HowTo Schema vs Plain HTML Steps for AI Citations (2026)

HowTo structured data (JSON-LD HowTo / HowToStep) and plain HTML numbered steps both try to make procedures machine-readable. In 2026, generative systems that cite sources still prefer pages they can fetch as real HTML, quote without contradiction, and map to a clear entity. Schema helps when it mirrors the visible page. Schema that invents steps, hides behind a JS shell, or disagrees with the article body does not create citations—and can confuse extractors.

This guide compares HowTo JSON-LD versus plain step lists for GEO/AEO work aimed at Gemini, OpenAI (ChatGPT), Claude, Grok, and DeepSeek. It complements (does not replace) FAQPage schema for GEO and JSON-LD TechArticle patterns.

GEOcheck.ai (ThinkPrompt Co., Ltd) scores those five engines. Perplexity is a citation target, not a scored engine. Public CTAs: homepage and homepage AI visibility analyzer. Sister products: Doctranslate.io, Mangaka.app. Distinct from geocheck.cc / GeckoCheck.

What each format is for

Format	Best for	Weak when
Plain HTML <ol> / numbered headings	Universal extractability; works even if JSON-LD is ignored	Buried under fluff intros; steps split across tabs
HowTo JSON-LD	Explicit step graph for parsers that consume schema	Orphan schema, mismatch with body, or non-procedural pages
FAQPage JSON-LD	Q&A pairs and objection handling	Used as a dumping ground for marketing claims
TechArticle / Article	Method posts and definitions	Empty articleBody or client-only rendering

Classic rich-result eligibility in Google Search is not the same as generative citation. Treat Search Gallery docs as a hygiene reference, not a GEO guarantee.

What AI crawlers and fetchers actually need first

Before schema debates:

1. SSR or static HTML with the steps in the first response (crawlable HTML for GEO)
2. robots Allow for the user fetchers that matter (OpenAI crawlers; training vs user fetchers)
3. Sitemap membership without soft-404 UUID noise (sitemap hygiene; soft-404 UUID pages)
4. Answer-first lead so the procedure is not below the fold of fluff (answer-first structure)

Schema cannot rescue a JS shell. Plain HTML steps on a crawlable URL beat perfect HowTo JSON-LD on an empty #root.

Plain HTML steps: the durable default

Use visible structure:

<h2>How to build a five-engine visibility panel</h2>
<ol>
  <li><strong>Lock a prompt set version</strong> — …</li>
  <li><strong>Run identical prompts</strong> on Gemini, OpenAI, Claude, Grok, and DeepSeek.</li>
  <li><strong>Score mention vs citation</strong> — …</li>
</ol>

Why this wins often:

• Extractors that ignore JSON-LD still see an ordered list
• Humans and models share the same source of truth
• Diffs and refreshes are obvious in HTML
• Internal anchors can deep-link to step headings (internal linking)

Requirements:

• Steps are complete in text (not “see video at 3:12” only)
• No contradictory alternate procedures on the same URL
• Dates or changelogs when steps change (citation decay / freshness)

HowTo JSON-LD: when it helps

Add HowTo when the page is a procedure and you want an explicit graph:

{
  "@context": "https://schema.org",
  "@type": "HowTo",
  "name": "How to build a five-engine AI visibility panel",
  "description": "Run a frozen prompt set across Gemini, OpenAI, Claude, Grok, and DeepSeek.",
  "step": [
    {
      "@type": "HowToStep",
      "name": "Lock a prompt set version",
      "text": "Version the prompt file and freeze wording for at least four weekly runs."
    },
    {
      "@type": "HowToStep",
      "name": "Run identical prompts on five engines",
      "text": "Execute the same set on Gemini, OpenAI, Claude, Grok, and DeepSeek."
    }
  ]
}

Rules that keep HowTo from becoming noise:

1. Mirror visible steps — same order, same meaning, no extra “secret” steps
2. One primary HowTo per URL — do not stack three competing procedures
3. Keep names short and factual — avoid keyword stuffing in name
4. Prefer text that stands alone — scrapers often drop images/videos
5. Validate — broken JSON helps nobody

Pair HowTo with Organization / Article signals when the brand entity is contested (Organization schema / sameAs; brand entity optimization).

HowTo vs FAQ vs plain steps: choose by intent

Buyer intent	Prefer
“How do I measure …?”	Plain steps + optional HowTo
“What is … / Is X true?”	Definition block + FAQPage
“Best tools for …?”	Comparison table + TechArticle; HowTo secondary
“Why did my citations drop?”	Diagnostic checklist (plain HTML); schema optional

Stuffing HowTo onto a comparison roundup rarely improves citations. Stuffing FAQ onto a pure procedure can help objections—if the Q&A matches the visible FAQ section.

Engine-aware notes (practical, not mystical)

• Gemini — Benefits from clean HTML and consistent structured data that Search-related pipelines already understand; still verify SSR.
• OpenAI — User fetchers need live step text; schema alone does not prove fetchability.
• Claude — Often synthesizes procedures from prose; clear numbered lists reduce paraphrase errors.
• Grok / DeepSeek — Compete on public density; thin HowTo with three vague steps will lose to a rival’s detailed guide.
• Perplexity — Watch whether your how-to URL appears among sources; keep that canary separate from product scores.

Measurement: did schema move citations?

On a frozen how-to prompt class:

1. Record citation URL (if any) per engine
2. Note whether the cited URL is your how-to page, a FAQ, a third party, or none
3. After shipping HowTo JSON-LD that mirrors existing HTML, wait for fetch/re-crawl windows, then re-run the same prompts
4. Attribute wins only when the cited URL is yours and the step content matches

Do not claim “schema caused +X% AI traffic.” Report citation rates and cited URLs.

Implementation checklist

• [ ] Procedure lives in visible HTML first
• [ ] HowTo JSON-LD mirrors those steps exactly (or omit HowTo)
• [ ] No HowTo on non-procedural pages
• [ ] FAQPage used only where visible FAQs exist
• [ ] Page returns article body without JS execution
• [ ] robots/sitemap allow the URL without soft-404 neighbors
• [ ] Entity strings match homepage and Organization schema
• [ ] CTAs point to geocheck.ai / homepage AI visibility analyzer—not retired funnels
• [ ] Panel re-run after publish; results logged by prompt set version

Anti-patterns

1. Schema-only steps — JSON-LD present, body empty or behind interaction
2. Mismatch — schema says 5 steps, page shows 8 different ones
3. HowTo on every blog — dilutes signal; looks spammy to humans too
4. Video-only procedures — inaccessible to many fetchers
5. Invented time/cost fields — hallucinated totalTime / estimatedCost that marketing made up
6. Ignoring FAQ vs HowTo fit — wrong type for the intent

FAQ

Does Google still show HowTo rich results?

Eligibility and UI surfaces change. Regardless of rich-result UI, keep HowTo honest and mirrored—GEO value is extractability and consistency, not a guaranteed SERP enhancement.

Should we drop FAQPage if we add HowTo?

No. Different intents. A guide can have a procedure (HowTo) and a short FAQ section (FAQPage) if both are visible and truthful.

Is Microdata better than JSON-LD?

JSON-LD is easier to maintain for most teams. Consistency with the visible page matters more than the encoding flavor.

Will HowTo help zero-click answers?

It can improve accuracy of narrated steps even when chips are hidden. Pair with mention/accuracy scoring (zero-click brand demand).

Next step

Pick one high-traffic how-to URL. Confirm SSR shows numbered steps. Add HowTo JSON-LD only if it mirrors those steps 1:1. Validate JSON, request re-crawl where you can, then re-run your how-to prompt class on Gemini, OpenAI, Claude, Grok, and DeepSeek. If citations still miss you, fix density and entity clarity before adding more schema types.

Start from the homepage AI visibility analyzer and keep category peers on the homepage AI visibility analyzer.