The Discovery Layer Is Fracturing
For the past two and a half decades, SEO was the answer to one question: how do people find what you built?
You wrote content. Google indexed it. Users searched for a problem. Your page ranked. Traffic flowed. The model was simple, predictable, and ruthless. Businesses lived or died by their ability to own search queries.
That assumption is breaking. And the fault line isn’t in Google’s algorithm. It’s in how LLMs like ChatGPT, Claude, and Perplexity are becoming the new discovery interface.
When someone asks ChatGPT “how do I hire a marketing agency,” they’re not clicking through to a search results page. They’re getting an answer synthesized from dozens of sources, with citations attached to maybe three of them. Traffic that would have spread across 10 blue links now concentrates on whoever the LLM deemed most authoritative. And the mechanism that decides authority is almost invisible to most builders.
That mechanism is citation. And citation behavior in LLMs is nothing like Google PageRank.
Why Traditional SEO Signals Don’t Transfer to LLMs
Google’s ranking system rewards:
– Backlinks (votes of confidence)
– Keyword relevance (topical match)
– Domain authority (age, scale, traffic history)
– User signals (click-through rate, time on page)
– E-E-A-T (experience, expertise, authoritativeness, trustworthiness)
LLMs, during training, absorb patterns from the internet. But at inference time, when you ask them a question, they’re not running a ranking algorithm. They’re predicting the most likely next tokens based on statistical patterns learned from their training data.
– Recency doesn’t matter much. A well-written 2019 article trains the model just as much as a 2024 one.
– Traffic volume is invisible. Whether a page gets 100 visits or 100,000 is noise to the model weights.
– Backlinks matter only insofar as they influenced the model’s understanding of authority during training. New links post-training don’t help.
– Structural signals (internal linking, site architecture) barely register.
What does matter: presence in high-quality, diverse training data; consistency of voice; clarity of claims; and density of facts that align with the model’s worldview.
If you’re operating a directory or content site, this is the real shift. You’re not optimizing for a crawler that understands links. You’re optimizing for a statistical model that learned patterns from billions of tokens.
Structured Data and the Citation Problem
The first tactical move is to make your content machine-readable in a way LLMs can trust.
Structured data (JSON-LD, microdata) has been Google’s solution for giving crawlers a semantic layer. It tells a search engine: this is a review, this is an author, this is an organization, this is a rating. Google uses it for rich snippets and E-E-A-T signals.
LLMs don’t parse JSON-LD the way crawlers do. But they do train on HTML that contains structured markup, and models like GPT-4 with vision can ingest both text and semantic context. More importantly, when an LLM is prompted to cite aclear, coherent block of information. Structured data increases that traceability
If you’re running a directory of 47,000+ agencies with 4.2M+ aggregated reviews, the difference between unstructured text and properly marked-up claims is enormous. An LLM asked “which agencies have the highest ratings” will be far more confident citing you if it can extract:
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Agency Name",
"aggregateRating": {
"@type": "AggregateRating",
"ratingValue": "4.8",
"reviewCount": "127"
}
}
'''Then if those facts are buried in prose.
Start treating structured data not as a Google ranking signal, but as a citation mechanism for LLMs. Make your claims extractable.
IndexNow, Bing, and the Real-Time Index
Google’s index updates in waves. It can take weeks for new content or updates to fully propagate. That lag mattered when users refreshed their search results daily and clicked the top three blue links.
It matters less when ChatGPT Search pulls from Bing’s index, and Bing’s index is continuously updated
IndexNow is a Microsoft-led standard that lets you ping Bing (and other search engines) the moment you publish or update content. Instead of waiting for a crawler to find your site, you push a notification: “page X changed at timestamp Y, go re-index it.”
Bing uses IndexNow to keep its index fresher. And because ChatGPT Search has access to Bing’s index, faster re-indexing means LLMs get access to your updates sooner.
If you’re publishing data-driven insights or aggregated information that changes (ratings, reviews, new directory entries), IndexNow is no longer optional. It’s the difference between an LLM training on day-old data versus real-time data.
Implement it. It takes 10 minutes and costs nothing.
llms.txt: The Coming Standard for AI Discovery
This is where things get speculative, but the pattern is clear.
OpenAI and Anthropic are exploring how to make crawling and citation more efficient for LLMs. One emerging idea is llms.txt, a convention where sites publish a plaintext file at `/.well-known/llms.txt` containing:
– Your brand voice and guidelines
– Information your site considers most valuable
Think robots.txt, but for language models. You’re not trying to block LLMs from scraping you (you want them to). You’re instructing them on how to cite you, what to trust, and what matters.
No major LLM has formally adopted llms.txt yet. But the standardization pressure is real. MCP (Model Context Protocol) servers, developed by Anthropic, are beginning to formalize how AI systems connect to data sources. Within a year or two, sites that publish machine-readable metadata about their data will have an advantage in LLM discovery.
The Practical Playbook for Builders
If you’re running content or a directory and want to survive the shift from SEO to AI search discovery:
1. Audit your structured data. Make sure every claim, rating, review, and fact is marked up with schema.org types. Treat it as a citation mechanism, not a ranking signal.
2. Implement IndexNow. Push updates to Bing in real-time. Keep your index as fresh as possible.
3. Build in public and with data. LLMs trust specificity and numbers. If you have aggregated data (reviews, ratings, geographic trends), publish it openly. Make it auditable.
4. Diversify your training footprint. Don’t rely on organic search to distribute your content. Publish your insights in multiple formats, in multiple venues. The more places an LLM saw your idea during training, the more likely it is to cite you.
5. Prepare for MCP. If you’re building a data service that LLMs might want to query, start thinking about how you’d formalize that connection. MCP servers are the next layer of AI-native integration.
6. Document your authority. If you aggregate or synthesize data from thousands of sources, make that process visible and auditable. LLMs will cite you more confidently if they understand where your knowledge comes from.
The Real Shift
This isn’t about SEO dying. It’s about discovery fragmenting. Some traffic will still come from Google. But an increasing share will come through LLM intermediaries, and the rules for being cited are still being written.
The builders who understand how LLMs train, cite, and update will own the next decade of discovery. It requires different thinking than optimizing for a search algorithm. But the fundamentals are the same: clarity, specificity, authority, and making your information as machine-readable as possible.
The question isn’t whether your site can rank. It’s whether an LLM will trust it enough to cite it.
