Which Schema Types Actually Improve AI Citations (and Which Don't)

Schema markup is one of the few AEO investments that can be scoped precisely, which is what makes the prioritization question tractable. A content team can decide to ship four schema types and skip eight, and the decision holds for years. The difficulty is that most guidance treats every type as equally worthwhile, which is not what the available data suggests. The best schema types for AI visibility cluster at the top of a fairly steep curve, and several widely recommended types sit at or below the baseline for pages carrying no markup at all.

This piece ranks the major schema types by the strength of the evidence connecting them to AI citations, and identifies the ones that do not justify engineering time. It is a prioritization overview rather than an implementation walkthrough, and it does not revisit whether structured data influences AI answers at all, a question the evidence review on schema and AI citations treats separately.

One caveat belongs at the top rather than buried in a closing section, because it reframes everything below. The strongest quasi-experimental test available, an Ahrefs difference-in-differences analysis of 1,885 pages that newly deployed JSON-LD against roughly 4,000 matched control pages, found no positive effect. Citations in Google AI Overviews declined slightly and significantly over the 30-day measurement window, while changes on AI Mode and ChatGPT were statistically indistinguishable from zero. Adding markup to a page that already ranks well and already earns citations does not appear to move the needle. The type-level ranking that follows should therefore be read as guidance on how to build pages correctly from the outset, not as a list of retrofits that will lift an established page.

The Best Schema Types for AI Citations at a Glance

The table below consolidates the type-level pattern across the four datasets that report it. Citation rates come from an AirOps analysis of roughly 353,800 pages and describe how often pages carrying each type were cited by ChatGPT; they are descriptive correlations, not measured lifts. The odds ratios come from a logistic regression on 1,508 real-estate sites and control for other technical factors.

Two patterns in that table deserve attention before the type-by-type detail. The spread between the top and bottom types is roughly 16 percentage points, which is a meaningful gradient rather than noise. And the bottom of the table sits below the overall no-schema baseline, which in the same dataset was near 32% against 38.5% for pages carrying JSON-LD of any kind. A schema type performing below the baseline is not evidence that the type causes harm; it is evidence that the type tends to appear on page templates that are cited less often for unrelated reasons.

How to Read the Evidence Behind the Ranking

The datasets underneath this ranking differ in what they can support, and conflating them produces the overconfident advice that dominates this topic.

The AirOps figures are descriptive. Pages that ship structured data are disproportionately pages built by teams with technical maturity, editorial standards, and link equity, so the correlation between markup and citation is real while the causal direction remains unestablished. The regression on real-estate sites is a cleaner controlled signal because it holds other technical factors constant, though its sample is a single vertical in a single market, which limits how far the finding generalizes. The Ahrefs analysis is the only design that isolates the effect of adding markup, and it found no positive effect.

The reconciliation is straightforward once the mechanism is examined rather than assumed. A live fetch test across seven AI platforms found that only Gemini reliably parsed JSON-LD during retrieval; six of the seven, ChatGPT and Perplexity among them, treated the markup block as ordinary page text. This is the most direct available evidence on the question of which schema helps AI citations, and it points away from a direct mechanism for most engines. Schema's influence runs largely through Google's index and entity understanding, which then shapes what the engines drawing on those results retrieve and cite. That indirect path is slower and weaker than a direct one, and it explains why a 30-day measurement window would find nothing.

The practical implication is that structured data types for LLMs should be treated as foundational hygiene rather than as a citation trigger. Schema on thin content does nothing measurable. Schema on authoritative content functions as a tiebreaker. This is consistent with the broader research on how LLMs select and extract citations, where extractable content and clear structure carry more weight than any single markup signal.

Tier One: The Foundational Schema Types for AEO

These four types belong on the sites of nearly every organization publishing content, and they represent the defensible core of any schema markup priority list.

BreadcrumbList

BreadcrumbList shows the highest observed citation rate in the AirOps dataset and appears in the top-performing combination in a separate analysis of citation lift by schema pairing. The mechanism is hierarchical rather than semantic: breadcrumbs reinforce site structure and the position of a page within it, which Google uses directly for result display and entity context.

The case for it rests less on effect size than on cost. Breadcrumbs deploy sitewide from a template, they require no per-page editorial work, and they carry no risk of mismatch between markup and visible content. The honest read is that its position at the top of the table partly reflects the kind of site that ships breadcrumbs, namely one with a considered website structure for AI search. The correlation is confounded, and the type is still worth shipping, because near-zero marginal cost makes the effect size question close to irrelevant.

Organization

Organization schema consolidates brand name, logo, contact details, and linked social and reference profiles into a single canonical entity definition. Its citation rate sits near the top of the observed range, though no controlled test isolates its contribution.

Its value is structural and long-term rather than page-level. Entity disambiguation is the precondition for a brand being named accurately in a generated answer, and the mechanism connecting markup to that outcome runs through Google's Knowledge Graph, which several engines inherit understanding from. Organizations with ambiguous names, recent rebrands, or common-word brands have the most to gain, since the failure mode being corrected is the engine conflating the brand with a different entity. The related question of how LLMs recognize and match entities explains why this matters more than its modest citation-rate delta suggests.

Article, NewsArticle, and BlogPosting

Article markup and its variants supply authorship, publication date, headline, and primary image. These are the fields that feed freshness and authorship assessment, and they align with the signals answer engines appear to weight when selecting among comparable sources.

One qualification matters for expectation-setting. Google's own documentation indicates that Article rich results are no longer surfaced in traditional results for most publishers, so the remaining value accrues on AI surfaces, where the effect is indirect. Publishers should still ship it, because the marginal cost on a templated blog is negligible and the fields it exposes are the ones that recency and freshness signals depend on. Teams evaluating this specific type in more depth can consult the dedicated treatment of Article schema for blog posts.

FAQPage

FAQPage carries the strongest single piece of type-level evidence in the entire body of research. In the controlled regression, sites with FAQPage markup showed odds of ChatGPT surfacing information about them roughly thirteen times higher than sites without it, holding other technical factors constant, and the coefficient was significant at conventional thresholds. Its observed citation rate sits near the top of the descriptive dataset as well.

The mechanism is plausible without being established. Question-and-answer pairs map onto a structure that language models handle natively, and they align with the retrieval patterns that conversational queries produce. The complication is that ChatGPT does not parse the markup during retrieval, per the fetch test, so the association is almost certainly running through the visible content that FAQPage markup tends to accompany rather than through the JSON itself. Pages with genuine FAQ sections are pages with self-contained question-shaped passages, and those passages are what gets extracted.

That reading has a direct operational consequence. FAQPage markup on a page without visible question-and-answer content captures none of the benefit and invites a policy problem, since Google has issued spam guidance against markup that does not correspond to what the reader sees. The investment that pays is the citable FAQ section itself, with the markup describing it accurately.

Tier Two: The Context-Dependent Types

These types are correct investments for the sites they fit and wasted effort everywhere else. Their placement in a schema markup priority list depends entirely on the template mix of the site.

Product exposes price, availability, identifier, and rating, which are fields that shopping-oriented AI features extract actively. It shows an odds ratio around four in the controlled regression and a citation rate near 40% in the descriptive data, and a separate analysis reports schema-marked product pages being cited at over twice the rate of unmarked equivalents. For commerce, this is a tier-one type wearing a tier-two label; for everything else, it does not apply.

HowTo carries the largest reported single-type lift in the research, with one analysis attributing a citation increase above 100% for procedural content relative to unmarked equivalents. The mechanism is legible: discrete steps map cleanly onto the step-by-step format answer engines prefer for procedural queries. The constraint is strict. Content must be genuinely step-based, and applying HowTo to non-procedural pages is a quality violation rather than a shortcut.

LocalBusiness has no large correlation study behind it and strong practitioner consensus in front of it. It feeds local entity understanding and the location-specific answers that engines generate for proximity-dependent queries. Schema alone is insufficient here, since the markup must agree with the business profile and directory data it is competing against.

Review and AggregateRating appear in the top-performing combinations and are frequently nested inside Product markup. The binding constraint is policy rather than performance: reviews must be genuinely collected from customers, and self-served or incentivized ratings violate Google's requirements.

WebSite with a SearchAction property sits near the top of the observed citation rates, which almost certainly reflects that the pages carrying it are homepages. Google retired the rich result it once powered, so it is best classified as basic hygiene on a single page rather than an active lever.

Tier Three: The Types to Skip

Three types recur in schema recommendations without evidence supporting their inclusion.

SoftwareApplication posts a citation rate near 33% in the descriptive dataset, below the overall JSON-LD average and close to the no-schema baseline. This is the clearest negative signal in the type-level data. A SaaS marketing page is not an app listing, and marking it as one supplies fields the engines have little use for in that context. The type is defensible on a genuine software download or app store entry and difficult to justify elsewhere, which is a notable finding given how routinely it is recommended to SaaS teams.

Person used standalone posts the lowest citation rate observed, near 30%. Author identity does matter for expertise assessment, but the value is realized when Person is nested as the author property of Article markup rather than deployed as a page type of its own. Author bio pages are rarely the pages an engine wants to cite, and the E-E-A-T signals that author identity supports are established through the byline relationship, not the standalone entity.

Speakable remains a beta feature in Google's documentation, scoped to voice and news readouts, with no documented adoption by any AI engine. Outside news publishing, it does not warrant consideration in 2026.

VideoObject is a partial exception rather than a skip. It does not appear in the top tier of the citation data, and its value is eligibility for video results rather than a documented citation lift. Pages where video is central should ship it; pages where video is incidental should not.

Why the Ranking Shifts by Engine

Any type-level ranking averages across engines that behave differently, and the averaging hides a distinction that matters for prioritization.

The consequence is that schema's direct effect is concentrated on Google surfaces, and its effect elsewhere is mediated by rankings. The answer to which schema helps AI citations therefore depends on which engines a given audience actually uses. For a team whose buyers ask questions primarily in ChatGPT, structured data types for LLMs are a second-order investment behind content structure and extractability. For a team whose visibility depends on AI Overviews and AI Mode, schema sits closer to the critical path. These claims describe engine behavior as of mid-2026, and the direction of change is toward more parsing rather than less, so the ranking should be revisited rather than treated as permanent.

Which Types to Prioritize by Site Type

The defensible reading of the evidence, applied by situation:

• Publishers and content-led blogs. Article, Organization, and BreadcrumbList sitewide, Person nested as the Article author, FAQPage on pages with genuine Q&A sections. Skip VideoObject unless video is core to the format.
• E-commerce. Product with Offer and AggregateRating on detail pages, BreadcrumbList and Organization sitewide. SoftwareApplication only on actual software SKUs.
• SaaS and B2B. Organization, BreadcrumbList, Product on the pages with transactional intent, FAQPage where real questions are answered, Article on editorial content. SoftwareApplication is the type most often recommended and least often justified here.
• Local and multi-location businesses. LocalBusiness and Organization, BreadcrumbList, FAQPage, and Review where verified reviews exist.

Across all four situations, two schema types for AEO carry nearly every site, namely BreadcrumbList and Organization, with a third determined by the dominant template, namely Article, Product, or LocalBusiness. That is a shorter list than most schema guidance produces, and the evidence does not support a longer one. Beyond two or three well-matched types per page, the combination research shows diminishing returns, and mismatched types introduce quality risk rather than incremental gain.

The larger point is one of sequencing. Schema is a foundational signal, correctly built once and rarely revisited. The variable that separates cited pages from uncited ones is the content underneath the markup: whether it answers the question directly, whether its passages survive extraction, and whether the brand is a recognizable entity in the first place. Teams that treat structured data as the AEO strategy rather than its substrate consistently find that the markup validated cleanly and the citations never arrived. The technical SEO checklist for LLM retrieval situates schema among the other retrieval prerequisites it works alongside.

Verifying the Decision Against Real Citation Data

A prioritization decision of this kind is testable, and the test is not schema validation. Validators confirm that markup parses; they say nothing about whether an engine cited the page. The measurement that resolves the question is citation rate by page and by prompt, tracked across engines over a period long enough for indirect effects to propagate, which the research suggests is considerably longer than 30 days.

DeepSmith tracks that directly. Its AI Visibility module monitors mention rate, citation rate, and share of voice across ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode, reports which specific pages earn citations and which prompts drive them, and shows which competitor pages win the prompts a brand is targeting. Schema is one input among many, and the platform does not implement markup; what it provides is the outcome data that indicates whether any of these investments, schema included, changed what the engines cite. Teams that want that baseline before committing engineering time to a markup project can start a free trial and see real citation data for their own prompts.

Ranking schema types is worthwhile, and it remains a smaller decision than the ranking's popularity suggests. The best schema types for AI visibility are cheap enough that the argument for shipping them does not depend on effect size, and the weakest are cheap enough that skipping them frees only modest resources. The consequential work sits in the content the markup describes, and in measuring AI search citations well enough to know which changes mattered.