“Add schema markup” has become one of the most repeated pieces of advice for getting cited by ChatGPT, Google’s AI Overviews and other answer engines — it shows up in nearly every AEO and GEO checklist published in the past year. The logic sounds reasonable enough: structured data literally structures your content for machines, so machines should make more use of it. A large controlled study published in 2026 tested that claim directly against nearly 1,900 real pages, and the result is worth knowing before any Malaysian business spends development time chasing it.
What the schema-for-AI-citations advice actually assumes
The reasoning behind most AEO and GEO guides runs roughly like this: AI Overviews, AI Mode and chat-based answer engines all need to extract facts from a page quickly and confidently, and JSON-LD schema — Article, FAQ, Product, Organization, HowTo — exists precisely to spell those facts out in a format a machine can parse without guessing. If AI systems reward pages that are easier to parse, schema should, in theory, be one of the more reliable AEO wins available: cheap to implement, technically unambiguous, and something a developer can tick off in an afternoon. It’s also one of the few AEO recommendations that comes with a specific, concrete instruction rather than vague advice to “optimise for AI,” which is probably why it has spread as widely as it has.
What a large-scale controlled test actually found
Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, and compared their AI citation activity against roughly 4,000 matched pages that added nothing, using four separate statistical methods — including difference-in-differences analysis — to isolate schema’s effect from platform-wide trends. The schema types were pooled together (Article, FAQ, Product, HowTo and Organization), and only genuine HTML-embedded JSON-LD was counted, not JavaScript-injected markup.
| Platform | Citation change after adding schema |
|---|---|
| Google AI Overviews | −4.6% (statistically significant decline) |
| Google AI Mode | +2.4% (indistinguishable from normal variation) |
| ChatGPT | +2.2% (indistinguishable from normal variation) |
The AI Overviews decline worked out to roughly 12 fewer citations a day, on average, for pages that added schema. Neither AI Mode nor ChatGPT showed movement outside normal noise. Across all three platforms, adding schema produced no reliable citation uplift — and for AI Overviews specifically, it correlated with a small decline.
Then why does the advice persist?
The same study found that 53% of pages already being cited by AI systems carry some form of schema markup, and that statistic is almost certainly where the “add schema” advice originates. But it describes correlation, not cause. Sites that bother implementing schema tend to be the same sites investing in technical SEO, original content and link building at the same time — the citation is being earned by the content and authority behind it, with the markup along for the ride rather than driving anything.
Where schema markup still earns its keep
None of this makes schema markup pointless — it just belongs back in its original job. It still drives the rich results that show star ratings, FAQ dropdowns and product pricing directly in classic Google search listings, which remain useful regardless of what AI systems do with them. Organization and LocalBusiness markup still gives search engines and directories a clean, unambiguous statement of who a business is, which supports the kind of entity clarity that does appear to matter for AI systems (more on that below). It simply isn’t the lever that gets a page chosen as a citation. Keep using it for SEO. Stop budgeting it as an AEO strategy.
The llms.txt file is the same story, told a second time
A near-identical pattern has played out with llms.txt, the plain-text file some sites publish to summarise their content for AI crawlers. Asked directly about it on Reddit in June 2026, Google’s John Mueller was blunt:
“I don’t think anyone knows – it’s purely speculative for now (the file has existed for years, yet none of the AI systems use it — what does it mean?).”
John Mueller, Google, June 2026
His point was direct: no major AI system currently reads or requires an llms.txt file, despite the format having existed for years. He pointed instead to WebMCP, a Google-backed standard that lets AI agents actually perform tasks on a site rather than parse a summary of it, as the more concrete direction worth watching. His practical advice for site owners was simpler still — make sure AI crawlers aren’t blocked from reaching a site, rather than spending time on a file no system currently reads.
What actually seems to move the needle
Strip out the two techniques that don’t hold up under testing, and a clearer pattern emerges from the available research:
- Content written to directly answer one specific question early, rather than building up to it over several paragraphs
- Ranking well organically in the first place, since AI inclusion still correlates strongly with organic search visibility
- Being described consistently — name, service, location — across a business’s own site and the other sources answer engines already reference
- Being mentioned by independent sources at all: press coverage, directories, comparison pages and review sites
Most answer engines still lean heavily on what other sources say about a business, not on what that business’s own markup claims about itself — which is a harder thing to engineer than adding a schema tag, and probably why it gets far less attention.
What this means for a Malaysian business’s SEO priorities
- Check that AI crawlers (Google-Extended, GPTBot, PerplexityBot and similar) aren’t accidentally blocked in robots.txt — that quietly forecloses citation regardless of anything else done well.
- Rewrite key pages so the direct answer to a reader’s actual question appears in the first few sentences, not after several paragraphs of scene-setting.
- Keep using schema markup for the rich results and entity clarity it’s genuinely built for, without expecting it to double as an AEO strategy on its own.
- Redirect the time budgeted for schema or llms.txt tinkering toward getting mentioned elsewhere — press, partner sites, comparison content, directories — since third-party citation still appears to carry more weight than anything published on-page.
- Spot-check a handful of real customer queries in ChatGPT, Google AI Mode and Perplexity periodically to see whether the business is actually named, rather than assuming structured data is quietly working in the background.
Two questions worth asking before changing anything
Should existing schema markup be removed?
No. It still supports rich results and entity clarity in traditional search, both of which remain worth having regardless of what the evidence shows about AI citations specifically.
Does this mean AEO and GEO don’t matter?
No — it means the highest-leverage AEO work is content structure, organic ranking and independent mentions elsewhere, not a specific markup tag added to a page template.
This is really a wider pattern in AEO and GEO advice at the moment: the tactics getting the most airtime tend to be the easiest ones to implement, not the ones with evidence behind them. The slower work — content built around real expertise, a technical foundation that doesn’t block AI systems out, and a reputation built across sources a business doesn’t directly control — has more evidence behind it, even though it doesn’t fit neatly on a checklist. That’s also roughly how MRVS treats SEO, AEO and GEO as a connected discipline rather than a set of tags to add: search visibility, content quality and where a brand gets mentioned across the rest of the web all move together, in much the same way that AI Overviews have already changed what counts as a good SEO outcome for Malaysian businesses.
