Which LLMs should I track for brand visibility?
Which LLMs to track for brand visibility, how we choose the engine set for clients, and why real-UI monitoring and API tracking are not the same thing.
Clients ask which LLMs they should track, usually expecting a list of ten. We start with fewer. The right set depends on where your buyers look, and the goal is a stable panel you can compare month to month, not maximum coverage.
Our default starting set
For most programs we begin with ChatGPT (Search), Google AI Overviews, Google AI Mode, Perplexity and Gemini. Then we add Microsoft Copilot where the audience is enterprise, and Claude where buyers are developers or analysts. Regional or niche engines come later, only if the client's own data shows demand.
We avoid quoting market-share statistics to clients. They change quickly and many are not well sourced. Instead we check the client's own analytics. Promptwatch's visitor analytics shows AI-referred traffic and conversions by source, including ChatGPT, Perplexity, Gemini, Claude, Copilot, DeepSeek, Kimi and Mistral, which gives a real picture of where visits come from.
Real interface versus API
This distinction decides whether a number means what you think it means. Real-UI monitoring captures what a person would see in the live product, with search and formatting included. API tracking queries the model directly. Both are useful, but they are not interchangeable, and a client deserves to know which they are looking at.
The platform we run client programs on is Promptwatch, and it is explicit about this. Real-UI monitoring covers ChatGPT (Search), Google AI Overviews, Google AI Mode, Microsoft Copilot, Gemini, Perplexity, Amazon Alexa and Meta AI, plus the coding agents Codex and Claude Code. Claude (Haiku 5.5), DeepSeek, Mistral (Large), Grok 4.5, GPT 5.6 Luna, GPT OSS 120B, Kimi K2.7 Code, GLM 5.2, Inkling and OpenCode are tracked via API. We tell clients that Claude and Grok results are API results. Every paid plan includes all active models.
How we set up the panel
- Write the prompt set first: topic, persona and country.
- Run it against the starting engines daily. Promptwatch runs prompt panels daily per model, country and language.
- Keep the set stable for at least a month before reading trends.
- Look at share of voice per engine, not an average across all of them. An average hides the engine that matters.
- Compare against the competitors the client actually loses deals to.
When to add a model
We add an engine when one of three things happens: client analytics shows visits from it, sales hears prospects mention it, or the category is one where that engine is known to be used. Coding agents are a special case. If the client sells developer tools, Promptwatch's tracking of what Codex, Claude Code and OpenCode recommend is worth including.
Why crawler data belongs next to model coverage
If one engine names a competitor and not your client, the cause could be content, corroboration or access. Promptwatch's IP-verified crawler logs show which bots actually reach the site, so you can rule access in or out. Logs start on the Professional plan. Without them you are guessing.
What we avoid promising
We do not tell clients that tracking more models means more visibility. We do not claim any engine will name a brand. And we do not compare a real-UI number with an API number as if they were the same measurement.
The short answer
Track the live-search engines your buyers use first, add Claude, Copilot and others by audience, mark API-tracked models clearly, and keep the panel stable so trends mean something. Then read each engine separately and connect the findings to crawler access, citations and visits.
We revisit the engine list every quarter and write down why each engine is in or out.