LLM SEO: Track What AI Says About Your Brand | Bernard Huang

With Bernard Huang, Founder of Clearscope and Mushi Labs · Streamed live · 64 min

What this episode covers

Bernard Huang explains how to track what AI assistants say about a brand by sampling prompts built from query fan-outs, since answers vary from run to run. He argues that temperature-zero runs and search-grounded prompts give more repeatable baselines, and that informational answers rarely change while commercial and fresh prompts do. His main conclusion is that brands should test each model separately and focus on the pages and third-party mentions AI models validate against.

Key takeaways

  • ✦Build AI visibility prompts by sampling a seed keyword about 20 times and turning the clustered query fan-outs into prompts.
  • ✦Track each model separately, since Gemini, ChatGPT and Claude cite differently.
  • ✦Run prompts with Google search grounding enabled, because trackers that skip search mostly sample training data.
  • ✦Make pricing, use case, feature and industry pages clear, since these are what AI assistants most often cite for Clearscope.
  • ✦Avoid bought links, which the description says led to a manual Google penalty for an early client.

Key moments

  1. 15:56Why Bought Links Backfired
    “the shortcut to rankings is to buy links.”
  2. 32:18Temperature Zero for Stable Answers
    “running a prompt at temperature zero gets you, in our opinion, like the closest to what the training data looks like”
  3. 36:50Query Fan-Outs Drive Validation Searches
    “I want to prevent inaccuracies and hallucinations, so I will search.”
  4. 39:09Gemini Suppresses Self-Promotional Content
    “if we at Clearscope are writing content on what are the best SEO tools, Gemini is actively suppressing that.”
  5. 45:10Structured Data for AI Agents
    “the best thing is either a markdown file or JSON file for your data attributes.”
  6. 55:34Sample Each Prompt 20 Times
    “you take a prompt and then you run it 20 times.”

Quotes are from the episode transcript, lightly cleaned of auto-caption errors.

Questions answered in this episode

How do you track what AI says about your brand?
Run each prompt many times, around 20, and measure how often brands, query fan-outs and citations appear, since single answers vary. Bernard says the frequencies over a data set become stable enough to use as a baseline.
Why do some AI prompt trackers give misleading results?
Bernard says many trackers run prompts without Google search grounding, so they only sample training data. He says this makes the results unlike what a user would actually see when the AI searches.
How do you decide which prompts to track?
Sample a keyword about 20 times in Gemini to collect its query fan-outs, cluster them into topics, and turn those topics into prompts. Bernard says he weights prompts by how often each fan-out appears.

More episodes on AI Search & LLMs