Structured Data for AI Citations: The Schema Markup B2B Sites Need

Everyone selling structured data for AI citations quotes the same correlation: pages that get cited by AI engines are far more likely to carry JSON-LD. Ahrefs found cited pages were almost three times more likely to have it than non-cited ones. Then Ahrefs ran the experiment that matters. They tracked 1,885 pages that added schema between August 2025 and March 2026 against roughly 4,000 control pages. Citations in ChatGPT and Google AI Mode moved about 2%, which is indistinguishable from noise. Google AI Overviews citations fell 4.6%.

So does schema markup matter for AI search or not? Yes, but not in the way most guides promise, and not for the reasons they give. I have watched teams spend a full quarter on schema and then wonder why their citation rate did not move. This post is the version I wish they had read first: what the evidence actually supports, which schema types are worth the engineering hours for a B2B site, and a rollout plan that lets you measure the result instead of guessing.

Here is the framework we use, the Schema Evidence Ladder.

What the 2026 Studies Say About Schema and AI Citations

The honest answer is that the evidence conflicts, and the conflict is informative. Four findings matter.

Source	What it measured	Result	How far to trust it
Ahrefs, May 2026	1,885 pages that added JSON-LD vs about 4,000 controls, Aug 2025 to Mar 2026	AI Mode +2.4%, ChatGPT +2.2% (noise), AI Overviews down 4.6%	Best controlled test available
Ahrefs, same study	Cited vs non-cited pages	Cited pages almost 3x more likely to have JSON-LD	Correlation only; good sites do everything well
OtterlyAI	Sitewide rollout on 2,000+ URLs	AI Overviews citations up 1,500%, AI Mode up 377%; down on ChatGPT, Gemini, Copilot; no change on Perplexity	Vendor run, no public methodology to audit
Google documentation	Official guidance	No special markup required for AI Overviews or AI Mode	Authoritative on Google, silent on others

Two caveats change how you should read the Ahrefs result. First, every page in the test already had more than 100 AI Overview citations before schema was added. The study tells you what happens to pages the engines already like. It says little about a page that has never been surfaced, where schema might help with crawling and parsing in the first place.

Second, the study grouped schema types together. It cannot tell you whether Organization markup behaves differently from FAQPage markup. Nobody has published a clean controlled test of individual types yet, and any vendor quoting a precise lift for one type is quoting a number I cannot trace to a primary source.

Microsoft is the one major platform on record saying otherwise. Fabrice Canel of Bing said on stage at SMX Munich that schema helps Microsoft’s LLMs understand content. Since ChatGPT’s search retrieval has leaned on Bing, that matters for B2B buyers who research there. Google says structured data is useful for the features it powers but is not required to show up in AI Overviews.

Put those together and you get my working position. Schema is an eligibility and clarity layer. It makes your pages easier to parse, your brand easier to identify, and your rich results possible. It is not a citation lever, and it will not rescue a page whose opening paragraph buries the answer. If you want the factors that do carry weight, we ranked them in AIO ranking factors, what Google AI Overviews actually weight.

Why the Correlation Misleads Almost Everyone

Teams see “cited pages are 3x more likely to have schema” and conclude schema causes citations. The causation runs the other way, or more precisely it runs through a third thing.

Sites that implement clean JSON-LD tend to have a technical SEO owner, an editorial process, maintained author pages, fast templates and a reason to keep content fresh. Those sites also earn links and write extractable answers. Schema is a marker of operational maturity, and the maturity is what gets cited.

You can test this on your own site in ten minutes. Take your five most cited pages from your last citation check and look at what they share. In my experience it is rarely schema. It is a direct answer in the first 60 words, a named author, a table, and a clear primary source for each number. If you have not run that check, start with our 10 point citation readiness audit; schema is one line item in it, not the whole thing.

That does not make schema useless. It means the right question is not “does it raise citations?” but “what job does each schema type do, and is that job worth the cost?”

The Schema Evidence Ladder

The ladder ranks schema work by how strongly the evidence supports it and how cheaply it ships. Climb in order. Do not start at the top because it sounds advanced.

Tier 1: Identity. Organization and WebSite markup, site wide. This tells every crawler who you are.

Tier 2: Content. Article (or BlogPosting) with author, datePublished and dateModified, plus BreadcrumbList. This tells engines what each page is and how fresh it is.

Tier 3: Entity. Person schema for authors, and sameAs links connecting your brand and people to verified profiles. This helps engines resolve who is speaking.

Tier 4: Extraction. FAQPage and HowTo, where the content genuinely exists on the page. This is the layer most guides oversell.

Tier 5: Skip for now. Speakable, niche vocabulary types, and anything you cannot keep accurate.

Tier	Schema types	Evidence strength	Effort	Ship when
1 Identity	Organization, WebSite	Moderate: Microsoft confirmed, Google uses it for knowledge panels	Low, one template	Week 1
2 Content	Article, BreadcrumbList	Moderate: powers freshness and hierarchy signals	Low to medium	Week 1 to 2
3 Entity	Person, sameAs	Plausible, untested in controlled studies	Medium	Week 2 to 3
4 Extraction	FAQPage, HowTo	Weak for citations, fine for hygiene	Low if content exists	Week 3
5 Skip	Speakable, exotic types	Very weak	Varies	After a measured win

Tier 1: Identity (Organization and WebSite)

Every B2B site should have one Organization block on every page, usually injected from the site template. It should carry your legal or brand name, URL, logo, a short description, and the sameAs array. WebSite markup adds the site name and, optionally, a search action.

Why this tier comes first: engines attribute claims to entities. If your brand name appears on your site as “Acme”, on LinkedIn as “Acme Inc.” and on G2 as “Acme Software”, you are fragmenting your own entity. Organization markup is where you declare the canonical version once.

Here is a minimal Organization block. Keep it boring and accurate.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Example Co",
  "url": "https://www.example.com",
  "logo": "https://www.example.com/logo.png",
  "description": "One plain sentence about what the company does and for whom.",
  "sameAs": [
    "https://www.linkedin.com/company/example-co",
    "https://www.crunchbase.com/organization/example-co",
    "https://www.g2.com/products/example-co"
  ]
}

Common mistakes here: logos behind redirects, sameAs pointing at profiles you do not control or that have been abandoned, and a description written as ad copy. Write it the way a neutral encyclopedia would.

Tier 2: Content (Article and BreadcrumbList)

For blog posts, use Article or BlogPosting with these fields filled in properly: headline, author (linked to a Person), datePublished, dateModified, publisher (your Organization), and mainEntityOfPage.

The field that earns its keep is dateModified. Freshness is one of the few signals with consistent support across studies, and the markup should match a visible “Last updated” line on the page. If the schema says you updated last week and the page shows 2024, you have created a contradiction that costs trust. We cover the visible side of this in the AEO audit.

BreadcrumbList is dull and useful. It tells engines where a page sits in your content hierarchy, which helps when a cluster of posts supports a pillar page.

Tier 3: Entity (Person and sameAs)

This is the tier I would invest in beyond the basics, with the caveat that it is plausible rather than proven.

AI engines answer questions about who knows what. A post with a named author whose Person markup links to a LinkedIn profile, a bio page and two external articles gives an engine something to cross-check. A post attributed to “Admin” gives it nothing.

Build the Person block like this:

{
  "@context": "https://schema.org",
  "@type": "Person",
  "name": "Akif Kartalci",
  "jobTitle": "Founder",
  "worksFor": { "@type": "Organization", "name": "Momentum Nexus" },
  "url": "https://www.example.com/authors/akif-kartalci",
  "sameAs": [
    "https://www.linkedin.com/in/example-profile"
  ]
}

The rule for sameAs: link only to pages that say the same thing about the same entity. If the LinkedIn headline says one thing and your author page says another, fix the profiles before you fix the markup. The markup cannot launder inconsistent facts.

Tier 4: Extraction (FAQPage and HowTo)

FAQPage is the most popular schema recommendation for AI search, and the one with the weakest case. Google restricted FAQ rich results to authoritative government and health sites in 2023 and cut back HowTo rich results at about the same time, so the original SERP payoff mostly disappeared. For AI engines, the controlled evidence is thin.

I still ship it, for a different reason: the questions on the page are good for readers and for passage retrieval, and the markup costs almost nothing when it mirrors visible content. Our own blog template renders the faqs frontmatter as a visible FAQ section and generates the FAQPage markup from the same source, so the two cannot drift apart. That is the standard to hold: one source of truth, markup generated from it, never hand-edited separately.

Rules for this tier:

• Visible or nothing. Every question and answer in the markup must appear on the page.
• Real questions. Use phrases a buyer would type or speak, not a keyword list with question marks.
• Self-contained answers. 45 to 90 words, answer in the first sentence.
• No stuffing. Five honest pairs beat fifteen padded ones.

Tier 5: Skip for now

Speakable was designed for Google Assistant on news content and has little reason to exist on a B2B SaaS blog. The same goes for obscure vocabulary types. If a markup type does not map to something visible on the page or a feature you can name, it is maintenance debt.

Implementation: How to Ship Without Breaking Things

Schema fails quietly. A template change can strip your Organization block on 400 pages and nothing alerts you. Treat it as production code.

Step 1: Inventory what exists

Crawl your site and extract every JSON-LD block. Count types per template: homepage, blog post, product page, author page. Most B2B sites find three problems: duplicate blocks from a plugin plus a theme, outdated sameAs links, and Article markup with no author.

Step 2: Choose one source of truth

Generate schema from your CMS fields, not from a plugin that guesses. The author field feeds Person. The updated date feeds dateModified and the visible date. The FAQ fields feed both the on-page section and FAQPage. One field, two outputs, zero drift.

Step 3: Validate before and after deploy

Use Google’s Rich Results Test and the Schema Markup Validator on one URL per template. Then add an automated check to your release process that fails the build when a required type is missing from a template. A lightweight check is enough: fetch the page, parse the JSON-LD, assert the types exist.

Step 4: Match markup to page

Open five pages and compare the markup with the visible text. Titles, dates, author names, and FAQ text should be identical. Mismatches are the single most common reason markup is ignored.

Step 5: Document the contract

Write down which schema each template emits and who owns it. When a redesign happens in eight months, that document is the difference between a smooth migration and a silent loss.

A 30 Day Rollout That Lets You Measure the Result

The mistake in most schema projects is changing everything at once, then crediting or blaming schema for whatever the citations did. Here is a sequence that produces evidence.

Week	Work	Measurement
1	Run a baseline: 20 to 30 buyer questions across ChatGPT, Perplexity and Google AI Overviews. Record which of your pages get cited. Inventory existing schema.	Baseline citation count and share of voice
2	Ship Tier 1 and Tier 2 on half your blog posts, chosen at random. Keep the other half untouched as a control.	Validation passes on all treated templates
3	Ship Tier 3 on author pages and treated posts. Add FAQPage where visible Q&As already exist.	Rich Results Test clean; no markup warnings
4	Re-run the prompt set. Compare treated and control groups.	Citation delta, treated vs control

Give it 4 to 6 weeks before judging, because engines recrawl on their own schedule. Keep model changes in mind: if ChatGPT updates its retrieval in week three, your whole set moves and only the control group tells you what was schema. For the monitoring side, including how to track brand mentions without paying for a tool, see our guide to AI citation monitoring.

If the treated group beats the control, you have your own evidence for your own site, which is worth more than any vendor chart. If it does not, you have still gained cleaner entity data and rich result eligibility, and you can stop arguing about it.

Where Schema Fits Next to the Levers That Do Move Citations

Schema competes for engineering time with changes that have stronger support. Keep the order straight.

Lever	Evidence	Typical effort
Direct answer in the opening 60 words	Strong across multiple analyses	Hours per page
Inline citations to primary sources and specific statistics	Strong; Ahrefs content tests showed gains	Hours per page
Named authors with verifiable credentials	Moderate	Days, once
Freshness with a visible updated date	Moderate to strong	Quarterly review
Identity and Content schema (Tiers 1 and 2)	Moderate, mostly via parsing and eligibility	1 to 2 weeks
Extraction schema (Tier 4)	Weak	Days

If your opening paragraphs still start with three sentences of context before the answer, fix that before touching JSON-LD. The full playbook for those levers is in our answer engine optimization guide.

Mistakes I See B2B Teams Make

1. Treating schema as the strategy. It is a hygiene layer. A quarter spent on markup with unchanged content produces tidy code and flat citations.
2. Marking up content that is not on the page. Hidden FAQs, invented review ratings and fake author credentials get ignored at best and penalized at worst.
3. Letting two systems write schema. A theme and an SEO plugin both emitting Organization blocks with different logos is common. Pick one.
4. Quoting vendor lift numbers. “Up to 30% more visibility” with no method is marketing. Run your own control group.
5. Forgetting dateModified on refresh. You update the stats, the visible date changes, the markup does not. Now the machine-readable freshness signal is wrong.
6. Pointing sameAs at dead or unmanaged profiles. An abandoned Crunchbase page with an old name teaches the engine the wrong thing.
7. Never re-validating. Template changes break schema silently. Add an automated check.

What to Do This Week

Pick one template, your blog post template, and open a post in the Rich Results Test. Check three things: is there an Article block with author, datePublished and dateModified; does it link to an Organization with a working sameAs; do the FAQ pairs match what a reader sees. Fix whatever fails, then set up your 20 to 30 question baseline so the next change is measurable.

Structured data is cheap insurance, not a growth engine. Ship the tiers in order, hold back half your pages as a control, and let your own numbers decide how much more to invest.

If you want help working out where your site stands on citation readiness, book a free growth audit and we will map the gaps against your specific pages. You can also try our free AI growth tools at app.momentumnexus.com.