AI Search Ranking Factors: What's Documented, What's Measured, What's Guesswork

Ten factors come up in advice and research, ranked here by how far we trust them. Only the first two are documented by engines, three more are measured in studies, and the rest are mixed, inferred or contradicted. Nobody publishes weights, so the order is our judgment, not an algorithm.

Engines document access rules, not ranking weights. OpenAI, Perplexity and Anthropic each say blocking their search crawler can keep you out of answers, and Google says a page must be indexed and snippet-eligible. Google adds that unique, useful content will likely matter more than any other suggestion in its guide.

• OpenAI: OAI-SearchBot surfaces sites in ChatGPT search. Opted-out sites are not shown in answers, though they can appear as navigational links. GPTBot, the training crawler, is a separate setting.
• Perplexity: PerplexityBot surfaces and links sites in Perplexity search results. It is not used to train foundation models.
• Anthropic: disabling Claude-SearchBot may reduce your visibility and accuracy in Claude search results.
• Google: AI Overviews and AI Mode retrieve from the Search index and use query fan-out. A page must be indexed and snippet-eligible to appear as a supporting link. Blocking Google-Extended does not affect Search inclusion.

For most brands this gate is already open. In AEOeye's crawler access study (October 2026; 391 well-known sites, a convenience sample), 1 of 193 SaaS sites blocked any AI retrieval bot in robots.txt, against 27 of 38 news publishers. But 24 of 88 e-commerce and consumer homepages challenged our self-identified crawler, so check firewall and CDN rules as well as robots.txt.

One more documented fact: Google says no third-party tool has access to its internal ranking or AI systems. That includes ours. We measure outcomes, not weights.

Three have published evidence: cited, specific claims; third-party coverage; and fresh, well-structured pages. Each study has limits that matter. Read them as evidence of direction, not weights you can drop into a spreadsheet.

Specific, cited claims. The GEO study (KDD 2024) built a 10,000-query benchmark (results on its 1,000-query test split) and a research engine using GPT-3.5-turbo and the top five Google results. Adding statistics, quotations or cited sources lifted a rewritten source's visibility by 30–40% on one metric and 15–30% on another; keyword stuffing gave little to none. On Perplexity.ai, quotation addition scored 29.1 on position-adjusted word count against a 24.1 baseline, and keyword stuffing 21.9. The limit: a simulated engine measuring share of an answer, not whether you get retrieved.

Third-party coverage. A University of Toronto study (data collected August 2025) compared Google with a search-enabled GPT model. For US software-product queries, AI search cited 72.7% earned media and 26.7% brand-owned sites; Google showed 45.4% earned and 43.7% brand. In consumer-electronics queries it also leaned away from community platforms such as Reddit, which sits awkwardly with the popular claim that Reddit dominates AI citations. One engine, product-style queries: do not build a plan on this alone.

Fresh, well-structured pages. GEO-16 audited 1,100 URLs behind 1,702 citations from Brave Summary, Google AI Overviews and Perplexity for 70 product-intent prompts on B2B SaaS pages. Metadata and freshness correlated most with citation (r = 0.68), then semantic HTML (0.65) and structured data (0.63). The authors warn the design is observational and may be confounded.

Entity authority and backlinks are the two big ones. Both are plausible, and we found no AI-specific test of either. Treat them as cheap hygiene, not levers you can promise results from.

• Entity clarity. A consistent name, category, description and profile links help machines tell you apart from lookalikes. Google says Organization markup can help disambiguate your organization in Search, but we found no evidence that answer engines weight it. Do it, because it takes an hour. See what entity SEO is.
• Backlinks. Microsoft calls them essential for discoverability, and Google's AI features sit on its core ranking systems, so links likely help indirectly. Unlinked mentions are a separate question, covered in brand mentions vs backlinks. Google warns that seeking inauthentic mentions is not as helpful as it seems.
• Structured data. Google says structured data is not required for generative AI search, while Microsoft recommends it. Write the visible text first, then add schema that matches it (schema guide).
• llms.txt. Google says it ignores the file. In our study, 173 of 330 checkable sites publish one. We publish one too, and we think it is overrated as a lever.

Self-contained sections are the one place Microsoft and Google disagree: Microsoft describes Copilot breaking pages into smaller pieces, while Google says there is no need to chunk. We side with Google on not fragmenting pages and with Microsoft on writing sections that stand alone.

Fix access, then make claims specific and sourced, then earn coverage where engines already look, then measure repeatedly. That order follows evidence strength, and the first two steps cost only time.

1. Check the gate. Unless you have a reason to block them, allow OAI-SearchBot, PerplexityBot and Claude-SearchBot in robots.txt, check CDN and firewall logs for blocked bots, and confirm key pages are indexed. Our AI robots.txt generator writes the rules.
2. Rewrite your ten most important pages with sourced specifics. The passage-level method is in what makes content quotable by AI.
3. Find where your category's answers come from. Ask engines your buyers' questions, list the third-party sources they cite, and earn honest coverage there. How AI assistants choose brands explains the shortlist.
4. Keep dates honest. Update pages when facts change and show the update date.
5. Measure more than once. Research on AI search measurement finds answers vary across runs, prompts and time, so a single answer is an anecdote.

AEOeye's free audit asks three buyer questions on ChatGPT live and shows who is named instead of you. The $29 report adds Claude, Gemini, Google AI and Perplexity and ten questions. For daily monitoring across hundreds of prompts, dedicated trackers such as Profound, Peec AI or Otterly do that; we do not.