions, listing excerpts, and email snippets. Jetpack AI auto-generates "summaries" alongside titles and meta descriptions (search/AI WordPress reviews, 2025–26). In structured systems this is typically a Transform/AI Action that reads a body field and writes an excerpt or summary field — again, field-to-field rather than blob-to-blob, which keeps the summary a first-class, queryable piece of content (usable for cards, RSS, search snippets, and llms.txt-style machine summaries).

Feature 4 — Tone & brand voice

This is the single most commercially important control feature in 2026, because it is what makes AI output publishable rather than generic.

• Jetpack AI offers one-click tone switching across a range (formal, provocative, humorous, etc.) (jetpack.com).
• AEM can "autonomously create content that is brand aware — incorporating brand tone of voice, style guidelines, and other specific requirements," explicitly addressing the "is this brand-aware, legally compliant, and tailored per audience?" question (Adobe Business blog, 2025).
• The broader industry pattern is a Brand Kit / brand-voice profile: a persistent constraint applied to every AI interaction. As Contently describes it, "If your brand avoids superlatives, the AI avoids superlatives. If your brand uses specific product terminology, the AI uses those terms exactly" (Contently, 2025). Contentstack similarly markets brand-voice persistence so a junior editor's output "matches the same voice standards as a senior strategist."

The reliability mechanics that make brand voice actually stick are: (1) a stored style guide injected into every prompt (system-prompt-level guardrail), (2) protected/forbidden term lists, and (3) automated post-generation checks for voice, tone, terminology, and style. Sanity's Translate action even supports "style guides and protected phrases to maintain voice and brand terminology across languages" (Sanity, 2025) — the same primitive applied to localization.

Feature 5 — Translation & localization

Translation is the most mature, lowest-risk AI authoring feature because output is verifiable against a source.

• Storyblok AI Translations generates "professional-level translations" across a long language list (Arabic, Chinese, French, German, Japanese, Korean, Spanish, and more), with extended language support shipped through 2025; there is a dedicated ai-translate Management API endpoint (Storyblok, 2025).
• Sanity's Translate Agent Action is purpose-built for multilingual versions and respects style guides + protected phrases.
• Contentful ships translation as one of its pre-built AI Action templates with per-locale variables.

Because headless CMSs model locales as first-class data, AI translation here writes into locale-specific fields/documents, preserving structure across languages — far cleaner than translating rendered HTML. The remaining concern is terminology drift and untranslatable brand terms, which protected-phrase lists exist to solve.

Feature 6 — Alt-text generation (accessibility)

Image alt-text is a high-volume, low-glamour task that AI handles well and that regulation increasingly demands.

• Storyblok launched AI alt-text suggestions directly inside the image upload and image-editor interfaces in March 2025, and can generate alt text in languages other than English (Storyblok changelog, March 2025). It also integrates third-party Alttext.ai for the same job.
• Contentful lists alt-text generation as a pre-built AI Action template.
• The driver is WCAG 2.2 conformance plus the European Accessibility Act (in force June 2025), making alt-text a compliance feature, not just convenience. The quality caveat: vision-model alt-text describes what is in the image but not always why the image is there in context — editorial review still matters for decorative-vs-informative judgment.

Feature 7 — SEO meta, titles & descriptions

Auto-generating SEO titles, slugs, and meta descriptions is near-universal. Jetpack AI "auto-generates compelling titles, meta descriptions, summaries" (reviews, 2025–26). Yoast SEO, RankMath, and AIOSEO in the WordPress ecosystem all added AI title/meta generators. In structured CMSs this is an AI Action that reads the body and writes the seoTitle/metaDescription fields, with character-count validation enforced by the schema. The 2026 nuance is GEO/AEO (Generative Engine Optimization / Answer Engine Optimization): meta and titles are increasingly tuned for citation by AI Overviews and answer engines, not just blue-link CTR — which feeds directly into the next feature.

Feature 8 — FAQ & schema.org structured-data generation

A standout 2025–2026 capability: generating valid JSON-LD structured data (especially FAQPage/Question/Answer) directly from page content.

• The payoff is measurable: industry analyses report pages with FAQPage markup are ~3.2× more likely to surface in Google AI Overviews, and FAQ schema has one of the highest citation rates among structured-data types (Frase.io; SearchScaleAI, 2025–26).
• The recommended pipeline is now AI-assisted: feed raw HTML to a fast model (Gemini 3 Flash is cited as the 2026 tool of choice) which "extracts entities and outputs validated JSON-LD" (DigitalApplied, 2026). Best practice constrains FAQ answers to ~30–50 words and questions to ~80 characters.
• The hard control concern: schema must match rendered content, and you must "audit the rendered page, not just the CMS settings" (Opace, 2025). AI that hallucinates a Q&A not present on the page risks Google structured-data penalties. JSON-LD is the safe, Google-preferred output format.

Feature 9 — Repurposing & cross-channel

The newest frontier: take one canonical piece of content and emit channel-specific variants — social posts, email, ad copy, push notifications.

• AEM Generate Variations produces multiple title/description/CTA variants from a fragment for different audiences and channels (Adobe, 2025).
• Sanity Content Agent (late 2025, led by co-founder Even Westvang) "understands your content structure, finds what needs fixing, and makes the changes, hundreds at a time, staging everything for your review" — repurposing/normalization at corpus scale with a review gate (Mastra customer story; Sanity, 2025).
• The structured-content advantage shines here: because the source is typed data, repurposing is "one hybrid index, many surfaces" — generate once, transform per channel, each variant stored as structured content.

Cross-cutting: quality and control concerns

Across every feature above, the same risk surface recurs.

1. Hallucination. Industry analyses cite AI-generated content containing factual errors in ~15–20% of outputs (MIT-cited figure, via Contently 2025). The mitigations that work in a CMS context:

• RAG grounding — connect the LLM to your own content lake / knowledge base so output is grounded in your real data, not training-data guesses (Contently; Devoteam, 2025). Sanity's Content Lake and Contentful's entry data make in-CMS RAG natural.
• Field-scoped generation to bound blast radius (Sanity target paths).
• Schema validation as a hard gate — structured generators reject outputs that don't fit field types.

2. Mandatory human review. Contently's 2026 governance survey found platforms that "treat human oversight as optional scored lower because AI hallucination rates remain significant without expert validation." Every credible production workflow keeps a human-in-the-loop approval step before publish — typically the existing editorial/workflow gate, now applied to AI drafts.

3. Brand-voice & terminology drift — solved with persistent Brand Kits, protected-phrase lists, and automated post-generation style checks.

4. Provenance & disclosure (now a legal requirement). EU AI Act Article 50 transparency obligations take effect August 2, 2026 (confirmed against the European Commission's own application timeline as of 2026-06-05 — the date was not moved by the 2026 "Omnibus" digital-simplification package, which postponed certain high-risk Annex III deadlines; treat "Art. 50 unaffected by the Omnibus" as Probable, pending the final amending regulation). Art. 50(2) requires providers of generative AI to mark outputs (audio/image/video/text) in a machine-readable format detectable as artificially generated; Art. 50(4) requires deployers to disclose deepfakes and AI-generated public-interest text; penalties reach €15M or 3% of global turnover. The Commission's implementing Guidelines on Article 50 and a voluntary Code of Practice on marking and labelling of AI-generated content are, as of 2026-06-05, still in draft (slated for Q2 2026; the draft-Guidelines consultation closed June 3, 2026) — so the operative technical standard is not yet finalized. The emerging de-facto answer is a dual-layer label: a visible "cr" Content Credentials icon for humans, plus invisible C2PA Content Credentials metadata (Provider Name, System Version, Creation Timestamp, Unique Identifier) for machines. (As of 2026-06-05: the current published C2PA spec is 2.4; Content Credentials 2.3 was the Feb 9, 2026 headline release, and C2PA reports 6,000+ members.) Industry/ecosystem reporting (press-sourced, not yet confirmed at the vendor's primary page) says OpenAI joined the C2PA steering committee in May 2026 and now embeds both a C2PA manifest and a Google SynthID watermark in ChatGPT/API/Codex images, alongside Adobe and Google's existing C2PA/SynthID adoption — SynthID remains the most widely deployed watermark. CMS implication: platforms increasingly need to stamp AI-assisted assets with C2PA credentials and log what AI contributed, what humans modified, and who approved publication — both for the EU AI Act and for regulated-industry audit trails.

5. Cost & rate limits. Generation is metered. Jetpack AI, for example, is $8.33/mo (annual) for 100 requests/mo with a 20-request free tier; structured platforms meter Agent Action / AI Action calls. This makes prompt efficiency and caching (generate once, transform/repurpose from the result) a real cost lever, not just a quality one.

Where the field is heading

The trajectory is from assistive (sidebar that suggests text you paste) → embedded (AI Actions on each field) → agentic (Content Agent / WordPress AI Assistant that operate across the whole content set, find issues, and stage hundreds of changes for review). The constant across all three stages — and the recommendation for anyone evaluating or building these features — is that the generator must be schema-aware, brand-constrained, RAG-grounded, and gated by human approval with a provenance log. That combination is what turns a flashy demo into something safe to publish at scale.

Key Takeaways

• Generation is now table stakes; the 2026 differentiator is schema-aware, field-scoped generation (Sanity Agent Actions, Contentful AI Actions, Storyblok AI) versus blob-into-editor generation (WordPress/Jetpack).
• The core feature set is consistent across platforms: draft, expand/rewrite, summarize, tone/brand-voice, translate, alt-text, SEO meta/title, FAQ/schema.org JSON-LD, and cross-channel repurposing.
• Brand voice is the highest-value control feature, implemented via persistent Brand Kits, protected-phrase lists, and automated style checks; Sanity even applies protected phrases to translation.
• Translation and alt-text are the most mature, lowest-risk features and are increasingly compliance-driven (WCAG 2.2, European Accessibility Act).
• FAQ/schema.org JSON-LD generation has real SEO/GEO payoff (FAQPage markup ~3.2× more likely in AI Overviews) but must match rendered content to avoid penalties; audit the rendered page.
• Hallucination runs ~15–20% of outputs without grounding; the fixes are RAG grounding, schema validation as a gate, and field-scoped blast radius.
• Human-in-the-loop approval before publish is non-negotiable in every credible production workflow.
• EU AI Act Article 50 (effective Aug 2, 2026 — date confirmed unchanged as of 2026-06-05) mandates machine-readable AI-output labeling; the de-facto standard is dual-layer disclosure with C2PA Content Credentials (published spec 2.4; 2.3 was the Feb 2026 release) + visible "cr" icon, plus audit logs of human edits and approvals. The Commission's implementing Guidelines + voluntary Code of Practice on marking/labelling are still draft (Q2 2026), so the operative technical standard is not yet finalized.
• The architectural arc runs assistive → embedded → agentic (Sanity Content Agent, WordPress AI Assistant), but the safety recipe stays the same: schema-aware + brand-constrained + RAG-grounded + human-gated + provenance-logged.

Key References

• Automattic / Jetpack. Jetpack AI Assistant — AI Content Generator for WordPress. 2025. https://jetpack.com/ai/ — Feature list (drafts, tables/lists, tone, titles, meta, translation) and pricing tiers.
• WordPress.com. Introducing the WordPress AI Assistant — Now Built Into WordPress.com. Feb 17 2026. https://wordpress.com/blog/2026/02/17/wordpress-ai-assistant/ — Site-aware, layout-aware embedded assistant.
• Sanity. Agent Actions: AI building blocks for structured content. 2025. https://www.sanity.io/blog/agent-actions-ai-building-blocks-for-structured-content — Generate/Transform/Translate, schemaId, target paths, instructionParams.
• Sanity. Content Agent: AI-powered content operations at scale. 2025. https://www.sanity.io/content-agent — Corpus-scale changes staged for human review.
• Contentful. AI Actions (Help Center). 2025. https://www.contentful.com/help/ai-automations/ai-actions/ — Per-field AI Actions, variables, prompts, custom LLM selection.
• Contentful. Templates for AI Actions. 2025. https://www.contentful.com/help/ai-automations/ai-actions/ai-actions-templates/ — Pre-built translation, SEO, alt-text, grammar templates.
• Storyblok. AI Alt-text generation. March 2025. https://www.storyblok.com/cl/2025-march-ai-alt-text-generation — Alt-text in image upload/editor, multilingual.
• Storyblok. AI Translations / Extended language support. 2025. https://www.storyblok.com/cl/2025-april-extended-language-support-for-ai-translations — Professional translations, ai-translate API.
• Adobe. Introducing new generative AI capabilities in Adobe Experience Manager Sites. 2025. https://business.adobe.com/blog/the-latest/introducing-new-generative-ai-capabilities-in-adobe-experience-manager-sites — Firefly-powered variations, brand-aware generation.
• Adobe Experience League. Generate Variations. 2025. https://experienceleague.adobe.com/en/docs/experience-manager-cloud-service/content/ai-in-aem/generate-variations/generate-variations — Content Fragment variation generation (title/description/CTA).
• Contently. What AI Governance Should Look Like Inside a Content Team: Top 10 Platforms for 2026. Dec 27 2025. https://contently.com/2025/12/27/what-ai-governance-should-look-like-inside-a-content-team-top-10-platforms-for-2026/ — Hallucination rates, human oversight, Brand Kit guardrails.
• Frase.io. Are FAQ Schemas Important for AI Search, GEO & AEO? 2025. https://www.frase.io/blog/faq-schema-ai-search-geo-aeo — FAQPage citation rates and AI Overview visibility.
• DigitalApplied. Schema Markup AI Generation: Complete Guide 2026. 2026. https://www.digitalapplied.com/blog/schema-markup-ai-generation-guide-2026 — AI-driven JSON-LD generation pipeline and best practices.
• TrueScreen. EU AI Act Transparency: Obligations for Businesses in 2026. 2026. https://truescreen.io/articles/eu-ai-act-transparency-obligations-businesses/ — Article 50 effective Aug 2 2026, penalties, machine-readable labeling.
• C2PA Viewer. EU AI Act and C2PA: What Article 50 Requires for AI Content. 2026. https://c2paviewer.com/articles/eu-ai-act-content-credentials — Dual-layer disclosure, Content Credentials metadata fields.