The old model: Google's PageRank principle
Google's core algorithm, originally developed by Larry Page in 1996, treats links like votes. Whoever receives many links from trustworthy sites is rated as relevant and ranked higher. This mechanism was the foundation of the entire web for 25 years.
The result: SEO became a link-building game. Agencies built links, ranked pages, and earned from traffic. Content was often secondary — as long as it was optimized for the right keywords and equipped with enough backlinks. The game was learnable, scalable, and highly profitable for service providers.
The new model: semantic understanding
AI search engines like Perplexity, ChatGPT Search, and Google's Gemini-powered AI Overviews work fundamentally differently. They use large language models (LLMs) that understand text semantically — they don't just count keywords, they grasp meaning, context, and connections.
What that means in practice: a document that answers the question "How do I reduce support costs with AI?" clearly and completely will seem more important to an AI search engine than a page with a hundred backlinks that circles the topic vaguely.
How Perplexity works internally
Perplexity crawls the web in real time when a query is submitted. It doesn't look for the page with the most links — but for the page that answers the question most precisely. The mechanism resembles a very well-trained researcher who finds and synthesizes the most relevant documents within seconds.
- Real-time web search via its own crawler infrastructure
- Semantic relevance scoring — not just keyword matching
- Source selection based on clarity, recency, and recognizable authority
- Direct citation with a source link — sources are visible
How Claude and ChatGPT work with training data
In non-search mode, Claude (Anthropic) and ChatGPT (OpenAI) have static knowledge from their training cutoff. Websites that were frequently cited during training have a higher presence in answers. That makes regular, high-quality content more important than ever — not for rankings, but for the likelihood of being part of the next training corpus.
Content that is referenced across many reputable sources ends up in the training corpus of future AI models more often — a feedback loop of visibility that reinforces itself.
Claude, ChatGPT, Perplexity — different strategies
The three most important AI search platforms follow different approaches, which suggests different GEO optimization strategies:
Claude (Anthropic) is LLM-native and, in its search function, cites reliable, well-structured sources. Anthropic explicitly trains on E-E-A-T-like quality signals — authorship and structured metadata play a particularly large role.
Weiterlesen — kostenlos
Den vollständigen Inhalt freischalten
Trag deine E-Mail-Adresse ein und bestätige sie: Du abonnierst den Signal-Forge-Newsletter von FORGE und erhältst sofort Zugang zu diesem und allen weiteren registrierungspflichtigen Inhalten. Die Abmeldung ist jederzeit möglich.
Schon registriert? Der Link aus deiner Bestätigungs-Mail schaltet dieses Gerät wieder frei.
ChatGPT Search combines the Bing index with its own crawl and OpenAI-specific relevance scoring. Recency and authority are weighted similarly to Claude, but the Bing foundation means classic SEO signals are not irrelevant.
Perplexity is an AI search engine with explicit source attribution. Every answer shows source links — which makes Perplexity the most direct vehicle for GEO visibility. Whoever gets cited on Perplexity is visible to the user.
The 5 differences from Google
| Attribute | AI search engines | |
|---|---|---|
| Result type | Ranking list (10 links) | Direct answer with sources |
| Primary signal | Backlinks + PageRank | Semantic relevance + E-E-A-T |
| Author identification | Optional, helpful | Critical — machine-readable required |
| Recency | Freshness signal present | Fresh content carries significantly more weight |
| Content format | Keywords + structure | Complete answers + citation-worthiness |
What AI models recognize as "trustworthy"
From my work on various projects at the intersection of AI and content, five reliable signals can be identified that AI models treat as trust indicators:
- Clear authorship: Name and credentials visible,
schema.org/Personmarkup present - Specific, answerable questions: No vague buzzword text, but direct answers to concrete questions
- Consistent expertise: The author always writes about the same topic — not a general store
- Citable facts and figures: With a source, not "studies show that..."
- Structured HTML: H1 → H2 → H3, semantic elements (
<article>,<section>) instead of<div>spaghetti
Comparison: one query on Google vs. Claude
To make the difference tangible, here is a concrete example of the same query on both platforms:
Query: "How do I optimize my blog for AI search engines?"
Google result:
→ 10 links of varying quality
→ User clicks link 1-3, reads an article
→ Conversion depends on the website
Claude result:
→ Direct answer: "For GEO optimization I recommend:
1. Implement Schema.org Article markup
2. Clear authorship with Person schema
3. Complete answers instead of teaser content"
→ Sources: [forgeproject.eu/blog], [other sources]
→ User has the answer — may click through to a sourceConsequences for your content strategy
You no longer optimize for an algorithm — you optimize for a language model trained to react the way a careful human reader would. In concrete terms, that means:
- Write for real questions, not for keywords
- Make authorship unmistakable — name, credentials, Schema.org
- Use clear structure — H1, H2, H3 nested logically
- Answer questions completely — no teaser articles that lure the reader to the next page
- Publish regularly — recency is a weighting factor
In the next article: the GEO audit — the 5 concrete factors you can check and improve on your website today.