Schema markup hasn't aged well as AI citation advice.

If you've read more than three articles on getting cited by ChatGPT, you've read "add schema markup" as one of the top recommendations. It's the most-repeated piece of GEO advice in the entire category. It's also the piece of advice that's held up worst as actual AI citation patterns have become observable. This is an opinion piece — laying out why the schema-first framing has aged badly, what schema actually does that's useful, and what the same effort would buy you elsewhere.

To be clear up front: schema markup is not a scam, not a waste of time, and not actively harmful. It's a real tool with real effects. The issue is that for three years it's been positioned as the primary lever for AI citation, and the data we now have says it isn't even in the top three.

What the 2023-2024 advice was

If you read GEO advice from 2023 or early 2024, it sounded like this: AI engines parse structured data to understand your site's entities. Adding Organization schema tells the AI who you are. FAQPage schema lets the AI extract questions and answers directly. Product schema tells the AI about your offerings. Therefore, the more schema you have, the more visible you are to AI engines.

The theory was plausible. Structured data is machine-readable. Language models can parse JSON-LD. Search engines have spent a decade promoting schema. Why wouldn't AI engines lean on it?

The answer is that the way AI engines actually retrieve and cite content doesn't go through your schema at all in 95% of cases. The retrieval pipeline pulls text chunks from a search index. The cross-encoder reranks those chunks based on relevance to the user's question. The language model writes the answer from the resulting chunks. Schema lives in a different layer that's mostly used by traditional search ranking and rich-result eligibility — it isn't fed to the model as primary citation material.

What we can actually observe

When we audit sites that are getting cited by ChatGPT, Claude, and Perplexity, the citations consistently come from text passages in the page body — not from schema. The model quotes paragraphs, lists, headings, and table contents. It never quotes JSON-LD. You can verify this yourself: ask ChatGPT for a recommendation in any category, click through to the cited sources, and look at what the model said about each one. The phrasing maps to body text, not schema.

In our audits across 2026, we've watched sites with zero schema get cited more frequently than competitors with comprehensive schema. We've also watched sites with extensive schema fail to be cited because their body text was structured in ways the cross-encoder couldn't chunk usefully. Pattern: schema doesn't compensate for body content that doesn't answer questions directly, and content that does answer questions directly gets cited fine without schema.

This isn't unique to our audits. Otterly AI's 2026 YouTube citation study found that schema correlation with citation rate was statistically weak compared to content structure (TLDR-first paragraphs, named entities, specific claims). The Princeton GEO research from 2024 — still the most-cited empirical work in the category — found that adding statistics, quotations, and inline citations were the top-performing interventions, lifting AI visibility by up to 40%. Schema markup wasn't among the top-tier measured effects.

What schema actually does for AI citation

Schema has two real, measurable effects on AI citation, and neither is what the 2023 advice claimed.

Effect one: schema makes your page easier to chunk correctly. The RAG pipeline retrieves your page and then breaks it into 200-800 token chunks that get scored for relevance. FAQPage schema in particular gives the chunker explicit boundaries — each Q&A pair becomes its own logical chunk. Without schema, the chunker uses heuristics on your HTML structure, and those heuristics often pull wrong boundaries (mid-paragraph, mid-list, across sections). Schema doesn't directly affect what's cited, but it does affect whether your chunks contain the answer-bearing material vs. your navigation.

Effect two: schema improves Bing's indexing of your page, which feeds AI retrieval. Bing uses schema as a quality signal for its own ranking, and Bing's index is what ChatGPT pulls from for browsing-mode citations (~87% citation overlap). Better Bing ranking → more likely to be pulled into the retrieval candidate set → more likely to be cited. So schema helps citation, but indirectly, through a path that runs through traditional search ranking rather than through the language model.

Notice what these two effects have in common: neither is "the AI reads your schema and uses it as fact source." That framing is what 95% of GEO articles still claim, and that framing is wrong.

What the same effort buys you elsewhere

Schema implementation takes real time. A full pass — Organization, WebSite, BreadcrumbList, FAQPage, Article, Product as appropriate — on a 50-page site is a couple of days of work, more if you're doing it manually rather than via a CMS plugin. That same time invested elsewhere has more leverage:

Restructure your top 10 pages to answer first. The single biggest factor in AI citation is whether the first 300 words of your page contain a citable answer to the obvious user questions. Most pages bury their answer behind 800-1500 words of setup, framing, and brand-voice introduction. The cross-encoder selects chunks based on direct answer-relevance — if your answer isn't in the early chunks, you're not cited. Restructuring 10 pages to lead with the answer takes a similar amount of time as a comprehensive schema pass, and the citation impact is dramatically larger.

Verify your site in Bing Webmaster Tools and submit a sitemap. Bing's AI Performance dashboard (launched February 2026) is the only first-party citation tracking that exists. It shows you exactly which sub-queries triggered your site as a candidate and which pages got cited. Most sites we audit have 30-60% of their pages missing from Bing's index entirely. Fix that and you've moved more retrieval-stage candidate slots than any schema work can deliver.

Build entity consistency across your pages. AI engines build a mental model of "what your site is" by reading multiple pages and looking for consistent self-identification, consistent categories, consistent named entities. A site that calls itself a "CRM" on one page and a "sales platform" on another splits its entity weight. Picking one identity and reinforcing it across every page is more impactful than any single piece of structured data. This work is also impossible to automate — your competitors won't be doing it.

Update content dates on still-valid pages, refresh stale ones. Freshness is a heavier signal in RAG retrieval than in traditional search. Content older than six months drops in citation rate sharply. Going through your top 20 pages and either updating the last-modified date (where the content is still valid) or substantively refreshing the body (where it isn't) is high-leverage maintenance work.

The case for still doing schema

None of this is an argument against schema. It's an argument against schema being the headline recommendation. The actual case for schema in 2026 looks like this:

• Yes, add Organization and WebSite schema. They take 10 minutes to add via JSON-LD, they help Bing's indexing, and they're zero-maintenance. Just do it.
• Yes, add BreadcrumbList schema to deeper pages. Helps both traditional search and the chunker's structural understanding. Most sites already have a breadcrumb component — wiring it to schema is a one-time effort.
• Add FAQPage schema selectively where you actually have FAQs. Don't manufacture FAQ content just to add the schema. The schema only helps if the underlying questions and answers exist in the body and address real user intents.
• Skip the rest unless you have a specific reason. Product schema for ecommerce, Article schema for news, Course schema for education — these have specific rich-result benefits. If you're not going for those rich results, the schema isn't doing much for AI citation that wouldn't be done by good body content.

The problem with the "add lots of schema" framing isn't that schema is bad. It's that it positions schema as the leverage point when it isn't, and steers people away from the actual leverage points: body structure, Bing indexing, entity consistency, freshness.

Why this matters for the GEO category

A lot of GEO advice still leans on the 2023 mental model of AI engines as "machines that read structured data and respect entities." That model was reasonable when nobody had observed actual citation patterns at scale. It became less reasonable through 2024 as Bing's AI Performance dashboard, OtterlyAI's tracking, and the Princeton research started producing empirical data. By 2026, the model is mostly wrong, and continuing to give schema-first advice misallocates attention and effort.

The category needs to catch up. Recommendations should be ordered by observed citation impact: body content structure first, retrieval pipeline health (Bing indexing, freshness) second, entity consistency third, schema fourth. Most GEO content still ranks them in roughly the reverse order. That's the calibration error worth fixing.

What we recommend at Reffed

Our audit checks all of this — body structure for answer-first patterns, Bing index status, schema implementation, entity consistency across pages, freshness signals. We surface findings in order of likely citation impact, not in order of "what's easiest to recommend." Schema usually shows up as a yellow flag rather than a red one, because for most sites it's not the top issue.

If you want to see how this maps to your own site, run a free audit — you'll get a per-engine citation score across ChatGPT, Claude, Perplexity, Google AI Overviews, Gemini, and Microsoft Copilot, plus the structural and indexing findings ordered by expected impact. Most sites discover the priorities differ from what they've been told.

If you want to go deeper on the why-and-how, the free Foundations course walks through the actual mental model — what AI engines do at each stage, how citation decisions get made, and which levers move which outcomes. The paid Quickstart ($147 founding) covers the full execution playbook.