We asked four AI engines the same 12 SEO questions. They almost never cited the same sources.
Across 48 answers, four AI engines cited 158 different pages and 139 of them only once. One engine leaned on Google's own docs for three quarters of its sources; another almost never did.
The short answer
We put 12 common SEO questions to four AI models with live web search and collected every source they cited. The 48 answers cited 158 different pages, and 139 of those pages were cited only once. On a typical question, two engines shared roughly one domain in eight. The bigger surprise was taste in sources: OpenAI’s model took 74% of its sources from Google’s own documentation, while Anthropic’s model took 2%.
What did we test?
On 12 November 2026 we sent the same 12 questions to four models through OpenRouter, each using its maker’s own web search:
| Engine | Model | Search |
|---|---|---|
| Perplexity | Sonar Pro | Perplexity’s built-in search |
| OpenAI | GPT-6.1 Sol | OpenAI’s native web search |
| Anthropic | Claude Sonnet 5.5 | Anthropic’s native web search |
| Gemini 4 Carbon | Google’s native search grounding |
Every call used the same instruction: answer a website owner in under 120 words, use web search, then list every URL relied on after a line reading SOURCES. The questions were the kind people actually type, from “How do I get my website on the first page of Google?” to “Should I block AI crawlers like GPTBot in robots.txt?” and “Does an llms.txt file help with AI search visibility?”. The full list, the exact prompt and every answer are in the raw data.
All 48 calls returned an answer with sources. Gemini needed retries on four questions (its reasoning used up the token budget), and every retry succeeded. Gemini sometimes listed Google redirect links instead of pages; we resolved all 32 of them the same evening. Before counting, we stripped tracking parameters and trailing slashes so that one page counts once.
How much did the engines agree?
Very little. For each question we compared the set of domains each pair of engines cited. The average overlap (shared domains divided by all domains the pair cited) was 0.12. Only one question, “How long does SEO take to show results?”, had a domain that all four engines cited: Search Engine Land.
| Measure | Result |
|---|---|
| Answers | 48 (12 questions × 4 engines) |
| Citations, after removing duplicates within an answer | 190 |
| Different pages cited | 158 |
| Pages cited only once | 139 (88%) |
| Different domains cited | 109 |
| Average domain overlap between two engines on one question | 0.12 |
| Questions where all four shared a domain | 1 of 12 |
Two questions, on getting mentioned in ChatGPT answers and on keyword research for a small business, had no page or domain in common between any two engines.
Which sources did they reach for?
Google’s developer documentation was the one constant. developers.google.com appeared in 23 of the 48 answers, more than twice as often as the next domain. After it came SEO trade press and tool vendors:
| Domain | Answers citing it (of 48) |
|---|---|
| developers.google.com | 23 |
| searchengineland.com | 10 |
| semrush.com | 9 |
| ahrefs.com | 6 |
| seo.com | 5 |
| support.google.com | 4 |
The engines differed most in how much they leaned on that primary source. Counting each domain once per answer, this is the share of each engine’s sources that were Google-owned:
OpenAI’s model cited the fewest sources, 2.4 pages per answer from just 7 domains across all 12 questions, and most of them were Google Search Central pages. Anthropic’s model went the other way: 39 domains, almost all of them SEO agencies, tool vendors and independent blogs, with one Google page in 12 answers. Perplexity cited the most pages per answer (5.6) and the widest spread of domains (48). Gemini sat in between and cited Google’s own pages less often than OpenAI’s model did.
The robots.txt question shows the split clearly. OpenAI’s and Google’s models answered from Google’s documentation and OpenAI’s crawler pages. Perplexity’s and Anthropic’s models answered from hosting companies, agencies and developer blogs.
What does this mean for being cited?
Three readings hold up within the limits below.
- There is no single “AI answer” to optimise for. When engines agree on roughly one domain in eight, a page cited by one engine is usually invisible to the others. Visibility has to be checked per engine.
- Primary documentation is the one shared source. On questions Google has written about, its own pages were the most likely citation across engines. A page that only restates Google’s guidance competes with Google itself.
- The long tail is wide open. 139 pages were cited once. Independent blogs and small agencies appear next to Semrush and Ahrefs, which matches the controlled studies that find content, not only domain size, shapes which source is used.
What are the limits?
This is one run of 48 questions on one evening, so treat it as a sketch rather than a measurement.
- API models, not the consumer apps. We called the models through OpenRouter with each maker’s search tool. ChatGPT, Gemini, Claude and Perplexity as people use them have their own retrieval settings and may cite differently.
- One sample per question. Temperature was left at each provider’s default (two of the models do not accept it), so a second run would likely return different sources.
- Only written URLs count. We asked each model to list its sources and counted what it wrote. The engines’ structured citation data was not available through the tool we used, and Perplexity referred to more sources in its text than it listed.
- Our wording shaped the answers. “Answer in under 120 words” favours short source lists, which may understate how many pages each engine reads.
- We did not judge quality. Every listed URL was fetched; none returned 404, but 23 blocked automated requests and were not checked further.
The raw answers, the question list and the analysis script are published so anyone can recount or repeat the run.