What actually gets you cited by AI

Every other factor is conditional on this one. If your content is not in the retrieval corpus — because it renders only after JavaScript, because a crawler is blocked, because a misconfigured canonical tag consolidated your pages into one — nothing else you do matters. This is unglamorous and it is where most real failures live.

Analysis of millions of AI citations in 2026 found cited URLs skew meaningfully fresher than the web at large — roughly a quarter fresher on average. This is one of the few levers where the effect is measured rather than argued, and it is cheap to act on: real publication and modification dates, and content that has genuinely been revised rather than merely re-stamped.

Source: Ahrefs, 2026

A large factorial study published in 2026, running over 250,000 trials, found that a passage’s relevance to the query and its position within the retrieved context dominated the other variables tested. In plain terms: answering the specific question directly, early in the passage, beats almost every stylistic tactic.

Correlational work in 2026 found brand mentions tracked AI visibility considerably more closely than backlinks did. Correlation is not causation and the direction is arguable — well-known brands get mentioned and cited for the same underlying reason — but the practical implication is sound: presence in the wider corpus matters, and it is not something you can build entirely on your own domain.

Reddit is the most-cited domain across several engines, accounting for roughly 16.7% of ChatGPT citations in one 2026 analysis, with Wikipedia and YouTube close behind. Engines diverge sharply here — Gemini cites Reddit at a negligible rate — so this is a per-engine consideration rather than a universal tactic. It is also a reminder that your owned pages are not the only surface that determines whether you appear.

Source: Ahrefs, July 2026

Retrieval systems operate on passages, not whole pages. A paragraph that depends on three preceding paragraphs cannot be lifted into an answer without becoming wrong or unintelligible. The mechanism is clear and the practice costs nothing; what lacks direct evidence is any specific formatting prescription. Write sections that stand alone because it is obviously sensible, not because a study proved a particular word count.

Answer engines stake their credibility on the claims they repeat, which gives them reason to prefer sources that show their work. Original research also earns the brand mentions and platform presence discussed above. The evidence here is more mechanistic than measured, but the downside risk is zero.

A 137,000-site study found 97% of valid llms.txt files received zero requests in the month measured, and Google has documented that it does not use the file for Search or AI Overviews. It has real value as tooling for coding agents pointed at documentation — Google added it to Lighthouse’s agentic-browsing audits — but it is not a citation lever. Publish one if it helps agents use your docs; do not expect visibility from it.

Source: Ahrefs, June 2026

Google has publicly stated that structured data is not a ranking or inclusion factor for AI Overviews. Practitioners including Mike King at iPullRank dispute that guidance as self-serving. What is not contested: schema drives rich results in classic search, resolves your organization and authors to known entities via sameAs, and feeds the product data that agentic commerce protocols consume. Implement it for those reasons, which are solid, rather than for a citation effect that is disputed.

Rewriting pages into artificial fragments to pre-chunk them for retrieval is not supported by evidence, and Google has explicitly said it is not required. Worse, aggressive citation-focused rewrites can impair retrieval by stripping the context that made a passage meaningful. Write well-structured prose; let the retrieval system do its own chunking.

A critical survey of 45 studies published in July 2026 found that several widely-cited optimization-lift figures were conditional on a source already being present in the retrieval corpus, with the gains eroding once competitors applied the same techniques. Treat any specific percentage improvement claim in this field with suspicion — including ones that flatter tactics you already use.

Source: Martinez et al., July 2026

Once a reliable proxy, now much weaker. The share of AI Overview citations drawn from top-ten organic results fell from roughly 76% to about 38% inside a year, and ChatGPT’s most-cited pages overlap Google’s top ten only about 6.8% of the time. Ranking still correlates with being retrievable and credible; it no longer predicts citation.

Source: Ahrefs, 2026