How to Track Your Brand's Presence in AI Search

Start by writing down the actual questions your buyers ask, not the keywords you'd target in traditional SEO. AI answer engines respond to conversational, often comparative phrasing - "what's a good tool for X," "how does Y compare to Z," "is X worth it for a small team." Semrush's comparison study of Google AI Mode against traditional search and other LLMs points to how differently these interfaces surface results compared to a ranked list of blue links, which is a useful reminder that the prompt itself, not just the topic, shapes what gets cited.

Build your panel in three buckets: category questions (no brand names mentioned), comparison questions (your brand against named competitors), and branded questions (your brand name included). Each bucket answers a different thing - category questions tell you whether you get surfaced at all, comparison questions tell you whether you win the comparison when you're already in the conversation, and branded questions tell you whether the engine gets basic facts about you right. Aim for at least 15-20 prompts per bucket before you call the panel usable; fewer than that and any single week's result is too noisy to act on, for the same reason a small sample size limits what you can conclude, as Nielsen Norman Group's confidence interval reference lays out.

Once the panel is written, freeze it. Don't add or remove prompts between runs unless you're deliberately expanding coverage, and if you do expand, keep the old and new prompts in separate columns so you're never comparing a 20-prompt week against a 35-prompt week and calling the difference a trend.

Once you have a frozen prompt panel, you need a consistent way to read each answer and record two separate facts: did the answer name your brand, and did it link to your brand's page. These are not the same signal. An answer can describe your product accurately without a link, and it can include a link buried in a source list without ever naming you in the prose. Conflating the two into a single "visibility" number hides which problem you're actually solving.

EdenRank records whether an answer names the brand and whether it links to the brand page, and separates mention rate from citation rate on the same prompt set. That separation is the practical reason to track the two facts independently rather than as one blended score: a team chasing a rising mention rate but a flat citation rate needs to fix internal linking and structured data, not top-of-funnel content, and the reverse fix wastes a quarter if you can't tell which number moved.

If you're building this by hand instead, the discipline is the same - a spreadsheet with one row per prompt per run, one column for "named: yes/no," and one column for "linked: yes/no." What EdenRank adds is doing this consistently across a panel without a person re-reading every answer by hand each cycle, but the underlying logic - separate the mention from the citation - is what actually produces a usable number, whether you build it yourself or not.

Before you trust any number this system produces, validate it against a case you already know the answer to. Run your branded prompt bucket and confirm the engine's answers describe your product correctly - if it gets your category wrong or names a discontinued product, that's a data-quality problem with the source material the engine is drawing from, not a tracking bug, and no amount of re-running the panel fixes it. Fix the underlying page first, then re-test.

Watch for undeclared crawlers and blocked access as a separate failure mode. The Hacker News discussion of undeclared crawlers and no-crawl directives is a useful read here: some crawlers used by AI systems don't identify themselves clearly, and a misconfigured robots.txt file can silently block the ones that do, which means your brand could be invisible in an engine's answers not because of content quality but because the page was never fetchable. Google's Search Essentials documents the baseline crawl requirements every page needs to meet before any of this tracking matters.

When a number looks wrong, work through it in order: check whether the prompt wording changed, check whether the panel size changed, check whether the engine itself changed its answer format (some engines rolled out new citation-display behavior without notice), and only after those three are ruled out should you treat the change as a real shift in your brand's presence. TREC's 2024 RAG track test collection is a useful reference for how formally structured evaluation of retrieval-augmented answers actually is - it's a reminder that even research teams built dedicated test collections rather than trusting single-run comparisons, which is the same discipline worth applying here.