The work
Chief of staff — the one agent you talk to. It answers you, proposes goals and projects, hands work to the right crew, makes sure it gets done, and reports back. It does not do the work itself. See How we work. Project — an area of work inside a workspace, with a crew of agents that keeps working on it: your blog, your site’s technical health, the comparison questions. It has a short name, a focus, a feed of what happened and why, and optionally a goal, a deadline and a monthly cap. See Projects and Project and workspace below. Crew — the agents staffed on a project (Chief of staff, Researcher, Writer, Web engineer and others), each with what wakes it. You describe the work; the Chief of staff staffs it. You never define an agent. Deliverable — a finished piece of work: an article, a page, a fix, a social post, a listing fix, product copy or an image. Deliverables land in your own systems and are listed on the Deliverables page, with where each one lives and what it did. Articles, page fixes and outreach pitches are what a budget commits us to deliver. Goal — an objectively measurable target: a metric, where it is read from, a baseline, a target and a date. Each goal is read against a pace line and is ahead, on pace, behind, or unmeasured when its source cannot be read. The first goal is usually to be seen at all: named in a number of your prompts, on the assistants you chose. A goal reports where a number stands; it never claims the number moved because of us. Goals are under Analytics → Goals. Inbox — where the agents put what needs you. Every item that asks something of you is one of three kinds:- Approve: a yes or no on something the agents made. A draft can be edited before you approve it, and a decline asks why.
- Answer: a fact only you have, given in the row.
- Do: something only you can do, such as connecting an integration or sending a drafted email from your mailbox, then marked done.
The tracked objects
Prompt — one question, phrased the way a buyer would phrase it. Your plan sets how many you can track at once: 50, 150, 350 or 750 slots. Topic — a named group of related prompts, so you can read a whole buying question as one line instead of twenty. Model — a model surface Attensira asks, identified by a stable channel id:chatgpt, perplexity, google_ai, claude, gemini. The channel is deliberately stable, so when the vendor swaps the underlying model behind that surface, your history stays continuous.
Workspace — the container for prompts, competitors, models, automations, and data. In the dashboard a workspace is called a brand: the brand switcher at the top of the sidebar picks which one every screen is about, and Add a brand creates another. A credential names your organisation and one workspace as its default; every MCP tool and REST endpoint takes an optional project to name another one in the same organisation.
Answer — one stored response from one Model to one Prompt on one day. Answers are the raw material every metric is folded from.
What we read out of an answer
Mention — the answer named your brand in its text, with no link attached. Citation — the model attached one URL, and that URL points at your own domain. Source — a domain bucket across citations. Sources cover every domain the models cite for your prompts, not only yours, so you can see which sites the answers lean on. Mentions and citations are counted independently. An answer can do either, both, or neither.The metrics
Every rate is
{value, n}. n is the number of successful runs behind it; failed runs are stored but excluded. value: null means not measured. value: 0 with n > 0 is a real, measured zero.{value, real} and pass a two-proportion z-test at 95%. Below the noise floor you get {real: false, value: null}: no change we can prove, which is not the same as no change.
Per-model figures carry tracked and readable. tracked: false means the model is not configured for this workspace and was never asked — reporting it as 0% invents a failure. readable: false means no reader exists for that surface.
The machinery underneath
Credit — the unit work is metered in: one credit is one cent at list price. Adding prompts, running an automation, and asking the agent all spend credits. Reading numbers you already have does not. See Credits. Automation — a saved piece of recurring work an agent performs on its own. A project’s crew works through automations: one for each thing that wakes a crew member. Chat — a one-off conversation with the Chief of staff about your workspace. Ongoing work belongs in a project. Agent model — the LLM your chat runs on. Attensira picks it for each turn; there is no model picker. This is a different thing from a Model: an agent model answers you, a Model is a surface Attensira measures.Project and workspace
The two words are easy to mix up, and the wire makes it worse, so here is the rule. A workspace is everything Attensira knows and does for one site: its prompts, competitors, models, connections, data, chats and automations, and its credit limits. Most organisations have one; an agency has one per client. A project lives inside one workspace and is one piece of ongoing work in it, such as “win the comparison questions” or “keep the docs cited”. It reads the workspace’s data and works under the workspace’s limits; it adds a crew, a feed, a focus and, optionally, a goal, a deadline, a monthly cap and an approval mode that can be stricter than the workspace’s. A workspace can have many projects, or none. On the wire,project has always meant the workspace: it is the project argument every MCP tool and REST endpoint takes, and project_id in identifiers. A project in the sense above is a work_area_id.
Words to wire fields
The wire format kept its original names, and this is deliberate: renaming fields would break every existing integration, so the product vocabulary changed and the contract did not. When the docs and the JSON disagree on a name, both are correct.
The same holds for enum values. The
get_analytics area enum accepts visibility, prompt, page, source, competitor, topic, platform, and bot. Those strings are the contract; pass them exactly as written. The older words group and engine were retired along with channel and are no longer accepted.
A platform is a stable identifier for an AI answer surface — ChatGPT, Perplexity, Google AI Overviews, Claude, Gemini. It is deliberately not the name of a model: the key survives the underlying model being upgraded, replaced or renamed, so two windows a year apart stay comparable.