You've done the SEO work. Your page ranks on page one. But when someone asks ChatGPT the same question your page answers perfectly — your content isn't in the response.
This isn't a ranking problem. It's a citation problem. The cause is structural.
How LLMs source answers — this is not a search index
When Google crawls your page, it evaluates keywords, links, and authority signals against a ranking algorithm. When ChatGPT answers a question, it draws on patterns in its training data, weighted toward sources it encountered frequently and with consistent entity recognition.
You can't optimise your way into an LLM the way you optimise for a crawler. What you can do is make your content easier to cite — which means structurally legible to systems that process language, not just keywords.
This is what how generative engines surface content comes down to at an implementation level. The framework is called GEO — Generative Engine Optimization — and the core mechanic is less about domain authority and more about extractability.
Three gaps that make well-ranked content invisible to AI
Most pages that rank well but never get cited share the same three problems.
The first is entity ambiguity. If your organisation is called "IIT Bombay" on your website, "Indian Institute of Technology Bombay" in press coverage, and "IIT-B" in student forums, AI systems struggle to resolve these into a single entity. They build an internal graph of who's who — inconsistent naming means a weak node in that graph, which means lower citation probability.
The second is the absence of extractable direct answers. LLMs prefer content they can pull a clean sentence from. "There are several factors to consider when choosing a design exam" is hard to cite. "UCEED tests visual reasoning and design awareness; NID DAT tests creative instinct across two stages; NIFT tests fashion and textile sensibility" gives the model something it can actually use in a response.
The third is missing schema. FAQPage, Article, BreadcrumbList, EducationalOrganization — these aren't just Google signals. They declare the relationships between entities in machine-readable terms. Without schema, the model infers structure from prose. With schema, the structure is explicit. The difference in citation probability is not small.
The JSON-LD that addresses all three
Here's a minimal schema stack for any educational content page. Start with the Article block:
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "Your Page Title",
"author": {
"@type": "Person",
"name": "Jaydip Parikh",
"url": "https://tej9.com",
"sameAs": [
"https://www.linkedin.com/in/jaydipparikh",
"https://twitter.com/jaydipparikh"
]
},
"publisher": {
"@type": "Organization",
"name": "EDU SolPro",
"url": "https://edusolpro.com",
"sameAs": ["https://www.linkedin.com/company/edusolpro"]
},
"datePublished": "2026-05-10",
"dateModified": "2026-05-10"
}
The sameAs array is doing most of the work. It tells the model: these identifiers all refer to the same entity. That's what builds consistent recognition across training data.
Add a FAQPage block on any page that answers specific questions:
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What does UCEED test?",
"acceptedAnswer": {
"@type": "Answer",
"text": "UCEED tests visual reasoning, design awareness, and environmental and social awareness. It is conducted by IIT Bombay for admission to B.Des programmes at IITs and IIITDM Jabalpur."
}
},
{
"@type": "Question",
"name": "How is NID DAT different from UCEED?",
"acceptedAnswer": {
"@type": "Answer",
"text": "NID DAT tests creative instinct and design sensibility across two stages. Unlike UCEED, which is objective and structured, NID DAT includes a studio test that assesses spontaneous creative thinking."
}
}
]
}
The answer text matters here. Specific, declarative, dateable. Not "various differences exist." Specificity is what makes content citable rather than just readable.
Why comparison pages get cited more than any other format
Comparison pages are among the most-cited content types in AI responses. The structural reason: a well-built comparison table declares relationships explicitly, in a format closer to structured data than prose.
"UCEED: conducted by IIT Bombay, B.Des at 7 IITs. NID DAT: conducted by NID Ahmedabad, B.Des at 23 NID campuses." That's machine-readable clarity in plain text. Add BreadcrumbList schema and specific dateable facts throughout, and you have content an LLM can anchor a confident answer to.
A design entrance exam comparison that covers eligibility, exam pattern, difficulty, and career outcomes in structured table format — with schema on top — demonstrates this in practice. The model may have encountered both a structured and an unstructured version of the same content. It will cite the structured one because the extraction is clean.
If you're building content-heavy sites, format choices you've been treating as editorial are now infrastructure decisions. Table vs. prose. Schema vs. none. These have AI visibility consequences that didn't exist three years ago.
Where the gap is widest: education
The education sector has a specific version of this problem. Prospective students now ask AI systems the exact questions universities spent years optimising for on Google.
"Which engineering college in Gujarat should I apply to?" "What is the NAAC grade of this institution?" "How does this university's placement record compare?"
The data exists. Most universities have NIRF reports, NAAC assessments, placement statistics. They haven't structured that data in a way AI can extract and cite. This is a markup problem, not a content problem.
There's an interesting parallel to how university rankings methodology works in traditional systems: academic reputation (40% of QS weighting) and citation impact (20%) are the dominant signals. In AI citation, entity reputation and content specificity play analogous roles. Institutions well-represented in authoritative, consistently structured sources rank higher in both systems.
The full strategy for Indian institutions — covering entity clarity, FAQ structure, schema implementation, and the first-mover window that still exists right now — is in GEO for universities.
GEO is a layer, not a rewrite
Most developers treat this as a content project when they first hear about GEO. It isn't. The content already exists on most sites. What's missing is the structured layer on top of it.
Implementation is faster than it looks:
- Audit existing pages for entity consistency — an afternoon with a spreadsheet- Add Article + Author + Publisher schema to all long-form content — one template, deployed site-wide- Add FAQPage schema to pages answering specific questions — extract the most common query for each page, structure it as a Q&A block, add JSON-LD- Add BreadcrumbList to all category and comparison pages — usually one component update- Create an llms.txt at root — a plain text declaration of what the site covers, who it's for, and what its authoritative pages areNone of this requires new content. It requires structural decisions about how existing content is declared to systems that read it differently than humans do. ## Five things to add to any educational content page today
- Article schema with author sameAs pointing to LinkedIn and Wikipedia if the author has a profile- FAQPage schema on any page answering a specific question — one Q&A block is enough to start- EducationalOrganization schema on institution pages with name, url, sameAs, and address consistent across every page- BreadcrumbList on all comparison and category pages- An llms.txt file at root — three to five sentences on what the site covers and who it's authoritative forThese changes shift a page from content that exists to content that can be cited. The model likely encountered your content in training already. These changes tell it what to do with it.
Top comments (0)