Definition
AI search ranking factors are the signals that influence which sources an AI search system retrieves, which of them it cites, and which brands it names in the answer. Unlike Google's classic ranking systems, which Google has documented in part over many years, AI answer engines publish almost nothing about how they choose sources, so most of what is called a 'ranking factor' here is inferred from experiments and large-scale observation rather than confirmed.
It helps to separate the stages an answer goes through, because different signals act at each one:
- Eligibility. The page has to be reachable by the crawler the platform uses and present in its index. Google states that to appear as a supporting link in AI Overviews or AI Mode a page must be indexed and eligible to show a snippet, with no other technical requirement. OpenAI says sites that block OAI-SearchBot will not be shown in ChatGPT search answers. Robots.txt rules, CDN bot blocking and JavaScript-only content all fail here.
- Retrieval. The system runs one or more searches, often several sub-queries through query fan-out, against its index. Pages that rank for those sub-queries in the underlying engine are the candidates. Google's Liz Reid has said AI Mode uses Google's existing ranking systems to ground its answers, so classic SEO strength matters for Google's AI features.
- Selection and citation. From the candidates, the model picks passages that directly support what it wants to say. Clear, self-contained statements, specific numbers and structured comparisons are easier to use than vague or promotional copy.
- Brand naming. Which brands appear in the answer text depends on the retrieved pages and on the model's training data, so third-party coverage, reviews and consistent descriptions across the web influence mentions even when your own page is not cited.
The best public evidence on content signals is still the 2024 KDD paper that coined generative engine optimization. Testing on 10,000 queries, the authors found that adding citations, quotations from relevant sources and statistics to a page raised its visibility in generative engine responses by up to 40%, and by up to 37% on Perplexity, while keyword stuffing did not help (Aggarwal et al.). Observational data adds platform-specific patterns: in our tracking, listicles make up about 27% of the content Google AI Overviews cite, and engines differ widely in how many sources they cite per answer (average sources per response).
What does not appear to work is also informative. Google says no special schema, AI text files or markup are needed for its AI features, and our analysis found that llms.txt has no measurable impact on AI citations. Signals also shift with each model update. ChatGPT's average citations per response dropped across models around the GPT-5.3 rollout in March 2026, which changed citation odds for every site at once.
In practice, teams treat AI search ranking factors as a checklist of probable signals: crawler access and indexing, strong rankings for the sub-queries a prompt implies, information gain through original data, passages that answer one question cleanly, content freshness for time-sensitive topics, author authority, and consistent third-party mentions. Then they test, because the only reliable way to know what a given engine rewards in your category is to track its answers.
Examples of AI Search Ranking Factors
- A software vendor adds a table of measured load times and named test conditions to its comparison page and starts being cited on Perplexity for 'fastest CRM for large datasets', where it was absent before.
- A health publisher finds its pages are never cited in ChatGPT, checks its logs and sees OAI-SearchBot receiving 403 errors from a bot-protection rule; citations appear within weeks of the fix.
- An ecommerce brand notices that AI Overviews in its category mostly cite 'best of' listicles from review sites, and shifts effort toward getting its products included in those lists.
- A B2B team adds an llms.txt file expecting more citations, sees no change after two months of tracking, and redirects the effort to rewriting thin product pages.
