Optimizing for AI citations: How one site got 65% ChatGPT traffic

A new passport-photo tool achieved 65% of its sessions from ChatGPT and 45 citations from Microsoft Copilot within three months. This outcome stemmed from a deliberate strategy focused on machine-readable content and structured data rather than traditional SEO.

A new passport-photo tool, barely three months old, recorded 65% of its total sessions from chatgpt.com. Concurrently, Microsoft Copilot cited the site 45 times over the same period, according to Bing Webmaster Tools. This traffic profile diverges sharply from conventional organic search, where the site ranks on pages 3-5 for competitive terms and receives only 6% of its sessions from Google organic search. The founder attributes this performance to specific technical optimizations for AI answer engines, not domain authority or backlinks.

AI answer engines, including ChatGPT search, Perplexity, Copilot, and Gemini, operate differently from traditional search engines. They synthesize direct answers rather than presenting a list of ten blue links. This model prioritizes pages that clearly and verifiably answer user questions, are easily crawlable, and feature structured data for straightforward extraction. For a nascent domain, this approach bypasses the typical 6-12 month timeline required to build domain authority and backlinks for competitive Google rankings.

What they shipped for AI extraction

The founder implemented a multi-pronged technical strategy, focusing on explicit signals for AI systems. This involved a novel llms.txt file and specific Schema.org structured data types. The goal was to make the site's content maximally machine-citeable.

Implementing llms.txt

The first step involved creating an llms.txt file, an emerging convention found at llmstxt.org. This plain-text file, located at /llms.txt, provides AI systems with a curated summary of the site. The founder's llms.txt includes the brand entity, a canonical list of facts, the most important pages, the target audience, and specific guidance on how to cite the site. This file serves as a direct, low-cost method to present a clean, structured model of the site to AI engines, offering explicit instructions on its content and purpose.

Targeting Schema.org structured data

Not all Schema.org types are equally effective for AI extraction. The founder prioritized specific types that explicitly package answers for machine consumption. This included FAQPage with Question and acceptedAnswer pairs, placed on both the homepage and a dedicated facts page. Each Q&A pair functions as a discrete, extractable answer. Additionally, individual pages utilized Question as mainEntity, signaling the primary question each page addresses directly, with the answer attached. This provides the cleanest possible signal: "this page answers exactly this."

Another critical implementation was the Dataset Schema.org type on the page documenting the site's public data API. Google Dataset Search and various AI engines are known to treat Dataset JSON-LD as a citation-worthy source. Finally, SpeakableSpecification was used to mark specific parts of the page suitable for text-to-speech, catering to voice assistants and further enhancing machine readability. The founder provided code examples for the per-page Question as mainEntity implementation, emphasizing its role in signaling a page's primary answer.

What we'd change for broader application

The founder's success with a passport-photo tool highlights a specific advantage for highly transactional, fact-based services. The effectiveness of FAQPage and Question as mainEntity relies on content that provides definitive, concise answers. This strategy may prove less impactful for sites with more nuanced content, subjective opinions, or complex narratives that are not easily distilled into discrete Q&A pairs. Founders building content platforms or B2B SaaS with extensive feature sets might find direct application of this schema more challenging, requiring significant content restructuring.

The llms.txt convention, while promising, remains an emerging standard. Its long-term adoption and the extent to which major AI models will consistently honor its directives are not yet fully established. Relying heavily on an evolving standard introduces a degree of risk; future changes in AI crawling or interpretation could diminish its efficacy. Furthermore, the founder noted aspects they

Pull quote: “”