Kevin Sullivan

AI search research

Studying how search systems understand and recommend information.

AI search covers the tools that summarize, cite, and recommend information instead of only ranking links—Google AI Overviews, ChatGPT, Claude, Gemini, Perplexity, and similar systems. Some of this work also goes by GEO (generative-engine optimization) or AEO (answer-engine optimization). I use those terms when they help match how a team already talks about the problem, but the underlying work is the same: understanding how these systems select, interpret, and present information, then turning that research into practical SEO and content decisions.

The goal is not to turn limited observations into universal rules. It is to ask better questions, build clearer diagnostics, and create more useful implementation priorities.

What I’m investigating

Working questions, not settled answers.

  1. Which sources get cited, summarized, or omitted for a given brand or topic, and why?
  2. How do entity signals and information architecture affect whether a brand is recognized and referenced correctly?
  3. What differs in inclusion, framing, and attribution across Google AI Overviews, ChatGPT, Claude, Gemini, and Perplexity?
  4. Which structural or content changes correspond with a measurable difference in later observations?

Featured AI-search projects

Where this research turns into work.

Three projects built to move from generative-search observations to decisions a team can act on.

01

Methodology

AI Visibility Diagnostic

A repeatable method for investigating how brands are found, interpreted, cited, and recommended across generative search—and translating those observations into implementation priorities.

  • Cross-platform analysis
  • Entity interpretation
  • Prioritized action
02

Operating system

AI-Assisted SEO Workflows

A human-in-the-loop system for standardizing recurring SEO analysis, improving consistency, and protecting the strategic judgment automation cannot replace.

  • Repeatable workflows
  • Human review gates
  • Scalable delivery
03

Experimental build

Turf Network

An industry resource and research environment used to test entity architecture, structured information, internal relationships, and AI-search retrieval.

  • 2,500+ organizations
  • 142 entity pages
  • Architecture experiments
See all work

Observation, hypothesis, limitation

Keeping what I’ve seen separate from what it means.

Generative-search analysis produces a lot of individual observations. The useful part is being disciplined about which of those observations are settled, which are working theories, and where the evidence runs out.

Observation

Patterns in which sources are cited, summarized, or omitted across generative-search responses—reviewed across hundreds of individual responses.

Hypothesis

A working explanation for why entity clarity and source structure appear to influence inclusion—treated as a hypothesis, not a proven ranking factor.

Limitation

The available evidence does not establish causality. Platforms change frequently, and sample sizes are small relative to the scale of these systems.

Research principles and limits

How I keep this research honest.

  • Observations are treated as data points, not established ranking factors.
  • Findings are described with the uncertainty they deserve—“observed,” “suggested,” or “a working hypothesis,” not “proven.”
  • Third-party estimates, when referenced, are always labeled SEMrush-estimated organic traffic—never presented as verified analytics.
  • Sample sizes and platform coverage are stated plainly so a reader can judge how far to generalize.
  • The goal is better questions and clearer diagnostics, not universal rules.

Next conversation

Hiring for AI search, GEO, or AEO?

I’m interested in senior SEO, AI-search, organic strategy, and search-focused consulting opportunities.

kevin@kevinsullivan.io