Schema Markup — Angelina Shylkina

Quick read · 30 seconds

The whole case, in four lines.

Problem

A new feature in a domain nobody on the team knew — the structured data behind SEO and AI discovery.

Insight

Six sessions surfaced two opposite users: marketers who feared schema, experts who distrusted AI writing it.

Design

Two doors, one flow. AI generation or manual entry, a type picker to sharpen the output, a validation step, and guidance in the panel.

Result

~78% publish once they start. Getting users into the feature is the next constraint.

02 · Context

Why the business wanted this.

The product is a landing-page builder, built for paid campaigns. Budgets were tightening and AI assistants were starting to decide what to surface from structured data — so organic discovery became the cheapest growth left.

Fig. 01 Schema is invisible to the visitor and central to being found.

What schema markup actually is

A block of code that tells search engines what the content on a page is. Without it, Google sees plain text and has to guess what is what.

A JSON-LD schema block wrapped in a script tag, describing a person, a music group, a location and a product as structured properties — **What the machine reads**Fig. 02 JSON-LD embedded in the page — every fact labelled with a type.

What the person seesFig. 03 The same page in search, with and without schema.

04 · Learning the domain

Nobody on the team had done schema before.

The team started this project without prior expertise in schema markup. So the first decision wasn't what the UI should look like — it was how to learn enough to make the right call.

Illustration: a team with a question mark climbs four steps labelled learn, research, understand and discuss, arriving at a validated block of code — Fig. 05 Learn, research, understand, discuss — before a single screen.

How we tested

Six moderated studies, run directly with users on a clickable prototype.

Illustration of a moderator taking notes at a laptop during a video call with six participants — **6 moderated studies**Run directly with users, remote and recorded.

Illustration of a numbered interaction script document with three numbered tasks and a checklist — **Interaction script**Find schema in settings, choose a build method — manual or AI — then save it.

Illustration of a Figma file with a prototype interactions panel open beside a wireframe page — **Clickable prototype**Generated in Figma AI as the artifact under test.

How the study was set up

Research goals. Evaluate usability and clarity, and learn how familiar users are with structured data.
Defined objectives. Identify what guidance and validation people expect before they trust the output.
Understood personas. Two skill levels — Norma and Diego — needing very different depths of guidance.
Invited participants. Six people: 1 SEO lead, 2 former SEO marketers, 2 support, 1 designer. Basic and advanced experience.

05 · What we learned

What the sessions surfaced.

4/5 Overall impression of the prototype. What was missing wasn't layout — it was comprehension and trust.

Schema feels technical.

Several people confused it with page layout or headings.

Confidence depends on experience.

Experienced users trusted it quickly. Newcomers needed more reassurance.

Validation happens elsewhere.

People want to check their work without leaving the product.

Save vs. publish is unclear.

People couldn't tell whether saving also published the schema live.

In the sessions · 3 insights

01 / 03

Confidence depends on the persona

Experienced users arrive with clear expectations of how schema should work and what it should return. People with little experience arrive curious, looking for guidance and a plain explanation of the point of it.

"Something I've heard of but never worked with."

Participant · basic experience

The schema markup section of the product with Generate schema and Type in manually buttons, annotated with a participant quote: something I have heard of but never worked with

They wanted plain explanations, in the product

Users asked for learning resources at the point of action — which meant validation had to live in the product instead of sending people out to Rich Results Test or the Schema Markup Validator. We watched them use both during the sessions.

"…a link to a help center article… before I decide whether my schema is good or not."

Participant · on generated output

Generated JSON-LD schema in the product with regenerate, cancel and save options, annotated with a participant quote asking for a help centre link before deciding whether the schema is good

Trust splits by who wrote it

Schema has to be validated, and experts told us how they already do it in their daily workflow. That gave us a way to serve both sides: a validation path for experts, plain education for everyone else.

"If I clicked generate, I trust it. If I did it manually myself, I don't trust it as much."

Participant · advanced experience

The manual JSON-LD entry field in the product with cancel and save schema buttons, annotated with a participant quote about trusting generated schema more than hand-written schema

Quotes condensed from session recordings. Raw recordings withheld under confidentiality.

06 · Risks into opportunities

Every risk we found had a design answer.

Risks

Norma and Diego need different things. Diego is confident and wants control; Norma isn't sure what schema even is. One generic flow loses one of them.
Users can't tell if their schema is correct. "Valid" means nothing to them, so they fall back on ChatGPT, Google, and schema.org.
No safe line between draft and live. People feared the wrong schema being applied automatically.

Opportunities

Guidance tailored by persona. Simple explanations for Norma, precise control for Diego.
Validation inside the platform. No round trip to an external validator.
A clearer save-and-publish state. Draft and live, named and visible.

Caught cheap, not caught expensive

Caught in discovery

Norma and Diego need different things
Users can't tell if their schema is correct
No safe line between draft and live

If it had reached production

A wave of support tickets from confused users
A distrusted feature, abandoned for ChatGPT
Live pages carrying broken or wrong schema

Takeaway

The feature had to make schema feel approachable to people who didn't know what it was — without dumbing it down for people who knew schema cold and didn't trust AI to write it for them.

Instead of guessing, the team learned in public — with the experts in the room.

07 · The design call

The team wanted AI to do everything. I pushed back.

Takeaway

One wrong AI output and we'd lose the experts — the loudest voice in any SEO community. So neither track got to be the secondary one.

Dual-track flow diagram: the marketer track picks a type and reviews AI-generated schema; the specialist track hand-edits JSON-LD with validation; both converge on the same published schema — Fig. 06 Two entry points, one shipped output — the pattern the whole feature rests on.

What shipped

The feature lives in the page's SEO settings, behind a plain-language explainer and a learn-more link. From that one panel a marketer can have schema generated from a bounded list of types; a specialist can write their own. Both paths land in the same editor, with a validation step, before anything is saved.

Fig. 07

Two doors, equal weight

Generate schema and Add schema sit side by side — same size, same position, inside the page's SEO settings.

The SEO settings panel in the shipped product: a schema markup section with a short explanation, a learn-more link, and two buttons — Generate schema and Add schema

Fig. 08

The manual door

Write or paste your own JSON-LD. The expert path, not a fallback tucked behind the AI one.

The manual path in the shipped product: an Enter your schema panel with an empty numbered JSON editor, a note not to include script attributes, and a validate link

Fig. 09

Constrained type picker

A curated list of page types bounds what the AI is allowed to write. The full schema.org vocabulary is never exposed.

The shipped schema type picker open, listing Auto-detect, Event, Organization, Person, Place, Local Business, Restaurant, Product and Offer, with an explore schema types link

Fig. 10

Auto-detect for anyone unsure

The AI reads the page and proposes the type — the default for Norma, who doesn't yet know which type she needs.

The shipped schema type field set to Auto-detect, with cancel and continue actions

Fig. 11

Editable output

Generated schema lands in the editor rather than on the page. Experts intervene before anything is saved.

The shipped review step: a your schema is ready panel showing generated JSON-LD in a numbered editor, with cancel and regenerate actions

Fig. 12

A validation step before saving

Preventive guidance in the panel — no <script> attributes — plus a one-click check against the spec.

The shipped review panel with the Validate in Schema.org link circled, alongside the preventive note about script attributes

Shipped UI, captured on demo pages — no customer data shown.

The risk with dual-track flows is dithering at the fork. We mitigated it by keeping both doors reachable from each other, and by unifying everything downstream — the editor, the review step, saving. Only the input method differs.

08 · Outcome

Once users are inside, they finish.

Schema funnel · 90-day window

SEO tab opened

~900100%

Schema started

~100~11%

Type picked or JSON pasted

~80~84%

Saved and republished

~75~94%

~78% of users who start a schema publish it. That's the metric the design was built to move.

Rates anchored to post-launch results; absolute counts rounded. Raw numbers withheld.

The dual-track flow works. The type picker isn't a bottleneck, and the review-before-save step removes the "do I trust this output" wall.
Discovery is the next constraint. Only ~11% of SEO-tab visits turn into a started schema — the entry copy has to explain why this user should care.
The handoff held. The partner team shipped the core architecture; persona-aware copy and inline help didn't make it in.

PRD target~20% of new indexed pages carry schema

PRD target~30% of indexed-page users add one

PRD target>1 page per adopter — not a one-time test

One feature, two users who wanted opposite things.

The whole case, in four lines.

Why the business wanted this.

What schema markup actually is

Two users, opposite needs.

Nobody on the team had done schema before.

How we tested

How the study was set up

What the sessions surfaced.

Schema feels technical.

Confidence depends on experience.

Validation happens elsewhere.

Save vs. publish is unclear.

Confidence depends on the persona

They wanted plain explanations, in the product

Trust splits by who wrote it

Every risk we found had a design answer.

Risks

Opportunities

Caught cheap, not caught expensive

Caught in discovery

If it had reached production

The team wanted AI to do everything. I pushed back.

What shipped

Two doors, equal weight

The manual door

Constrained type picker

Auto-detect for anyone unsure

Editable output

A validation step before saving

Once users are inside, they finish.

Schema funnel · 90-day window

What I'd do differently.

Doubling activations by removing friction from a follow-up automation flow.

Cutting template selection time from 15 minutes to 3.

Every export option converted better once people could find it.