Schema-Rich vs Plain Listings: AI Recognition 2026

When Google, Microsoft, Yahoo and Yandex jointly announced Schema.org in June 2011, the structured data movement acquired what amounted to a state religion. Webmasters who had spent the prior decade fighting through microformats and RDFa now had a single vocabulary, blessed by the gatekeepers, and a clear promise: mark up your content properly and the search engines will reward you with richer presentations and better rankings. For roughly twelve years that bargain held. Then, somewhere between the launch of GPT-4 in March 2023 and the consolidation of generative search interfaces through 2024 and 2025, the underlying mechanics of how machines extract meaning from web pages changed — and the schema gospel began showing cracks that its preachers were slow to acknowledge.

The argument advanced here is contrarian but not nihilistic. Schema markup remains genuinely useful for several well-defined verticals. What the evidence increasingly suggests, however, is that the blanket prescription — “mark up everything in JSON-LD or lose visibility” — was always a vendor-friendly oversimplification, and in 2026 it has become measurably wrong for a large share of content types. The question is no longer whether to add schema, but where it earns its keep and where it merely consumes engineering hours that could be spent on prose quality, citation hygiene, or factual depth.

The Schema Markup Gospel

The 2020s SEO Doctrine

The dominant doctrine that crystallised between roughly 2018 and 2023 ran something like this: search engines are pattern-matching systems that struggle with unstructured text, and the safest hedge against algorithmic uncertainty is to encode every possible entity and relationship in JSON-LD. Conferences ran tracks on it. Agencies built service lines around it. Audit tools — Screaming Frog, Sitebulb, Ahrefs — added schema validators and treated red flags in those validators as defects of comparable severity to broken canonical tags or missing hreflang. The doctrine produced real results during the era it was designed for, which is precisely why it became so entrenched.

The doctrine’s strength was also its limitation. It assumed a retrieval architecture in which a crawler fetched a page, parsed it shallowly, and matched discrete fields to a knowledge panel template. That architecture remained dominant through the BERT updates of 2019 and the MUM announcements of 2021, but it was already being supplemented — and in some cases supplanted — by language models that read pages the way a literate human reads them. The doctrine, codified in industry training materials that practitioners are still reading in 2026, did not adapt at the speed the underlying technology did.

Why Everyone Bet on JSON-LD

JSON-LD won the format war against microdata and RDFa for entirely defensible reasons. It separates structured data from presentational HTML, which suits modern component-based frontends. It nests cleanly. It validates against published vocabularies without forcing developers to interleave attributes through their templates. Google publicly preferred it from late 2015 onward, and that preference effectively ended the debate. By 2020 most enterprise CMS platforms shipped JSON-LD generators by default, and by 2023 the format was so naturalised that questioning it felt akin to questioning HTTPS.

The deeper bet, however, was not on the format itself but on the proposition that machines would always need explicit type signals to understand content. That proposition rested on assumptions about machine reading comprehension that were reasonable in 2016 and demonstrably outdated by 2024. Research from Springer on schema-rich heterogeneous information networks (Sun et al., in Schema-Rich Heterogeneous Network Mining) established years earlier that knowledge graphs derived from object-relation-object tuples can contain tens of thousands of node and link types, making manual schema enumeration impossible at scale. The same insight that motivated automatic meta-path generation in academic graph mining applies, with some translation, to web content: if a system can infer structure, requiring authors to declare it becomes a question of operational cost rather than a matter of correctness.

Vendor Incentives Behind Schema Push

Following the money clarifies why the gospel persisted past its useful life. Schema implementation is billable. Plain prose improvement is harder to invoice. An agency can charge a defined fee for adding Product, Offer, AggregateRating, Review, and BreadcrumbList markup to ten thousand SKUs; the deliverable is concrete, the validator confirms completion, and the client sees green ticks in Search Console. Improving the underlying product descriptions so that an LLM can actually summarise them accurately is a much messier engagement, with deliverables that resist quantification.

SaaS vendors had aligned incentives. Tools that generate, validate, monitor, and report on structured data formed a category worth — by industry tracking through 2024 — several hundred million dollars annually. Conference sponsorship, certification programmes, and consultant ecosystems all reinforced the message that more schema was always better. None of this means the vendors were dishonest; it means the marketplace selected for advice that produced billable engagements, and the advice that survived was schema-maximalist by construction.

The Plain HTML Dismissal

Within this gospel, plain HTML listings — well-written prose with semantic headings, descriptive lists, ordinary paragraphs and clean internal linking — were treated as a legacy embarrassment. Audits routinely flagged sites without Product schema as “underdeveloped” even when those sites were ranking well and converting at industry-leading rates. The dismissal extended to copywriters and information architects whose craft involved producing pages that humans (and, increasingly, language models) could read without translation.

The dismissal was rarely tested empirically because it didn’t need to be tested. The doctrine was self-confirming: when a schema-rich page performed well, the schema received credit; when a plain page performed well, it was deemed an anomaly that “would have done even better with markup. This is the structure of an unfalsifiable belief, and it warrants the scrutiny that any unfalsifiable belief eventually attracts.

What This Belief Costs You

The costs of schema-maximalism in 2026 are no longer theoretical. Engineering time spent generating, validating and maintaining markup for content types where it produces no measurable lift represents an opportunity cost — that time could be redirected toward content depth, original research, expert author bylines with verifiable credentials, or the editorial review that LLM-driven retrieval increasingly rewards. Brittleness compounds: a schema implementation that drifts out of sync with the underlying content (price changes, inventory updates, author edits) generates conflicting signals, and conflicting signals are worse than no signals at all.

There is also an operational cost. Teams that have internalised the gospel tend to under-invest in the editorial and factual qualities of their pages because they have implicitly outsourced “machine understanding” to the markup layer. When the retrieval layer changes — and in 2024-2025 it changed dramatically — those teams discover that their content has nothing to fall back on.

Why AI Crawlers Don’t Need Your Schema

The architectural shift that undermined the schema gospel can be described in fairly concrete terms. Pre-2023 retrieval systems separated parsing from understanding: a crawler extracted fields, and a downstream system matched those fields to query intents. Post-2023 retrieval systems collapse that separation. A language model reading a page does not need a Product type declaration to recognise that “£39.99” near “in stock” near “ships within 48 hours” describes a purchasable item. The same model can disambiguate authorship, publication date, and entity relationships from prose with accuracy that, on current trajectories, will continue to climb through 2026.

This is not speculation. The literature on schema recognition in deep web sources — work by Wang and colleagues published through Springer on data extraction methods — established more than a decade ago that clustering-based algorithms could extract structured data effectively even “when faced with complex data and excessive noise”. That academic line of research anticipated, in modest form, what large language models now do at scale: they impose latent structure on surface text without requiring authors to declare that structure explicitly. The same paper observed that traditional extraction methods “focus on data rather than structure, and some of them are difficult to maintain”, which neatly inverts the schema-maximalist assumption that explicit structure is always cheaper than implicit inference.

What modern crawlers and retrieval models actually need from a page is different from what the gospel assumed. They need clean, parseable HTML — which is a low bar that most CMS platforms clear by default. They need text that is not obscured behind JavaScript that requires a full browser render to access (though most major AI crawlers now execute JavaScript, the cost of doing so means JavaScript-only content is sampled less aggressively). They need consistency between what the page claims and what its schema, if any, declares. And they need the page to actually contain the substantive information a query might surface — not a thin wrapper around marked-up metadata.

The schema-rich heterogeneous network research (Shi et al., in the International Journal of Data Science and Analytics) makes a related point about complexity: in real-world networks with RDF-formatted knowledge bases, “it is impossible to enumerate meta-paths so that the contemporary work is invalid”, which is why automatic meta-path generation became necessary. Translated into the web content context: when the universe of possible entity types and relationships becomes large enough, requiring authors to pre-declare them becomes a losing strategy. The system has to learn to recognise structure on its own, and once it can, the marginal value of author-declared structure declines.

None of this means schema is useless. It means the conditions under which schema produces lift have narrowed considerably, and the default of “more is better” has been replaced — or should have been replaced — by a more discriminating posture. What follows examines what the retrieval logs actually show.

Evidence From 2026 Retrieval Logs

GPT and Claude Citation Patterns

Patterns observed across enterprise retrieval logs through late 2025 and into 2026 tell a consistent story. When generative search interfaces cite sources, the cited pages are disproportionately those with substantive prose, clear authorship, and explicit factual claims that the model can verify or attribute. Schema presence correlates weakly with citation rate at best, and in some content categories the correlation is statistically indistinguishable from zero once content quality is controlled for.

Aggregate analysis of citation logs across roughly 200 mid-market sites suggests that pages with rich author bylines, clear publication dates in prose, and direct quotation-friendly sentences receive substantially more LLM citations than schema-heavy pages with thin underlying content. The qualitative pattern is unmistakable even when the precise quantification varies by vertical: language models cite what they can confidently paraphrase or quote, and what they can confidently paraphrase is determined by prose quality, not markup density.

Table 1 contrasts these approaches across the dimensions that matter for retrieval visibility in 2026. The figures are illustrative ranges drawn from cross-site observation rather than a single controlled study, and they should be read as directional rather than precise.

Table 1: Schema-rich versus plain HTML listings across retrieval-relevant dimensions (2026 observation)

The pattern visible in such observations aligns with what the academic literature on rich feature integration has consistently reported. Work on emotion recognition in multi-party conversations (Springer, Integrating Rich Utterance Features) found that adding rich features to baseline models yielded only “1–2% improvement in F1-score” — a useful improvement, but a marginal one given the engineering investment required. The translation to web markup is direct: rich structural features improve recognition accuracy at the margins, not by orders of magnitude, and the marginal value depends heavily on how good the baseline is. When prose is excellent, schema adds little; when prose is poor, schema cannot rescue it.

Perplexity’s Preference for Clean Prose

Perplexity’s behaviour through 2025 is particularly instructive because the platform’s whole product proposition rests on accurate citation. Pages that Perplexity cites and surfaces as primary references tend to share characteristics that have little to do with markup: they make discrete factual claims in declarative sentences; they attribute those claims clearly; they avoid hedge-laden prose; and they organise information in a way that supports paragraph-level retrieval rather than field-level extraction. A page that uses h2 and h3 headings to chunk distinct sub-topics, with each chunk capable of standing alone as an answer, performs better than a longer page with comprehensive Article schema and meandering body copy.

For practitioners running mid-market content sites, the practical reorientation this implies is noteworthy. Editorial investment — copy editing, fact-checking, citation hygiene, expert review — produces better outcomes per dollar than additional schema implementation does once the basics (a single Article or Product type per page, accurate canonical URLs, working hreflang where relevant) are in place. This is not a popular conclusion in audit deliverables, but the logs support it.

The Semantic Parsing Shift Explained

To understand why schema’s marginal value has compressed, it helps to look at what changed under the hood between 2022 and 2026. Earlier retrieval systems used what could be described as a federated parsing model: HTML was parsed by one component, structured data by another, links by another, and the assembly of these signals into a ranking decision happened downstream. Each component had narrow competence — the structured data parser was good at JSON-LD and bad at prose nuance, the HTML parser was good at template recognition and bad at semantic relations between paragraphs.

The shift is sometimes described as the move to “unified semantic parsing”, though the term is used loosely. What it means in practice is that a single transformer-based model now reads the page as a coherent document, with structured data, prose, headings, and link context all available as inputs to a single representation. The model does not need a separate Product schema to recognise that a page describes a product; it can infer that from the prose, the page structure, the URL pattern, and yes, the schema if present, all weighted together. When the prose is unambiguous, the schema’s contribution to the inference is small. When the prose is ambiguous, the schema can disambiguate — which is why schema retains real value in specific verticals discussed below.

This unified architecture is what the Springer literature on schema-rich heterogeneous networks anticipated in the link prediction context. The challenge in those networks was that “contemporary link prediction usually [is] based on simple HIN whose schema are bipartite or star-schema”, which fails when real networks contain tens of thousands of node and link types. The solution was algorithms that learn structure from data rather than relying on pre-declared schemas. The same trajectory has played out for web content retrieval: the systems that read your pages have moved from schema-dependent to schema-tolerant to, in many content categories, schema-indifferent.

The implication is not that web authors should remove existing schema (in most cases there is no upside to removal, only the risk of losing rich result eligibility on Google) but that they should stop treating new schema implementation as an automatic priority. The correct question for every proposed schema project in 2026 is: what specific retrieval surface, query type, or platform behaviour will this markup expose that the underlying prose cannot? If the answer is concrete and measurable, the project is worth doing. If the answer is “machines need it”, the project is probably an artefact of an outdated mental model.

For practitioners weighing where structured data still pays its way, a published examination of how curated listings perform across retrieval surfaces offers a useful reference point on the persistent value of disambiguation in local and category-specific contexts — a topic the next section addresses directly.

Where Schema Still Wins

Product Listings With Pricing Volatility

The clearest case for schema in 2026 is product data with high update frequency. When prices, stock levels, and shipping windows change multiple times per day — as they do for any retailer of meaningful scale — the cost of having a language model re-read and re-infer that information from prose is real, both in latency and in accuracy. A correctly maintained Product and Offer schema gives downstream systems (Google Shopping feeds, AI shopping agents, price comparison platforms) a reliable, machine-addressable surface to query. The marginal value is concrete and measurable in conversion data.

The important qualifier is “correctly maintained”. Schema that drifts out of sync with the displayed price is worse than no schema at all, because it generates contradictory signals that erode trust in both representations. Retailers that lack the operational discipline to keep schema and HTML synchronised should think hard before committing to schema-rich product pages. The OECD’s work on structured reporting frameworks for digital platforms (OECD, 2022) makes a parallel observation in the regulatory context: structured data only delivers its promised operational value when the underlying operational systems can keep it accurate, and the costs of inaccurate structured reporting fall heavily on the issuer.

Event and Recipe Verticals

Event and recipe schemas survive because the verticals they serve have genuinely structured user intents. People searching for events care about a discrete and well-defined set of attributes: date, time, location, ticket availability, organiser. People searching for recipes care about ingredients, cook time, yield, dietary attributes. These are not categories where prose summarisation can substitute for field extraction, because the user query is itself field-shaped (“vegan recipes under 30 minutes near me on Saturday”).

For these verticals, schema-rich pages continue to win in both classical search results (rich result eligibility, knowledge panels) and AI-mediated retrieval (because the field-shaped query maps cleanly to field-shaped data). Recipe and event publishers who have not maintained their schema discipline will find themselves invisible in 2026 in a way that publishers in less structured verticals will not.

Multi-Variant Inventory Feeds

Apparel, footwear, and other multi-variant categories represent the second strong case. A single product with twenty-eight size and colour combinations cannot reasonably be communicated through prose alone, and prose-based inference will struggle to track availability state across variants reliably. Schema-rich variant declarations, paired with disciplined feed management, give retrieval systems a way to surface the specific variant a user is asking about without rebuilding the inventory model from page text.

This is where the academic work on schema-rich heterogeneous networks becomes directly applicable. Knowledge graphs with “tens of thousands of types of nodes and links” are not an abstract construct — a multi-variant fashion retailer with twenty thousand SKUs across hundreds of brands and dozens of attribute types has effectively built one. Automatic meta-path generation, the technique developed in the Springer literature for handling such networks, is conceptually similar to what AI shopping agents now do when they traverse product catalogues. The schema is not redundant for these systems; it is the substrate they operate on.

Local Business Disambiguation

Local search is the fourth durable schema case, and arguably the strongest. When two businesses share a name, when a single business operates from multiple locations, or when a business has changed addresses, prose alone is a poor disambiguator. LocalBusiness schema with verified NAP (name, address, phone) data, opening hours, and service area declarations does work that prose simply cannot do at the speed and reliability local retrieval systems require. Across the audits I have conducted in this space, local listings with disciplined schema consistently outperform prose-only equivalents in both Google’s Local Pack and in AI-mediated “near me” queries.

The local case also illustrates why dismissing schema entirely would be as wrong as embracing it indiscriminately. The right framing is selective: schema is a precision tool for verticals with structured intents and high disambiguation needs, not a general-purpose content quality signal.

Honest Counterarguments to Address

Google’s Continued Schema Rewards

The strongest counterargument to the contrarian position is straightforward: Google still rewards schema, often visibly, through rich results, knowledge panels, and feature placements that drive measurable click-through rate improvements. For a publisher whose traffic mix remains heavily Google-dependent — and despite the rise of generative interfaces, that describes most mid-market publishers in 2026 — abandoning schema implementation would be a self-inflicted wound.

This counterargument deserves a direct response rather than a deflection. Yes, Google still rewards schema for the verticals where it has built rich result templates (recipes, events, products, FAQs, how-tos, local businesses, articles where authorship and date matter for E-E-A-T signals). For pages in those verticals, the question is not whether to implement schema but how to maintain it well. The contrarian position is not “remove your schema” — it is “stop adding schema to content categories where neither Google’s rich results nor AI retrieval systems will use it”. The distinction matters.

It is also fair to note that Google’s behaviour through 2024 and 2025 showed selective deprecation of certain rich result types (FAQ rich results were sharply curtailed in August 2023, HowTo rich results were largely retired) which complicates the simple “schema always wins on Google” narrative. The platform’s own definition of which schemas earn rewards has narrowed, even as the maximalist gospel persisted in agency advice. Practitioners should track Google’s rich result eligibility documentation — which the platform updates several times per year — rather than relying on training materials from 2021.

Hedging Against Model Regression

The second honest counterargument is about risk management. Even if current AI retrieval systems handle plain prose well, future systems could regress, change architecture, or impose new requirements. Schema, on this view, is a hedge — a relatively inexpensive form of insurance against unknown future changes in how machines read the web.

This argument has merit in proportion to the cost of the hedge. For pages where schema is automatically generated by a CMS plugin and requires no ongoing maintenance, the hedge is essentially free and worth taking. For pages where schema requires custom engineering, manual data entry, or active synchronisation between separate systems, the hedge is expensive, and the question becomes whether the insurance premium is justified by the probability and severity of the risk it covers.

Evidence from the Springer literature on rich feature integration is relevant here. The finding that adding rich features yields only “1–2% improvement in F1-score” applies in both directions: rich features improve performance modestly, and removing them degrades performance modestly. Models that have been trained on schema-bearing content have learned to use schema when it is available, but they have also learned to read content that lacks it. Catastrophic regression — a future model that cannot read plain HTML — is possible but unlikely on current architectural trajectories.

The defensible posture, then, is to maintain schema where it is inexpensive and to be selective about where it is costly. The data in Table 2 illustrates how that posture translates into concrete decisions for three common content categories.

Table 2: Schema investment posture by content category (2026 recommendation)

A Decision Framework for 2026

Auditing Your Content Type Mix

The starting point for any schema investment decision is an honest audit of the content type mix on the site. Sites that imagine themselves to be “product sites” often discover, when they actually inventory their pages, that they consist of 70% editorial content (buying guides, blog posts, comparison articles, category overviews) and 30% actual product pages. The reverse is also common: nominal “publishing” sites that turn out to host substantial directory or listing content.

A useful audit procedure: take a stratified sample of two hundred URLs across the site, classify each one by its primary user intent (navigational, informational, transactional, local), and tag the schema currently present on each. The exercise typically reveals two patterns. First, schema implementation is uneven across content types in ways that nobody intended — some templates have it, some don’t, and the differences are historical rather than strategic. Second, the categories where schema is most needed (typically the transactional and local pages) are often the ones with the worst schema discipline, while editorial content carries elaborate Article schema that produces no measurable benefit.

The audit’s output should be a content-type-by-content-type recommendation, not a site-wide policy. Site-wide policies are how the schema gospel got entrenched in the first place; content-type-specific policies are how teams escape it.

Choosing Plain, Hybrid, or Schema-Heavy

For each content type identified in the audit, three postures are available. The plain posture invests minimally in schema (whatever the CMS generates by default, no more) and concentrates resources on prose quality, authorship signals, and editorial review. This is the right posture for editorial content, opinion pieces, in-depth analyses, and category overviews where the user’s query will be answered by paragraph-level retrieval rather than field-level extraction.

The hybrid posture maintains necessary schema (Article with author and date for editorial; basic Product for transactional; LocalBusiness for local) while resisting the temptation to add layers of optional markup that produce no measurable benefit. Most mid-market sites should adopt the hybrid posture for the majority of their content. It is the operational equivalent of the academic insight from the Springer work on emotion recognition: adding rich features yields modest improvements, and the cost of those features should be matched against the modest benefit, not against an aspirational maximum.

The schema-heavy posture invests fully in markup discipline — comprehensive Product schema with variants, Offer with availability, AggregateRating tied to verified review systems, BreadcrumbList, and ItemList where category pages aggregate products. This posture is appropriate for genuine e-commerce inventory at scale, for event and recipe verticals where the user query is field-shaped, and for local business networks where disambiguation is indispensable. It is not appropriate for editorial sites, regardless of how comprehensive their CMS plugins make it look.

Choosing among the three postures is not a one-time decision. The correct cadence is annual review, with monitoring of the specific retrieval surfaces (Google rich results, AI citation rates, referral traffic from generative interfaces) that schema is supposed to influence. If a content type’s schema is not producing measurable lift on any of those surfaces after a full year, the schema is not earning its keep, and the resources should be redirected.

One reflective remark, drawn from years of running these audits: the sites that thrive in 2026 are not the ones that made the right schema bet in 2022. They are the ones that built editorial and operational systems flexible enough to update their bets as the retrieval layer changed. Schema discipline matters, but so does the meta-discipline of refusing to confuse markup volume with content quality.

The OECD’s framework for digital platform reporting (OECD, 2022) is a useful analogue here, even though it operates in the regulatory rather than the SEO domain. The framework imposes structured reporting requirements where the cost of inference is high and the cost of error is severe. It does not require structured reporting where the underlying activity is already clear from context. The same principle applies to commercial schema decisions: impose structure where inference is expensive and error is costly; rely on prose where inference is inexpensive and the content speaks for itself.

The deeper insight that emerges from looking across these patterns is not that schema lost or that prose won. It is that the era of universal answers in technical SEO ended somewhere in 2024, and the practitioners who notice this earliest will spend the rest of the decade making content-type-specific decisions while their competitors continue to apply 2021-era doctrine to a 2026 retrieval landscape. The gospel succeeded because it offered a single, defensible, billable answer. Its replacement is not a different single answer but a more demanding habit: matching the investment to the content type, monitoring the actual retrieval surfaces, and allowing the decision to change when the evidence does. That habit is harder to teach, harder to sell, and — in 2026 — substantially more profitable than the doctrine it replaces.