Agent
The Agent is a consultant inside Trakkr that has already read your brand's visibility data. It answers questions about your numbers, investigates why they moved, and proposes changes you approve with a click. It writes the recommendation. You decide whether it ships.
What separates it from a general chatbot in another tab is what it can see. Every turn starts from a fresh snapshot of your brand: the prompts you track, where you rank on each one, who appears beside you, which pages get cited, what your perception data says, which crawlers hit your site this week. Ask why Perplexity dropped and it looks instead of guessing.
The full workspace is at /agent, with threads, routines and documents in a left rail. The same conversation is also a drawer you can open from any page with ⇧⌘K (Ctrl+Shift+K on Windows), and rows across Prompts, Citations, Competitors, Pages and Crawlers carry an Ask Agent action that opens the drawer with that row already attached. The Agent needs a paid plan.
What the Agent knows
Most of its context arrives before you type anything.
- Visibility scores and trends across all eight models, per prompt and overall
- Tracked prompts, their rankings, and the competitors that appeared beside you
- Citation sources, citation gaps, and which of your pages get cited
- Perception scores and the narratives models tell about you and each rival
- Crawler activity, plus AI visitor traffic when Google Analytics is connected
- Your Actions queue, connected sites, audit findings and brand profile
- What changed since the last report, so it can volunteer "you just completed this"
It can also read Trakkr's own documentation. Ask what a metric means or how a score is calculated and it searches these Learn docs, then answers with a link you can come back to instead of guessing from training data.
The rest you teach it. The brain icon in the top right opens a drawer with three tabs: Memory (headed "What I know"), Approvals, and Connections. Memory holds the facts you have told it: "we only sell B2B", "our real rival is Hoka, not Nike", "do not recommend FAQ schema, we tested it". The Agent reads every item at the top of every conversation and treats them as binding, so a recommendation that contradicts one never gets made. The panel also shows an understanding score, a rough gauge of how well briefed it is.
You can teach it mid-conversation too. Say "remember we only sell to companies with 50 or more employees" and a suggestion appears. Confirm it and every future conversation on that brand knows.
Answer, investigate, execute
Answer
Direct lookups against the snapshot. "What is my visibility on Claude?" "Which prompts am I losing?" "How many actions are pending?" These come back in one tool call, usually as a sentence with the number in it.
Investigate
Multi-step work that combines the snapshot with live page fetches and, when the question reaches outside your account, web search. "Why am I losing this prompt" can pull the prompt's history, fetch the competitor page beating you, fetch your equivalent, and name the structural difference. Tool calls stream into the message as they run, and you can expand any of them to read the raw output. That is how you check a claim came from a real lookup rather than from prose.
Execute
When the Agent recommends a change, it attaches a suggestion to that message. Confirm it and the change happens inside Trakkr. Dismiss it and nothing does.
The kinds it can propose cover most of the product: track or untrack prompts singly or in bulk, add or remove a competitor, remember or forget a fact, queue or complete or snooze or dismiss or pin an Action, apply or roll back a page fix, connect a platform, create or pause or resume an automation, tag and retag prompts, generate an article, run a live answer check, shortlist a placement, update an outreach record, export a data cut, build a document, or open the right screen in Trakkr.
Suggestions are how you approve
A suggestion is bound to the message that proposed it, and it is the only way chat changes anything.
| Card state | What you see |
|---|---|
| Awaiting | The verb button and Dismiss |
| Confirmed | A tick, the past tense, and Undo where the kind is reversible |
| Dismissed | A quiet terminal line you can restore |
| Failed | The error inline, with Retry |
Page fixes show the change before it ships: the old value, the new value, the URL. After you confirm, Trakkr applies it and then verifies it landed.
Most kinds undo cleanly. A few cannot. Generating an article reserves an article credit, an answer check spends real model calls, a report or pitch spends compute, and a sent email is sent. The card says so before you confirm.
Once you have approved the same kind of work several times without editing it, the card offers to stop asking, with Review in Approvals and Keep asking. Choosing to review only opens the Approvals tab, because opening a panel is not itself a permission change.
Threads and history
Every conversation is a thread, saved to the brand you are in. The rail's Threads tab groups them by recency with pinned ones on top, and each row can be renamed in place, pinned, or deleted from its context menu. A dot marks a thread holding a decision you have not answered. Search filters the list.
⇧⌘O starts a new thread, as does the plus in the header. Start fresh when the topic changes hard: it is cheaper for the Agent to begin clean than to drag an old investigation into a new question. Brand memory carries across threads. Conversation history does not.
With no thread open, the workspace shows a short reading of where the brand stands and a ranked list of moves, each with a rough effort estimate, plus Show 5 more ways in for the evergreen questions. Those rows come from your current data, so "Investigate ChatGPT drop" only appears when there is a drop.
Routines
A routine is an automation with a schedule. The rail's Routines tab lists the scheduled ones with their cadence and next run, and a dot when one is paused or blocked. Event-driven automations are not routines and are not listed here, because they have no cadence to show.
There is no second creation screen. Ask in chat for what you want watched and the Agent replies with a suggestion to create it. Running one early works the same way: the row menu writes the ask, you confirm, and the answer lands in a thread you can read later. Pause and resume act immediately from that menu. Cadence, delivery, check history and permission live on the automation record, reached through Manage in Automations.
How much a routine may do without asking is set per area of work on that record, and rolled up for the whole brand in the drawer's Approvals tab. Permissions covers the ladder in full.
Documents and email drafts
Ask for a monthly client report, an executive summary, a competitor briefing, an internal memo or a board update and the Agent builds it as a document on a canvas beside the conversation, section by section, with its figures drawn from your data. You can rewrite one section at a time, reorder or remove sections, refresh a section against newer numbers, compare versions, copy the whole thing as text, or share it on a public link.
Documents outlive the thread that made them. The rail shows the ones you are working on. The full library is at /agent/documents, filtered by All, Reports, Memos or Briefings, with archive and restore.
Emails the Agent composes work the same way. The card in chat is where you read and edit one, and /agent/drafts is the list of all of them, filtered by Drafts, Opened, Sent and Discarded. Trakkr does not send mail for you; you hand a finished draft to your own mail client.
Using Trakkr from another assistant
The Agent is Trakkr's own surface. If you would rather ask from ChatGPT, Claude, Cursor or another assistant, connect Trakkr's MCP server and query the same data there. It is available on every paid plan. See MCP.
Common questions
Is the Agent just ChatGPT with a different name?
No. A language model does the reading and writing, but the brand snapshot, the tools, the approval flow and the memory are Trakkr's. Ask a general chatbot why your Perplexity score fell and it can only reason about what you paste. The Agent looks it up.
Can it change my site or my account without asking?
Not from chat. Every chat change needs a confirmed suggestion. A standing automation can act on its own only at the top of the permission ladder, and only within the scope, revert rules and weekly limit you saved. No setting at any level lets the Agent change your plan, your billing, your API keys, or another brand.
Why does it refuse some questions?
It is scoped to your brand's AI visibility. Off-topic messages, empty ones, and attempts to override its role get a short redirect rather than a speculative investigation. Account and billing questions get a plain answer plus a link to the right Settings tab, never an action.
Why is the answer sometimes generic?
Either the question had nothing concrete to grab (a prompt, page, competitor, model or metric), or the snapshot is thin because your brand profile and memory are sparse. Naming a real thing in the question and filling in the memory panel both raise answer quality fast.
Does talking to the Agent use up my prompt allowance?
No. Conversations are separate from tracked prompts. The Agent actions that spend something say so on the card first: an article credit, a live answer check, or an agency report.
What if I taught it something wrong?
Open the memory panel and delete the item, or say so in chat: "forget that we are B2B only, we launched a consumer tier". A suggestion appears, you confirm, and the memory updates.