How to Track Brand Mentions in Claude

Tracking brand mentions in Claude means running a designed prompt set on a fixed cadence, recording whether the model names your brand, how it frames you, and — when web search fires — which of your pages it cites. The wrinkle that distinguishes Claude from always-searching engines: you must track two answer modes, because a response grounded in live retrieval and one generated purely from training data can describe your brand very differently, and both reach real users.

What exactly are you measuring?

Four fields per prompt run. Presence: did the answer name your brand at all? Framing: recommended, listed among options, mentioned with caveats, or criticized — plus the literal descriptor sentence used. Grounding: did Claude search the web for this answer, and if so, were any citations yours? Competitive context: which other brands appeared, and in what order. The grounding field is the Claude-specific one. A brand strongly present in searched answers but absent from non-searched ones has a training-data gap; the reverse pattern means your live content is underperforming the model's priors.

How do you design the prompt set?

Write prompts the way your buyers actually talk to an assistant — full sentences with context, not keywords. Cover four intents:

Aim for 50-150 prompts. Include factual prompts even though they feel trivial — they surface stale training-data claims (old pricing, discontinued features) that quietly misinform every user who asks. Version the prompt set and keep it stable across periods; edits reset your trend lines.

What cadence and sampling discipline do you need?

Sample each prompt 3-5 times per period, weekly for the competitive core. Claude's responses vary across runs by design, and its conditional search means grounding itself is a random variable — single samples produce trend charts made of noise. Record raw answer text, not just scores, so you can re-code sentiment definitions later without losing history. When Anthropic ships a model update, annotate the date on your dashboards: step-changes in mention patterns usually trace to model or index changes, not your content, and unannotated dashboards get misread.

How do you score sentiment and position?

Keep the rubric small enough to apply consistently: position (first recommendation / listed / trailing mention / absent) and valence (endorsed / neutral / hedged / negative), plus a free-text field for the descriptor phrase. The descriptor archive becomes surprisingly valuable — Claude tends toward careful, qualified language, so shifts like "a popular option" becoming "a popular option, though some users report slow support" are early warnings from the corroboration layer (reviews, forums) that fed it.

When do you automate?

Immediately, if the prompt set exceeds a few dozen. Manual weekly sampling of 100 prompts at 3 samples each is 300 conversations — nobody sustains that. Anthropic's web search API tool (May 2025) makes scripted batteries with parseable citations straightforward, and purpose-built tracking such as Menra's AI visibility monitoring runs the sampling, scoring, and trending across Claude and the other engines from one prompt set. The methodology — set design, sampling math, scoring hygiene — is covered engine-agnostically in our AI mention tracking guide; Claude just adds the grounded-versus-parametric split to the analysis.