Skip to main content

​
What is a model in Attensira?

A model is an assistant surface that answers your prompts — ChatGPT, Perplexity, Claude, Gemini, Google AI. In code and on the wire it is called a channel, and the channel id (chatgpt, perplexity, google_ai, claude, gemini) is deliberately stable. When a vendor swaps the underlying model behind their product, the channel id does not change, so your history stays continuous instead of splitting into a new series every release. That stability is the point of the concept. You are tracking the surface a buyer actually uses, not a specific model version they will never see the name of.
This is not the same thing as an agent model — the LLM that a chat inside Attensira runs on, listed at GET /v2/models. Agent models power the assistant you talk to. Model channels are the surfaces Attensira measures.

​
Which models does a new workspace track?

A new workspace starts on two: ChatGPT and Perplexity. That is the default channel set, and it is a starting configuration rather than a limit. You add the others yourself. Attensira does not offer five models to everyone. The readable set is defined in the engine, and it is not the same as the list of assistants that exist in the world: Google AI has no API reader. Whatever coverage exists for it comes from consumer reading only, and you should expect thinner, less regular data there than on the API-backed channels. Say this out loud in any report you build on it.

​
Are models gated by my plan?

No. Every plan can track every readable channel. What your plan governs is prompt slots and sampling depth, not which surfaces you are allowed to ask. The real constraint is the response budget. Asking more models multiplies the work: every model you add multiplies your prompt count by one more daily reading per prompt, per country, at your plan’s sampling depth. Five channels across 50 prompts at n=3 is a very different volume of readings than two channels at n=1 — and the budget is spent proportionally. You choose the allocation. Two common shapes:
  • Wide and shallow. Track four or five channels across a smaller set of prompts. Good when you do not yet know where your buyers are and want to find out.
  • Narrow and deep. Track two channels across many prompts. Good when you already know one surface drives your pipeline and you want statistically usable rates on a wide question set.
Credits and the response budget are currently inconsistent in the product — a known open bug we are fixing. Because of that, this page does not publish a table of how many readings a plan’s grant buys, and we make no promise that monitoring is free of credits. See Credits for what a credit is and what spends it.

​
What do “tracked” and “readable” mean?

Every per-model metric Attensira returns carries two flags, {tracked, readable}, and reading them wrong is the most common way to misreport a result. tracked: false means the channel is not configured for this workspace. It was never queried. There is no answer, because no question was asked. Rendering that as 0% invents a failure that did not happen — it says a model declined to name you, when in fact you never asked it. readable: false means no reader exists for that surface at all. This is a capability of the platform, not a setting in your workspace, and no amount of configuration turns it on. google_ai over the API is the live example. Both are distinct from a measured zero. Put together with the number shapes: This bites hardest in per-model breakdowns and in exports, where an untracked channel and a channel that never named you look identical unless you check the flags. If you are building a dashboard on get_analytics, branch on tracked and readable before you format a percentage. See MCP overview.

​
How do I change which models I track?

Change the channel set in your workspace settings. A newly added channel starts producing data from its next scheduled run and does not backfill, so its history begins the day you add it, exactly like a new prompt. Each prompt × model × country combination runs at most once per workspace-local calendar day, at your plan’s sampling depth — n=1 on Starter, n=3 on Growth and Business, issued concurrently. Failed runs are stored but excluded from n, so a channel having a bad day shows up as a smaller n, not as a drop in your rate. Turning a channel off stops future runs. It does not delete the readings already collected, which remain subject to your plan’s retention window. Turning it back on later resumes collection but leaves a gap in the series for the period it was off — and during that gap the metric is tracked: false, not zero.
Before adding a channel across the board, add it and wait two weeks. Rates on a new surface are noisy until n builds, and the two-proportion z-test at 95% that guards deltas will correctly refuse to call a change until it has the evidence.

​
A model is not a crawler

Models and AI crawlers are different systems and Attensira reports them separately. A model answers a prompt. A crawler fetches pages from your site. A high crawl rate does not produce mentions, and a model can name you having never crawled you. The two also fail differently, and crawler traffic cannot be measured the way web analytics measures humans: AI crawlers do not execute JavaScript. A client-side beacon, however it is written, has never seen one and never will — by the time the script runs, the only visitors left are humans with browsers. See AI traffic.