In this article 12 sections
- The shift in how people look for information
- Each LLM cites different sources
- How ChatGPT’s web search really works
- Comparison: where you need to be on each platform
- The Schema Markup factor
- Local SEO: don’t forget Foursquare
- The llms.txt protocol
- Content structure for citation
- Technical requirements to get cited
- What does NOT work
- Strategy by type of business
- The near future
The shift in how people look for information
According to Gartner, traditional search volume will drop 25% by 2026 because of AI chatbots and virtual agents. The question is no longer just “how do I rank first on Google?” but “how do I become the source ChatGPT and Perplexity cite?”
The good news: you don’t need to learn a new kind of SEO. The ChatGPT-4o system prompt leaks reveal that OpenAI uses Bing as its search engine and applies a reranker that evaluates search intent, domain terminology and source type. It’s semantic SEO under a new name. If you want to read more on the subject, I recommend Natzir’s piece on what the GPT-5 leaks teach us about SEO for AI (in Spanish).
Each LLM cites different sources
Something that surprised me when comparing platforms: the lists of most-cited domains in Google AI Overviews, ChatGPT and Perplexity barely overlap. Each one has its favorite sources.
The sources ChatGPT cites most
ChatGPT prioritizes reference sources with structured information, such as Wikipedia, G2, Forbes or Amazon.
The strategy: digital PR and presence in reference media.
The sources Perplexity cites most
Perplexity relies more on community and niche content, such as Reddit, YouTube or LinkedIn.
The strategy: specialized, regional content with a unique perspective.
Google AI Overviews
It follows SEO fundamentals + freshness + entities. If you already rank on Google, you have an advantage.
How ChatGPT’s web search really works
ChatGPT-4o’s leaked system prompt described something that changes the strategy: ChatGPT didn’t query its own index. Every time it needed external information, it searched Bing in real time. OpenAI also has its own search crawler (OAI-SearchBot), so it’s worth not blocking it in robots.txt.
That’s why many ChatGPT answers come without links. It isn’t a bug, it’s by design: if it can answer from its training memory, it doesn’t search.
When ChatGPT triggers web search
The system only searches the web in 4 specific scenarios:
- Real-time information: weather, sports results, stock prices
- Location data: “restaurants near me”, store hours
- Recent news: events from the last few days or weeks
- Niche topics: information that probably isn’t in the training data
If your content covers stable topics that already existed before the model’s cutoff date, ChatGPT will answer from memory without looking you up. If you cover emerging, frequently updated or very specific topics, you have a chance of being cited.
The freshness system (QDF)
ChatGPT uses a 0 to 5 scale to decide how recent the results need to be:
| QDF | Meaning | Example |
|---|---|---|
| 0 | No date preference | ”What is SEO?“ |
| 3 | Moderate preference for recent results | ”SEO best practices” |
| 5 | Only results from the last 30 days | ”Latest Google update” |
When QDF is high, older articles drop out of consideration even if they have better content. Keeping recent update dates is critical.
Parallel queries and boosting
ChatGPT can generate up to 5 parallel searches for each user request. If someone asks something complex, the system breaks the question down and searches multiple angles.
It also boosts important terms with the + symbol. If it detects that “ecommerce” is key to the search, it adds +ecommerce to prioritize results containing that term.
The multilingual factor
A little-known fact: for queries in Spanish or other languages, ChatGPT also runs searches in English as a backup. This means that if your content is in Spanish but competes with English-language sources on the same topic, both get considered.
Comparison: where you need to be on each platform
| Platform | Main index | Preferred sources | Local SEO priority |
|---|---|---|---|
| ChatGPT | Bing | Wikipedia, media, reviews (G2) | Foursquare |
| Perplexity | Multiple (Bing, Google) | Reddit, YouTube, specialized blogs | Google Business |
| Google AI Overviews | Current SERP top 10 | Google Business | |
| Claude | Its own web search (since 2025) | No reliable public data | N/A |
This table changes the strategy: if your audience uses ChatGPT, being on Bing is mandatory and a complete Foursquare listing helps. If they use Perplexity, your presence on Reddit and YouTube matters more.
The Schema Markup factor
If you had to implement a single technical optimization, I’d start with complete Schema:
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "Article title",
"datePublished": "2026-01-17",
"dateModified": "2026-01-17",
"author": {
"@type": "Person",
"name": "Author name",
"url": "https://yourdomain.com/author"
}
}Schema helps LLMs verify authorship, date and context.
Local SEO: don’t forget Foursquare
If you run a local business, watch out: several industry analyses have found that ChatGPT pulls from Foursquare for local business results, not just from Google Business Profile.
To show up in ChatGPT’s local searches:
- Create and verify your Foursquare profile
- Fill in category, hours and exact address
- Upload photos of the premises
- Encourage check-ins and reviews
Google Business Profile is still essential for Google, but for ChatGPT it’s worth covering Foursquare too.
The llms.txt protocol
It’s an open proposal (llmstxt.org): a Markdown file at the root with a summary of the site and links to its key content, designed for an LLM to read quickly. It doesn’t control access (that’s still robots.txt’s job):
# My Ecommerce Store
> Online electronics store with buying guides and comparisons.
## Guides
- [How to choose a laptop](https://mydomain.com/guides/how-to-choose-a-laptop.md): criteria by use and budget
## Comparisons
- [Best phones](https://mydomain.com/comparisons/best-phones.md)No major search engine has confirmed using it for citations, and Google has said it doesn’t use it. It’s cheap to maintain, but don’t expect miracles.
Content structure for citation
LLMs extract fragments. Structure matters:
<!-- Easy to extract and cite -->
<h2>How much does it cost to register a trademark in Spain?</h2>
<p>Registering a trademark costs between €150 and €300.
The OEPM base fee is €144.58 for one class,
plus €95.54 per additional class.</p>
<!-- Hard to extract -->
<p>To register a trademark you have to pay fees
that depend on several factors, such as the number of
classes and whether you're an individual or a company...</p>The pattern that works: question in the H2, direct answer in the first paragraph, concrete data.
Technical requirements to get cited
For an LLM to be able to cite you, your content has to meet at least one of these criteria:
| Requirement | ChatGPT | Perplexity | AI Overviews |
|---|---|---|---|
| In Bing’s index | ✅ Mandatory | ✅ Recommended | ❌ Not applicable |
| In Google’s index | ❌ Not applicable | ✅ Recommended | ✅ Mandatory |
| In training data | ✅ Alternative | ❌ Not used | ❌ Not used |
| Schema implemented | ✅ Helps | ✅ Helps | ✅ Improves the snippet |
| HTTPS | ✅ Mandatory | ✅ Mandatory | ✅ Mandatory |
If your site isn’t on Bing, ChatGPT literally can’t find you even if you have the best content in the world.
What does NOT work
AI content to rank in AI: the ultimate irony. Generic AI-generated content lacks E-E-A-T, and reranking systems filter it out. LLMs look for sources with real expertise, not content that sounds just like their own output.
Ignoring Bing: if your site isn’t on Bing, ChatGPT can’t find it. Check indexing in Bing Webmaster Tools.
Keywords over semantics: LLMs understand context, they don’t count keywords. Write for humans.
Undated content: LLMs prioritize content with clear publication/update dates. An article with no visible date is less likely to be cited on topics where freshness matters.
Strategy by type of business
Not every business should optimize for LLMs in the same way:
Ecommerce
- Priority: Product schema, reviews, technical specifications
- Tactic: be on G2, Trustpilot and comparison marketplaces
- Key channel: ChatGPT (it looks for products with specifications)
SaaS / B2B
- Priority: technical documentation, use cases, comparisons
- Tactic: publish in the sector’s reference media, and on GitHub if relevant
- Key channel: Perplexity (it looks for specific solutions)
Local business
- Priority: consistent NAP, opening hours, photos
- Tactic: Foursquare + Google Business Profile
- Key channel: ChatGPT for local (Foursquare), Google for AI Overviews
Content / Media
- Priority: E-E-A-T, clear authorship, visible dates
- Tactic: complete Article schema, verifiable author presence
- Key channel: all three (depends on the topic)
The near future
The ChatGPT-4o leak shows that OpenAI is constantly refining its search system. Some patterns we can anticipate:
- Greater integration with verified sources: E-E-A-T will keep gaining weight
- Penalties for generic content: rerankers are getting better at spotting content with no value
- Structure matters more: Schema and semantic HTML will become increasingly critical
- Multilingual competition: English-language content will compete with local content
The strategy that works today will keep working: useful, well-structured content with verifiable authorship that’s updated regularly.
Google sums it up in its helpful content guidance: create for people, not for algorithms. The same applies to LLMs, which in the end are trained to understand what humans find useful.