OpenAI's usage dashboard tells you the project consumed 14 million tokens last month. It doesn't tell you which chatbot, which page, which user, or which conversation pattern drove that volume. Anthropic's console is the same. Google's console is the same. The provider has zero visibility into your WordPress setup, so the breakdown stops at API key granularity, which is almost never the granularity you need to make decisions.

SleekAI captures token counts at the same moment the provider returns them and writes them into the conversation log alongside WordPress-specific metadata: which chatbot fired, which template page the visitor was on, which post type, which user role, which display condition matched. The admin dashboard slices token usage by any of these dimensions. Suddenly the question 'which page is eating my tokens' has a one-click answer instead of being a multi-hour spreadsheet investigation.

Generic chatbot tools either don't expose token counts at all, or expose them in a single aggregate that's worse than the provider dashboard because it lacks even the raw split between input and output tokens. SleekAI logs both, calculates the ratio, and shows that long-context bots with short replies tip heavily toward input cost while reasoning-heavy bots tip toward output cost. Two different optimization paths, only visible with the right granularity.

You cannot optimize what you cannot measure. Most teams running WordPress chatbots have zero visibility into which bot, which page, or which prompt pattern consumes the bulk of their token budget. They tweak the system instruction blindly and hope cost goes down next month.

Sometimes it does, often it doesn't, and either way no one understands why. With granular token reporting, the optimization process becomes a tight loop. Change the system prompt, watch the token average per turn shift in the dashboard within hours, decide whether the change paid off, iterate.

The prompt-caching opportunity is the clearest example. Anthropic and a few other providers offer aggressive caching on static portions of the input. If 80% of your input is a long fixed policy section, caching it cuts input cost roughly 90% after the first call in a session.

Without token reporting, this opportunity is invisible. With it, the dashboard literally surfaces a 'high cache-ratio candidate' tag on bots that fit the pattern. The same logic applies to model right-sizing, conversation length tuning, variable injection scope, and dozens of other smaller decisions.

Granular data turns chatbot operation from gut feel into a discipline.