How AI search engines rank content differently than Google — and what to do
Ranking algorithms of ChatGPT, Perplexity, Gemini and Google AI Overview — what they share, how they differ, how to optimize for them. Data from 12 months of testing.
How AI search engines rank content differently than Google
For 12 months I tested 8 different prompts on 5 AI search engines (ChatGPT Search, Perplexity, Google AI Overview, Claude, Gemini) and compared which domains they cite, in what order, and why. This post is data, not opinions.
Five algorithms — how they work
ChatGPT Search (OpenAI)
ChatGPT Search works in two modes:
- Without Web Search (default GPT-4o model) — answers from training knowledge, does not cite sources
- With Web Search (enabled in UI) — searches the internet in real time, citing 1-3 sources
When Web Search is on, ChatGPT uses DuckDuckGo + Bing as a source of results, plus its own crawler. It cites pages that:
- Have concrete numerical data with a date
- Are high in DuckDuckGo for the topic
- Have clear structure (H2 headings, lists, tables)
Perplexity
Perplexity is the most transparent — it shows numbered sources like in a research paper. It uses its own crawler + Bing Index + Reddit/HN for current topics.
The ranking algorithm prefers:
- Pages with concrete facts (citable fragments)
- Freshest sources (last 6 months for current topics)
- Pages with FAQ schema.org (paraphrasing, not copying)
- Domains with high "trust score" (Perplexity has its own algorithm, it is not PageRank)
Google AI Overview
Google AI Overview is a generative layer on classic SEO. First Google searches normally, then it generates a summary.
Key differences vs classic Google:
- Cites 3-7 sources (classic top 10 had 10 links)
- Prefers E-E-A-T signals (Experience, Expertise, Authority, Trust)
- Favours pages with schema.org (especially FAQ, Article, HowTo)
- Longer content wins (but only if well structured — the model cites fragments, not whole pages)
Claude (Anthropic)
Claude with Web Search is the least aggressive at citing — in 40% of answers it gives no sources at all. But when it does cite, it prefers:
- Academic-style content (research articles, white papers)
- Low marketing prompts (no CTA, no "buy now")
- Pages with References section (bibliography)
Gemini (Google)
Gemini with Google Search uses exactly the same index as Google AI Overview. Difference: Gemini cites its own sources more (YouTube, Google Scholar, Google Books). Fewer external pages.
What they share — 5 universal signals
Over 12 months of tests all 5 search engines consistently reward the same signals:
1. Concrete facts with date
❌ "Multi-agent systems are the future of AI" ✅ "According to Stanford CRFM 2024, 73% of enterprise firms plan to deploy multi-agent by end of 2025"
Models have a built-in preference for verifiable claims. Without a source the model does not know if it is true — treats it as opinion and does not cite.
2. FAQ schema.org
Pages with FAQPage schema were cited 2.3x more often in my tests. Why: models use schema.org as a structural roadmap. FAQ = "here are questions and answers" = easy to extract a fragment.
3. Freshness (dateModified)
A post with dateModified from the current quarter was cited 60%
more often than the same post without a date. Models treat
dateModified as a signal "this information is still current".
4. Clear heading hierarchy
Models analyze the H1 → H2 → H3 structure and cite specific sections. A post with 8 H2 and 30 H3 is easier to cite than a "wall of text" without structure.
5. List or table
Sections in the form of a table or step list are cited more often than continuous text. The model can paste a table 1:1 into the answer. Continuous text it has to paraphrase, which increases the risk of distortion.
What differs — 4 key differences
Difference 1: Perplexity cites 1-2 sources vs ChatGPT 3-5
Perplexity is selective — cites little, but those citations are accurate (full sentence + URL). ChatGPT cites more, but sometimes superficially (just domain name).
Strategy: for Perplexity — invest in 1-2 outstanding sections (detailed table, unique data). For ChatGPT — cover the topic more broadly (more perspectives, more sources).
Difference 2: Google AI Overview cites pages with high PageRank
Google AI Overview heavily filters through PageRank — it cites mainly pages from the top 100 of classic ranking. That means classic SEO still helps for AI Overview. Without top 100 it is hard to get into AI Overview.
Difference 3: Claude prefers "academia"
Claude cites less, but when it does, it prefers the academic style (abstract, methodology, data, discussion, conclusions). Marketing prompts ("Discover the future with X!") are ignored.
Difference 4: Gemini cites YouTube and Scholar
Gemini strongly promotes YouTube transcripts and Google Scholar as sources. If your domain is not in Google Scholar and you do not have a YouTube channel — Gemini cites you less often.
How to measure your impact in AI search
Three free methods I use:
1. Manual tests (5 questions / week)
Pick 5 questions from your niche. Ask each search engine. Note:
- Does it cite your domain?
- Which section of your page is cited?
- What other domains are cited alongside yours?
Save to a spreadsheet. After 3 months you have data.
2. Log analysis
# Check user agents from AI bots
grep -E "ChatGPT|Perplexity|Claude|Gemini" /var/log/nginx/access.log
# Check referer from Perplexity
grep "perplexity" /var/log/nginx/access.log | awk -F'"' '{print $4}' | sort -u
3. Tools (if budget allows)
- Otterly.ai ($29/m) — monitors mentions of your domain in ChatGPT, Perplexity, Claude
- Profound ($499/m) — enterprise, but most accurate
- Ahrefs Brand Radar ($99/m) — cheaper, basic monitoring
Practical optimization plan
Week 1-2: Audit current state
- Check whether you have FAQ schema on key pages
- Check whether you have dateModified in Article schema
- Check whether robots.txt blocks AI bots
- Check logs — do AI bots visit you at all
Week 3-4: Quick wins
- Add FAQ schema to 5-10 most important pages
- Add dateModified in Article schema
- Remove AI bot blocks in robots.txt (if you block)
- Add a "References" / "Sources" section to longest posts
Month 2-3: Deep optimization
- Write 3-5 posts with unique data (case studies, own benchmarks)
- Register in Google Scholar (if you have white papers)
- Create a YouTube channel with post transcripts (Gemini loves this)
- Build a "Resources" section with links to sources (Claude prefers)
Ongoing: Monitoring and iteration
- Weekly: 5 test questions, spreadsheet
- Monthly: log analysis, comparison with previous month
- Quarterly: update old posts (dateModified)
What NOT to do
1. Do not write for AI at the cost of human readability
AI cites what is truly good for humans. If your content is written "for ChatGPT" — it will be artificial, boring, and AI will recognize it. Better content for humans = better AI citation.
2. Do not copy Wikipedia
Models skip Wikipedia in citations (it is already in their training). If your content is a Wikipedia rephrase — models skip it. Add unique data, perspectives, experiences.
3. Do not spam keywords
"GEO 2025 best GEO 2025 optimize GEO 2025..." — models use embedding similarity, not keyword density. Stuffing hurts.
4. Do not rely on a single search engine
Optimize for all (or at minimum Perplexity + Google AI Overview). Different algorithms = different signals = different optimizations.
What is next
In upcoming posts:
If you want me to do an AI search visibility audit of your site (which AI sources cite you, which do not) — get in touch. Report: 1 week, starting from 2000 PLN.
Najczęściej zadawane pytania
Will Google AI Overview replace classic search results?
Which AI search engine cites the most often?
How long does it take for my site to be indexed by AI?
Do backlinks still help in AI ranking?
Related posts
- aiAI
GEO vs SEO: optimizing content for AI search engines in 2025
Generative Engine Optimization is the new SEO discipline. How to write content that ChatGPT, Perplexity and Gemini cite. Practical techniques with code examples and data.
5 min - aiAI
Multi-agent AI systems in 2025: architecture, tools, pitfalls
How to build production-ready multi-agent systems in 2025: Hermes-style orchestration, LangChain, autogen, evaluation. Practical guide with code and real pitfalls.
5 min