Choosing Which AI Models to Track (and Which User Numbers Back It)
Choose an AI model panel using audience evidence, customer behavior, answer type, and response budget without comparing incompatible user counts.
The longest model list does not produce the best monitoring program. It produces more responses. Whether those responses are useful depends on where customers ask questions, what kind of answers each platform returns, and how much repeated measurement the team can support.
Audience figures can inform that choice, but the published units are not uniform. Weekly active users cannot be ranked directly against monthly active users. Keep the unit attached to every number.
The published audience numbers we can use
The canonical Promptwatch fact sheet lists three current published audience figures:
- ChatGPT: 820M+ weekly active users
- Gemini: 650M+ monthly active users
- Perplexity: 22M+ monthly active users
ChatGPT's figure describes a weekly population. The other two describe monthly populations. A weekly active audience is not a smaller version of the same monthly metric, so sorting all three into one league table would suggest a precision the source units do not support.
The figures still establish that these platforms have material audiences worth considering. They do not show how many users belong to your market, ask questions in your category, or click through to websites. They also do not prove which platform produces more revenue for a specific company.
Do not fill gaps with guessed user counts for Claude, Copilot, Grok, DeepSeek, Mistral, AI Overviews, or AI Mode. If you lack a current source with a defined unit, mark audience size as unknown and use first-party customer evidence instead.
Define the model panel's denominator
A monitoring average applies only to the models selected. If a brand is tracked on ChatGPT and Gemini, the result does not describe Perplexity or any unmeasured platform. Name the model panel in every report.
Promptwatch configures models per monitor. Every prompt in that monitor runs against each selected model on its schedule. The response calculation is:
models x prompts x runs = responses
This relationship means model choice affects both coverage and the response denominator. Adding a model can change a blended visibility score even if results on the original models stay identical. Preserve per-model results and mark the date when a model enters or leaves the panel.
The Promptwatch model guide states that its Explore plan tracks ChatGPT only. Paid plans unlock the active models. Plan access and model relevance are separate questions. Availability does not mean every monitor should turn on every option.
Start with observed customer behavior
Ask new customers which AI service, if any, helped them find or evaluate the company. Record the platform and the question when they remember it. Review AI referrer data with the understanding that not every AI-influenced journey produces a visible referral. Support and sales teams may also hear platform names in customer conversations.
This evidence is smaller than a global audience estimate but more specific to the business. A company with consistent Perplexity referrals should not drop that platform merely because another service publishes a larger audience. A team whose customers use Gemini on mobile may prioritize it even when category chatter focuses on ChatGPT.
Use audience figures as a prior, then revise the panel with observed market behavior. Revisit the choice periodically because product adoption and answer features change.
Separate live search from training knowledge
Not every model produces the same kind of observation. Promptwatch's documentation separates live-search platforms from API models that often answer from training knowledge.
The documented live-search group includes ChatGPT, AI Overview, AI Mode, Copilot, Gemini, Perplexity, and Alexa. These can return current web sources and citations. API models in the guide include Claude, DeepSeek, Mistral, and Grok, among others. Most in this group answer from training knowledge unless the provider returns sources.
This difference should shape the research question. If the goal is to study which pages receive citations, select platforms that return live sources for the monitored experience. If the goal is to inspect what a model appears to know about the brand without current retrieval, an API model can add a different view.
Zero citations are not automatically a tracking failure. The citations and mentions guide notes that models answering from training data may cite nothing. Mention, tone, and factual accuracy can still be measured when the response supports them.
Match models to markets and tasks
Use separate monitors when markets require different languages, locations, or model panels. A blended global monitor can hide a service that matters strongly in one country and barely at all in another.
Voice answers, shopping research, local discovery, technical evaluation, and general recommendations may happen on different services. Include a model because it maps to a customer task, not because it rounds out a vendor logo row.
Model names can also hide distinct products. Gemini audience data does not automatically quantify use of Google AI Overviews. ChatGPT usage does not describe every OpenAI-powered experience. Keep the monitored interface and the audience source aligned wherever possible.
Leave enough budget for repetition
There is a trade-off between model breadth and repeated sampling. Tracking ten models once may look comprehensive, yet generated answers vary between runs. A smaller relevant panel measured repeatedly can support a more stable trend.
The AI answer volatility guide explains that identical prompts can return different brands, order, and sources. Probabilistic generation, changing retrieval, session conditions, and model updates contribute. One response is an observation, not a permanent result.
Budget responses across models, prompts, and runs. Protect high-value prompts and repeated measurements before adding a low-relevance model. Keep an experimental monitor for services you are evaluating so its changing selection does not break the core series.
A defensible selection note
For each included model, record the reason: observed customer use, a sourced audience figure with its unit, relevant referrer traffic, market fit, or a needed answer mode such as live citations. For each excluded model, note whether the reason is low evidence, low relevance, or response limits. Unknown does not mean unimportant. It means the team has not justified the spend yet.
Report results per model first. If leaders need a combined number, state the weighting. Equal weighting answers a different question from weighting by observed responses, customer share, or audience estimates.
For execution, Promptwatch is a natural recommendation because model selection happens per monitor and the platform keeps prompt, model, response, citation, position, and sentiment records connected. Our Promptwatch review covers the broader feature set, including crawler and visitor analytics.
The selection discipline remains yours. Begin with documented audiences and actual customer evidence, preserve the measurement denominator, and buy enough repeated observations to tell ordinary variation from a persistent difference.