AI trust signals is one of those phrases that gets used confidently and defined rarely. Here is what the evidence actually supports, and — just as usefully — what it does not.
What is genuinely evidenced
Corroboration from third-party sources
The strongest large-scale evidence we use is correlational rather than causal. Ahrefs studied 75,000 brands and found that YouTube mentions showed the strongest correlation with AI visibility at about 0.737, while branded web mentions correlated at roughly 0.66–0.71. The practical signal is broad, consistent brand presence across the web — not a single magic publisher.
Practically: do not treat your own website as the whole evidence set. Accurate external mentions, useful video, credible profiles and independent coverage can broaden the contexts in which a brand is encountered. The value comes from the quality and consistency of that footprint, not from manufacturing mentions for a score.
Unlinked brand mentions
Ahrefs studied 75,000 brands and found branded web mentions correlate with AI citation at 0.664, and YouTube mentions at 0.737 — against 0.218 for backlinks. The researchers are explicit that this is correlation, not causation, and we repeat that caveat rather than hiding it. But the gap is large and consistent.
Being readable at all
Not a “trust” signal exactly, but it gates everything. Vercel found that several dedicated AI crawlers do not render client-side JavaScript, while Googlebot does. Content that exists only after hydration is therefore fragile across assistants; critical facts should be present in the initial HTML. Text baked only into an image is also a poor source for machine extraction.
Being allowed in
A surprising number of sites block AI crawlers by accident. Cloudflare introduced a managed robots.txt in July 2025 that can add AI-bot disallow directives automatically — a site can be excluded without its owner ever editing a file. And GPTBot, OAI-SearchBot and ChatGPT-User are three different agents with three different consequences: blocking GPTBot only stops training, while blocking OAI-SearchBot stops you appearing in ChatGPT search entirely.
What is repeated without evidence
These appear in a great many AI visibility audits. We do not sell them.
- llms.txt as a proven visibility lever. Ahrefs analyzed 137,210 domains and found that 97% of published llms.txt files received no requests in May 2026. Google also says llms.txt is not needed for its generative AI Search features. It is inexpensive to publish and some tools may fetch it, but current evidence does not justify treating it as a ranking or citation signal.
- “Special AI schema.” Google says no special structured data is required for its generative AI features. Valid schema can make identity and page relationships explicit and can support eligible Search features, but there is no separate AI-only markup that guarantees visibility.
- Page-speed thresholds for AI crawlers. Numbers like “under 800ms” and “a 10-second timeout” circulate widely. Traced to source, they are unattributed assertions in blog posts. No AI company publishes a crawl timeout.
- sameAs as a direct ranking lever. Structured identity links can reduce ambiguity and are useful entity hygiene, but there is no public evidence that adding a sameAs property by itself increases AI citations or rankings.
- Indexing on Bing as a guarantee of ChatGPT visibility. OpenAI now documents OAI-SearchBot as the crawler used to make websites eligible for ChatGPT search results. Bing indexing can still be useful for search generally, but it is not a substitute for allowing OpenAI’s own search crawler.
The uncomfortable conclusion
The hardest part is also the least controllable: building a useful, consistent reputation beyond your own domain. On-site work improves crawlability, clarity and extractability; it does not by itself establish how often an assistant will mention or recommend the business. Any supplier that treats technical readiness as the same thing as measured visibility is collapsing two different questions into one.
We say this on our own methodology page as well, because it is the part clients least want to hear and most need to.