How AI Search Engines Choose Which Brands to Recommend

Why AI ≠ SEO

When we started tracking AI engine citations in early 2025, we assumed AI visibility would closely mirror Google rankings. After all, AI engines claim to use web search as a grounding layer. We were wrong.

In our analysis of over 10,000 AI responses across 400+ industry prompts, 80% of AI-cited URLs had no measurable Google ranking for the query. The correlation between Google position and AI citation probability is 0.18 — barely above noise. Something else is driving AI recommendations, and understanding that signal is the difference between being invisible and being the brand AI engines recommend.

The 7 Factors That Predict AI Citations

We ran regression analysis on citation data across ChatGPT, Perplexity, and Google AI to identify which signals actually predict citation probability. Here's what we found, ranked by predictive power:

Referring Domains: The #1 Predictor

The strongest predictor of AI citation is referring domain count — but not in a linear way. Our data shows a clear threshold effect at approximately 32,000 referring domains. Brands below this threshold compete on content quality alone. Brands above it receive a systematic multiplier across all AI engines.

This makes intuitive sense: AI models are trained on web data, and referring domains are a proxy for how "present" a brand is across the web. A brand mentioned on 50,000 sites has left a much deeper imprint in training data than one on 5,000. The good news is that this is a lagging signal — you're building toward it, not waiting for an algorithm update.

Practical implication: focus link building on breadth (unique domains), not depth. One link from each of 100 relevant sites outperforms 100 links from a single domain. Guest posting, community participation, tool creation, and original research distribution are the highest-leverage tactics.

Brand Search Volume: The Signal No One Talks About

The second strongest predictor (correlation 0.334) is monthly brand search volume — how often people search for your brand name directly. This was one of our most surprising findings. We initially dismissed it as a confounding variable (bigger brands are searched more and also get cited more). But even controlling for company size and referring domains, brand search volume remained independently predictive.

Our hypothesis: AI models use brand query frequency as a signal of real-world relevance. A brand people actively search for is a brand people know exists. This creates an important feedback loop: improving AI visibility increases brand searches, which increases AI citation probability.

Content Freshness: Different Rules for Different Engines

Content freshness matters — but not equally across all AI engines. Perplexity, which relies heavily on real-time web search, shows the most extreme freshness preference: 76.4% of its top citations were published or significantly updated within the last 30 days. Content older than 90 days has less than a 12% chance of being cited by Perplexity, regardless of quality.

Practical implication: for Perplexity and Google AI, publish frequently — even shorter, well-structured pieces outperform long evergreen content. For ChatGPT, invest in depth and authority. The dateModified field in your JSON-LD schema is the critical machine-readable freshness signal — update it whenever you revise content.

Content Structure: What AI Crawlers Actually Read

Content structure has a measurable and reproducible effect on AI citation rates. Pages using proper H2 → H3 → bullet hierarchy receive 40% more AI citations than equivalent content in paragraph form. This isn't about keyword placement — it's about giving AI models clean parse trees.

A striking finding: 44.2% of citations reference content from the first 30% of the page. AI engines, particularly when operating under token constraints, prioritize content that appears early. Put your most citation-worthy facts, statistics, and conclusions at the top — not at the end.

FAQPage schema markup deserves special attention for ChatGPT: we measured a 13× odds ratio for citation when FAQPage schema is present vs. absent. The mechanism appears to be that ChatGPT's browsing layer can directly extract structured Q&A pairs, reducing hallucination risk and increasing the model's confidence in recommending the source.

What Doesn't Work (And Why People Still Do It)

Our most counterintuitive finding: press releases and media coverage have near-zero correlation (0.07) with AI citation probability. Companies spending thousands on PR distribution expecting to improve AI visibility are largely wasting their budget.

The reason is structural: AI models are trained primarily on web content with high domain authority and engagement signals. Press release syndication sites (PRNewswire, BusinessWire) have low referring domain diversity relative to their content volume — they're essentially duplicate content farms from the AI model's perspective.

Similarly, traditional backlink building to individual URLs shows surprisingly weak correlation. A single high-authority article with 50 links from diverse domains will outperform 500 links concentrated in 5 sources. The diversity signal matters more than the volume signal.

Finally: AI citation lists are not stable. A SparkToro study found less than 1% overlap in brand recommendations between two identical queries run minutes apart. This means there is no "rank 1" in AI — there is only citation probability. Pheme tracks this probabilistically, averaging citations across hundreds of runs to give you a stable visibility score.

Practical Recommendations

Based on our analysis, here's the highest-leverage roadmap for improving AI visibility, ordered by impact per effort: