You spent weeks getting your JSON-LD perfect. Product schema, FAQ schema, review schema — all validated and green-lit in Google's Rich Results Test. Then an Ahrefs test found that five major AI systems ignored that structured data entirely and pulled their answers from visible HTML instead (source).
If you're deciding where to spend limited dev and content hours this quarter, that finding changes the calculation. Here's how to figure out what schema markup still earns you, and what to check instead if AI citation is the actual goal.
What the Ahrefs Test Actually Found
In the test described in Ahrefs' AI search optimization guide, five major AI systems ignored JSON-LD, hidden Microdata, and hidden RDFa on the pages checked. Instead, they relied on visible HTML — the text a human reader would actually see rendered in a browser (source). If a fact like a price or a product spec existed only inside a schema block or inside an image, the AI systems in that test didn't appear to pick it up.
That's a narrower finding than "schema is dead," and it's worth being precise about what it does and doesn't tell you. It covered a small number of AI systems, checked at one point in time, on the specific pages Ahrefs examined. It doesn't prove every AI system behaves this way, and it doesn't guarantee this behavior holds next quarter as these systems change how they retrieve and process pages. Treat it as a strong signal to verify on your own site, not a permanent law of AI retrieval.
Does This Mean Schema Markup Is Now Useless?
No, and treating it that way would cost you real, measurable wins elsewhere. Schema markup still does two jobs well, and neither depends on AI citation.
First, structured data still earns you classic rich results in traditional search: star ratings, FAQ accordions, breadcrumb trails, product pricing snippets. Google's rich-result systems parse that data through its own established pipeline, separate from how AI answer engines chunk and retrieve content.
Second, schema still helps with entity disambiguation. It tells search engines unambiguously that "Apple" on your page refers to the company, not the fruit, or that a listed price applies to a specific SKU. That clarity supports indexing and knowledge graph connections regardless of whether an AI system ever quotes your page directly.
What schema markup does not do, based on this finding, is guarantee that an AI system will surface a fact that exists only inside that markup. If your pricing lives in a JSON-LD Product block but nowhere in the paragraph text a visitor reads, you have a rich-result asset — not necessarily an AI-citation asset. The fix isn't ripping out your schema. It's shifting the priority: stop treating schema maintenance as your AI visibility strategy, and start auditing whether your key facts are duplicated in visible text.
The Audit: Checking Whether Your Facts Are Actually Visible
This is the practical work that replaces "add more schema" as your next priority. Run this checklist on any page where pricing, specs, or stats matter to your business. For more on this, see more on ai search data sources: what matters by business type.
-
Pick your highest-value facts first. List the specific numbers a customer or an AI system would need to answer a real question about the page: price, dimensions, a percentage from a study, a release date, a comparison metric.
-
Open the page and inspect the rendered DOM, not just the page source. In Chrome, right-click and choose "Inspect" to open the Elements panel, which shows the DOM after JavaScript has run. This gets you closer to what a crawler or AI retrieval system may encounter than raw "View Page Source" output, though behavior still varies by system and by how each one renders JavaScript.
-
Search the rendered HTML for each fact as literal text. Use Ctrl+F within the Elements panel, or copy the rendered HTML into a text editor and search it there. Flag any fact that appears only inside a
<script type="application/ld+json">block, or as an attribute inside a Microdata or RDFa<div>, with no matching text in a<p>,<span>,<li>, or heading a reader would actually see. -
Check every fact that lives inside an image, chart, or infographic. If your only mention of a statistic is a number baked into a PNG, no text-based system — human or AI — can read it. Watch especially for pricing tables rendered as images, a common enterprise site pattern.
-
Confirm the fact reads clearly out of context. Pull the sentence containing the fact and read it alone, as if it were the only thing quoted from your page. If it needs three prior sentences to make sense, a chunking system risks citing it incorrectly, or skipping it entirely.
-
Re-run this check after major redesigns or CMS migrations. Redesigns often move facts that were visible text into schema-only fields, or into a component library that renders data as an image or SVG, and nobody notices until the fact stops showing up in search or AI results.
A Worked Example Walkthrough
Here's how steps 2 and 3 play out on a hypothetical pricing page, for illustration only. See full coverage of discover's 'dive deeper' ai test: a publisher checklist for additional background.
Say a SaaS pricing page shows a plan card with "$49/month" in large text inside a styled component. You inspect the element, and the rendered HTML shows <div class="price">$49</div> next to <div class="period">/month</div>. That's a pass: the fact is visible text, findable by a text search, and readable without adjacent context.
Now imagine a second page — a comparison table where the pricing header cell is generated as an SVG graphic for cross-browser design consistency. You inspect that element and find <svg> markup with path data, no readable "$49" string anywhere in the DOM. A Product schema block elsewhere in the page's <head> does list the price correctly.
Also read: is advertising on chatgpt ads worth it for retailers? — background
That page fails the check. The price satisfies search engines parsing structured data and qualifies for rich results, but a retrieval system reading the page like a human would may never find it as text. This is the exact gap the Ahrefs finding points to (source). A schema validator won't catch it either, since a validator only confirms your JSON-LD is well-formed — it never checks whether the same fact also exists as visible text.
What to Prioritize Instead
Once the audit above shows you where facts are hidden, the fix is usually additive, not a rebuild. Add a plain-text sentence stating the fact near the graphic or schema block; you don't need to remove the image or the structured data.
Beyond patching hidden facts, prioritize making your visible text itself easy to extract. Ahrefs cites research from Kevin Indig showing that 44.2% of AI citations in that research came from the first 30% of a page, and that cited passages tended to use more named entities and more definitive language than typical prose (source). Practically, that means leading each section with the direct answer before the supporting explanation — a pattern often called "bottom line up front" — and writing self-contained sentences that still make sense when quoted in isolation.
Ahrefs also points to its own 2024 survey on the cost of SEO as a working example: the key stats shown in charts were also restated as text in the article body, and the piece has kept earning citations even after competitors published newer research (source). That's one company's case study, not a controlled experiment, but the mechanism it illustrates — duplicating chart data as text — lines up directly with the audit above.
What to Expect Next
Run the six-step audit on your five or ten most commercially important pages this week, not your entire site. Fix the hidden-fact gaps you find, leave your schema in place for the rich-result and entity value it still provides, and recheck after any redesign.
AI retrieval behavior will keep shifting, so treat this as a recurring spot-check rather than a one-time project. Neither schema nor visible-text fixes guarantee a citation. What they do is remove an unnecessary barrier between your facts and the systems trying to read them.
