Is this a chatbot on top of a dashboard?

No. It reads the same measurements the dashboard shows, but it can also act: add prompts, add a competitor, create or run an automation, and open a pull request against a connected repository. The chat is the interface; the work is the product.

What stops the agent doing something I did not want?

Three documented bounds. It can only use tools you have granted, and a grant is per workspace and revocable. Anything irreversible stops for your explicit yes, and an approval left unanswered for 7 days closes the run having changed nothing. And where it writes to a repository it opens a pull request through a first-party GitHub App — it never merges, and never pushes twice to the same branch.

Does changing the agent model change my numbers?

No. Every agent model reads the same stored runs. A different model may phrase an answer differently or take a different route through your data, but the underlying measurements are identical. If a figure looks wrong, check the sample size and whether the model was tracked, not the selector.

Why does a chart have gaps in it?

Because a day nobody sampled is not a day you scored zero. Attensira draws an unmeasured day as a break in the line and puts the number of unmeasured days in the chart caption. Sampling depth is n=1 on Starter and n=3 on Growth and Business, and each prompt is measured at most once per model per country per day.

Can I get the underlying rows out?

Yes. Every table and chart carries Copy and Download. Copy gives you tab-separated rows that paste into Sheets or Excel; Download gives you a CSV of the rows behind a chart rather than an image of it.

Does asking the agent cost anything?

Each turn is metered agent work and draws on your workspace's credit balance, because it runs a model and may run tools for you. Opening a dashboard, filtering a chart or exporting numbers you already collected does not spend credits.

The agent

An agent that works your AI visibility, inside bounds you set

The Attensira agent reads your tracked prompts, models, answers, mentions and citations, works out what changed and why, and does something about it. It is scoped to one workspace, it can only use tools you have granted, and anything irreversible stops for your explicit yes.

Get Started Book A Demo

Ask in plain language
“Which prompts lost citations this month?” — answered from your own tracked runs.
Scoped to one workspace
It cannot read another workspace's data, and cannot reach anything you have not tracked.
It shows its working
Every step it takes is listed. Failures and approvals never collapse out of sight.
Bounded by design
Granted tools only, and an explicit yes before anything irreversible.

By Karl-Gustav KallasmaaLast updated September 3, 2026Behaviour on this page is checked against docs.attensira.com

What the agent is

The agent is a conversation with your own workspace. You ask a question in plain language, and it queries the prompts, topics, models, answers, mentions and citations you already track, then answers with the numbers behind it. Each chat is a session you can reopen and continue where you left off.

That is a different product shape from a dashboard. A dashboard hands you charts and leaves the interpretation and the work to you. The agent reads the same measurements, tells you what moved and what it implies, and can then act — adding prompts, adding a competitor, creating or running an automation. You set the destination. It handles the route.

The distinction that matters commercially is not that an AI does work. Everyone selling into this category claims that. It is that this agent is bounded in ways you can inspect, revoke and audit — which is the only serious objection to running an autonomous agent against your own website in the first place.

The problem it solves

AI answer engines describe your category to your buyers every day, and you were not in the room. The measurement part of that problem is now well served: plenty of tools will tell you that your mention rate fell. Almost none of them do anything about it.

So the work lands back on a person who already has a job. Somebody has to read the chart, work out which prompts moved, check whether the movement is real or sampling noise, find the pages involved, decide what to change, and then actually change them. That loop is where AI visibility programmes quietly die.

The agent closes that loop. It reads the numbers, applies the same statistical rules the dashboard applies, proposes the work, and — where you have connected a repository — opens a pull request with the change in it. Nothing goes live blind.

How it works

The agent reads exactly what your workspace has measured, and nothing else. If a model is not configured for the workspace it was never queried, so there is no data to summarise, and the agent says so rather than reporting a zero. If a prompt does not exist, no answer about it exists either.

It also inherits the measurement rules, not just the data. Null means not measured. A measured zero means runs happened and found nothing. A delta that does not clear a 95% two-proportion significance test is reported as no change we can prove, not as a small arrow. Ask it for the run count behind a figure whenever a number surprises you.

When it replies, it does not retype your table into prose. It names a component from a fixed set and the app draws it, so the figures you read came out of the measurement rather than out of a sentence.

What it draws	What that is
Metric	One headline number with its movement
Table	The rows behind the number, capped at the top ten
Chart	A trend line over the window, one line per model
List, quote, prose	Ordinary reading blocks
Code, diff, link list	Reference blocks, including a proposed change

Gaps stay gaps, and copies stay data

A day nobody sampled shows as a break in the line, never as a drop to the floor. Drawing an unmeasured day as a zero would invent a collapse that never happened, so the caption under a chart says how many days in the window were not measured instead.

Every table and chart the agent draws carries Copy and Download. Copy puts the rows on your clipboard tab-separated, so they paste straight into a spreadsheet. Download saves a CSV of the rows behind the line — not a picture of the line, because a picture is not something you can pivot. Only the table or chart travels; the agent's reasoning and tool calls stay behind.

Which model answers, and why that does not move your numbers

Two different things get called “model” around this product, and Attensira keeps them apart. A Model with a capital M is a surface you track — ChatGPT, Perplexity, Claude, Gemini, Google AI. An agent model is the LLM running your chat.

You pick the agent model from a selector. Auto is the default and routes each turn to whichever model clears the bar for that question, so a short lookup runs somewhere cheap and a hard analysis runs somewhere capable. Named models from Anthropic, OpenAI, Google and DeepSeek are there when you want a specific voice.

Switching agent models changes how an answer is phrased. It does not change your measurements, because every agent model reads the same stored runs. It is not a way to get a second opinion on your visibility — if a number looks wrong, check the sample size and the tracked flags, not the model selector.

Effort is fixed per model and shown as a tag beside its name, so there is no second dial to get wrong.
Under Auto, billing follows what the call actually cost, never the model id that was sent.
A long question may return a running status and finish in the background; reopen the session to read it.

What keeps it in check

Three bounds, all documented, all inspectable, and all of them things you control rather than settings we default for you.

Bound	What it means in practice
Granted tools only	A granted plugin is the only way a tool enters the agent's toolset. Grants are per workspace and revocable; until you grant one, the agent cannot see those tools at all.
An explicit yes	Anything irreversible — publishing, sending mail, spending with a third party — stops and waits for you. An unanswered approval expires after 7 days and the run closes having changed nothing.
Pull requests, never merges	Where the agent writes to a repository it does so through a first-party GitHub App scoped to repositories you pick on GitHub's side. It opens a pull request, never merges one, and never pushes twice to the same branch.

A concrete run

This is the shape of an ordinary session — a question, the rules being applied honestly, and a stop where a person is required.

You: compare mention rate for the Pricing topic against the previous 7 days, and say which prompts were not measured.

Agent: reads prompts → reads answers → reads citations (steps shown, then collapsed)

Agent: Metric + Table + Chart, with 2 of 7 days marked unmeasured in the caption

Agent: “Two prompts moved. One clears the 95% gate; the other is inside the noise, so it is reported as no change we can prove.”

Agent: proposes a page change → pauses in your Inbox for an explicit yes

Nothing in that sequence is a guess dressed up as a number. The unmeasured days are labelled rather than filled in, the movement that could not be proven is not claimed, and the one action that touches your site waits for a person.

What it does not do

Attensira publishes its limits, because a tool that hides them costs you more later than it saves you now.

It cannot reach anything you have not tracked. An untracked model was never queried, so there is no data — and the agent will say that rather than report a zero.
It cannot read another workspace's data. Chats are scoped to one workspace.
It cannot produce a metric that does not exist. There is no visibility score, no sentiment score and no brand rank, and asking for one returns an honest gap.
It cannot attribute a mention or a citation to a visit, a signup or revenue. Nothing in the product claims that link.
Switching agent models does not give you a second reading of your visibility. The stored runs are identical.
Each turn is agent work and spends credits. Reading numbers you already collected does not.

Which plans include it

Included on every plan — Starter $129/mo, Growth $349/mo, Business $699/mo.

Chats are on every plan. What differs by plan is how much you measure: prompt slots, sampling depth (n=1 on Starter, n=3 on Growth and Business) and how long history is retained.

Plan facts on this page are read from the same ladder that prices /pricing, so they cannot drift from what you are actually charged.

Read further

Documentation

Chats — asking a good question, and which agent model answers

Documentation

Tables, charts and answers — what the agent draws and why gaps are gaps

Automations

Turn a question you keep asking into a standing instruction with a run history.

Inbox

Every stop the agent makes, in one list — and the 7-day approval expiry.

Plugins

The only way a tool enters the agent's toolset, including your own MCP server.

Credits

You pay for work the agent does, not for seats or dashboards.

FAQ

Questions people actually ask

Answers here match the product documentation. Where the product cannot do something, this says so.

Still deciding? Compare plans or talk to us.

An agent that works your AI visibility, inside bounds you set

Ask in plain language

Scoped to one workspace

It shows its working

Bounded by design

What the agent is

The problem it solves

How it works

Gaps stay gaps, and copies stay data

Which model answers, and why that does not move your numbers

What keeps it in check

A concrete run

What it does not do

Which plans include it

Read further

Questions people actually ask

Is this a chatbot on top of a dashboard?

What stops the agent doing something I did not want?

Does changing the agent model change my numbers?

Why does a chart have gaps in it?

Can I get the underlying rows out?

Does asking the agent cost anything?

What the agent is

The problem it solves

How it works

Gaps stay gaps, and copies stay data

Which model answers, and why that does not move your numbers

What keeps it in check

A concrete run

What it does not do

Which plans include it

Read further

Questions people actually ask

Is this a chatbot on top of a dashboard?

What stops the agent doing something I did not want?

Does changing the agent model change my numbers?

Why does a chart have gaps in it?

Can I get the underlying rows out?

Does asking the agent cost anything?