Original evidence reviews on how AI search actually chooses who to cite, recommend and name. We run this research to keep our own audit methodology honest, and we publish it, including the findings that argue against our commercial interest and the claims we decline to sell.
The SEO industry spent 18 months recommending llms.txt for AI visibility. Server logs from 137,000 domains, Google's own statements and the spec author agree: the major AI surfaces do not read it.
45% of consumers now ask AI to find local businesses. Google's AI grounds those answers on Maps and Business Profile data covering 250M+ places, which makes the GBP and review layer the heaviest controllable lever.
Across 75,000 brands, web mentions correlate with AI visibility roughly three times more strongly than backlink authority, and three independent research teams reached the same direction.
YouTube is the most-cited domain in Google AI Overviews and #2 across all LLMs. The mechanism is unglamorous: titles, transcripts and descriptions are what AI systems parse, not the video itself.
Google documents three signals for indicating your preferred image, and an eligibility floor most sites fail: CSS background images are never indexed. The popular stock-photo-penalty claim does not survive evidence grading.
Reddit is a top-3 AI citation source whose share halved in three months, driven by licensing deals and litigation. For most local UK businesses, the honest Reddit strategy is a viability check, then nothing.
61.7% of AI answer appearances link a source without naming it. Citation and mention are different, weakly-correlated outcomes, engines behave in opposite ways, and phrasing swings mention rates 30-50x.
AI Visible sells AI search visibility work: schema markup, audits and implementation for UK small businesses. That gives us a commercial incentive to believe every "AI optimisation" tactic works. The research programme exists to protect our clients, and our own reports, from that incentive. When the evidence said the audit was crediting a signal AI surfaces do not read (llms.txt), we removed it. When the evidence pointed at levers we do not sell (Google Business Profiles, reviews), we published that too.
The method, in short: a minimum of three independent primary sources with non-aligned incentives for any claim that drives a recommendation; a logged vendor-incentive ledger for every source; second-hand statistics excluded from findings; per-study confidence scores; and a limitations section written before publication, not after challenge. If you think we have got something wrong, we would like to hear about it: tell us here.
Our free AI Visibility Snapshot checks your site against the signals this research supports, and none of the ones it debunks. One clear page, within 24 hours.
Get My Free Snapshot →