Many guides on this topic are already out of date. OpenAI removed canvas from GPT-5.5 Instant and GPT-5.5 Thinking on May 28, 2026, and drafting moved into editable blocks inside the chat response. Structured outputs, meanwhile, remain the API route to machine-readable content for your CMS and other systems. Writing for these surfaces now takes two separate skills: briefing a model for a draft a person will edit, and specifying fields a system will consume. This guide explains both, shows where each fits, and gives your B2B content team a repeatable method that protects quality from first prompt to published page.
Writing for canvas now means writing for its successor. Writing blocks in ChatGPT are editable areas for draft text that sit inside the response, and they serve as the ChatGPT canvas replacement for most users. You can edit text directly, ask for a revision of a selected passage or the whole draft, and undo recent AI edits.
OpenAI extended blocks to long-form work in June 2026, including reports and blog posts, with a focused full-screen editor. Available actions vary by plan and device, so test what your team has.
One lesson from the original canvas launch still applies. OpenAI trained the model to make targeted edits when users selected text and to favor rewrites otherwise. Precise selections protect approved paragraphs.
OpenAI structured outputs are an API feature that constrains a response to a JSON Schema you supply. A JSON Schema is a formal description of the fields, data types, and allowed values a response must contain. The model cannot skip a required key or invent a value outside your list.
Think of a writing block as a shared draft on a desk, and a schema as a form with fixed boxes. The draft invites judgment, and the form demands consistency. When OpenAI launched the feature in 2024, it reported that its new model scored 100% on complex schema evals, compared with under 40% for gpt-4-0613.
Yes. JSON mode only ensures the response is valid JSON, while structured outputs also enforce your schema. OpenAI recommends structured outputs wherever they are supported. The risk for content teams: JSON mode can return tidy output with the wrong fields, which breaks your publishing script.
The deciding question is who reads the output first. A person needs an editable draft, while a system needs predictable fields.
| Factor | Writing blocks | Structured outputs |
|---|---|---|
| First reader | An editor reviewing a draft | A CMS, script, or application |
| Best for | Blogs, reports, proposals, emails | Metadata, FAQs, content briefs, product descriptions at scale |
| Control method | Direct edits and targeted revisions | Fixed fields, types, and allowed values |
| Access | ChatGPT interface | API, with developer setup |
| Main risk | Polished but generic prose | Correct shape with incorrect facts |
Follow this order, because each step reduces the rework in the next one.
A weak prompt says, “Write a blog on cloud migration.” A stronger one says, “Outline a guide for CIOs at mid-sized banks comparing phased and big-bang cloud migration. Use only the attached notes. Flag any claim without a source.”
Writers rarely build the schema, but they should define what goes in it. OpenAI’s own schema tips favor clear key names, clear descriptions for important keys, and evals to compare structures. Hand your developers a field brief built on these rules:
A response may not match your schema if the model refuses or hits a token limit, so route those cases to a person instead of your CMS.
Not directly. Search engines and LLMs assess what you publish, not how you produced it. Published structure still matters, though. A Semrush study of 304,805 LLM-cited URLs associated citation with clarity and summarization (+32.83%), E-E-A-T signals (+30.64%), and section structure (+22.91%). These are associations, not proof of cause.
Format should also follow intent. Wix Studio research reported by Search Engine Land found that articles took 45.5% of citations on informational queries. For the page-level detail, read our guide on building AI-ready content.
The two tools work best in sequence. People own the judgment calls, and schemas handle the repeatable parts.
Canvas changed, but the discipline behind good AI-assisted writing did not. Brief the model with evidence, edit by selection, and define fields with the same care you give a headline. Headquartered in Noida, India, TIS works with B2B teams across India, the US, and Europe through AI-powered content services and ChatGPT SEO services. Talk to our team about building a content workflow your editors and systems can both trust.
Related article: FAQ Schema Is Not AEO: What Actually Gets Your Answer Selected
Canvas is no longer part of GPT-5.5 Instant or GPT-5.5 Thinking, following OpenAI’s update on May 28, 2026. Drafting and code editing now happen in blocks inside the chat response. Zapier reported that paid users could reach canvas through legacy models for a limited time, so availability may differ by account. Check your own model picker before building a team process around the older side panel.
Writing blocks are editable drafts inside ChatGPT, made for people who want to revise text directly. Structured outputs are an API feature that forces a model’s response to match a JSON Schema, made for software that needs predictable fields. Use a block when an editor will shape the final copy. Use a schema when a CMS, script, or application will consume the result without manual reformatting.
Writers do not need to code, but they should own the field definitions. A useful approach is a simple table listing each field name, its purpose, its allowed values, and a plain English description. Developers then convert that table into a schema. This split works well because writers understand what good output looks like, while developers handle API setup, error handling, and testing inside your systems.
No. JSON Schema defines the shape of data a model must return through an API. Schema markup uses schema.org vocabulary on a web page to describe its content to search engines. The two serve separate purposes and never replace each other. Google also states that structured data is not required for its generative AI search features, though it still supports eligibility for rich results.
It can, if the published page is genuinely useful. Google allows AI assistance when the work meets its Search Essentials and spam policies, and its guidance prioritizes original, experience-led content over recycled summaries. Citation is never guaranteed on any platform. The strongest approach is to add first-party evidence, named expertise, and clear answers, then have a subject expert review every page before it goes live.