AI search referral traffic is no longer hypothetical. Across the Ignis client portfolio we've measured 8 to 22 percent of new organic visits arriving from ChatGPT, Perplexity, Claude or Gemini referrals in the last 90 days. The brands getting cited are not the brands ranking #1 on Google. The optimisation surface is different.
What AI engines optimise for when they cite a source.
Signal one. Factual density. The page makes specific claims with specific numbers and specific named entities. Not "many businesses see a lift", but "at VBC monthly site sessions went from 19,000 to 43,000 in nine months". Generic prose is invisible to AI citation. Specific claims are quotable, and the engines need quotable to cite.
Signal two. Schema.org structured data. AI engines parse JSON-LD aggressively because it's the cleanest way to extract entity relationships without natural-language ambiguity. Organization, Person, Service, FAQPage, Article, BreadcrumbList. The five schemas that move the needle. Cross-link them with @id refs so the engine can resolve "the founder of X" without inferring.
Signal three. Question-shaped headings. Pages organised around the questions buyers actually ask outperform pages organised around topics. "How does the partnership work?" beats "Our process". The H2s should read like the queries someone would type into ChatGPT. The engine matches against query semantics, not topic taxonomy.
Signal four. The llms.txt file. A small markdown file at /llms.txt that summarises what the site is about, who it's for, what the headline numbers are, and where the canonical pages live. Treat it as a press release for AI crawlers. Update it whenever the offer changes.
Signal five. Author + entity provenance. Pages with a named author who has a Person schema linked to a verified Organization get cited at higher rates than anonymous pages. The engines need to know "who is making this claim" before they propagate it.
Signal six. Backlinks from publications the engines trust. Different from Google. AI engines weight Wikipedia, Github, official .gov, .edu, and large industry publications more heavily than Google does for the same query. Earning one cite from a Wikipedia entry can produce more AI traffic than 100 mid-tier blog backlinks.
What does NOT help. Keyword stuffing (engines deduplicate aggressively). Long-tail content with no original substance. Generic stock images (engines can't verify them). Pages built with default templates (the engines see the template before they see your content).
The shift to AI search is not a future event. It already arrived. The pages we built in late 2024 that followed the recipe above are now responsible for 15 to 22 percent of inbound discovery for clients in the portfolio. The pages built three years ago to "rank on Google" are not getting cited.