Agent
:::summarybox learn What the Agent actually is, and how it's different from ChatGPT The data and tools it can reach for when you ask a question Good asks and bad asks, with examples What it can do for you: answer, investigate, execute Why chat stays click-to-confirm, and how standing automations stay separate
How it's different from generic AI
It's worth being explicit about what this isn't.
If you only want a creative writing partner, generic chatbots are fine. The Agent has a different purpose: figuring out why your AI visibility looks the way it does, and changing it.
How to ask well
The Agent rewards specific, anchored questions. The more your question names a real thing in your account, the sharper the answer.
Those work because each one ties to something the Agent can investigate or act on: a specific model, a specific competitor, a specific page, a specific prompt to track, a fact to remember.
The questions that produce weak answers tend to fall into a few buckets.
If your question doesn't name a prompt, page, competitor, model, metric, or fact about your brand, the Agent has nothing concrete to grab onto. Generic in, generic out.
What it can do
Three loose tiers, from least to most consequential.
Answer
Direct lookups against your snapshot. "What's my visibility score on Claude?" "Which prompts am I losing on?" "Show me my top citation sources." "How many actions do I have pending?" These come back fast, usually in one tool call, often as a single sentence with the number.
Investigate
Multi-step analysis that combines snapshot data with live page fetches and, when needed, web search. "Why am I losing on this prompt?" might pull the prompt's history, fetch the competitor page that's beating you, fetch your equivalent page, and explain the structural difference. "What's driving my growth this week?" might cross-reference the prompts that moved, the citations that landed, and the perception shift to give you one paragraph instead of three dashboards.
This is where the Agent earns its keep. The answer typically names a real URL, a real competitor, a real page section, and a concrete next move you can ship today.
Execute
When the Agent recommends a specific change, it surfaces a suggestion chip in the conversation. The chip is a draft of the action. You click confirm to apply it, or dismiss to throw it away. Nothing happens to your workspace until you click.
The kinds of action it can propose:
| Kind | Example |
|---|---|
| Track or stop tracking prompts | "Track 'best stability shoes for plantar fasciitis'" |
| Add or remove a competitor | "Add Brooks to your competitor list" |
| Remember or forget a fact | "Remember you're B2B only" |
| Manage an Action | "Add 'rewrite /pricing for clarity' to Actions" or "snooze this until Friday" |
| Apply, dismiss, or revert a Page fix | "Update the meta title on /running-shoes" |
| Connect a platform | "Connect the WordPress site for this Page" |
| Create, pause, or resume an exact rule | "When perception drops more than 5 points, ping me" |
| Create or pause a standing automation | "Check our cited Pages every Monday" |
| Check an automation now | "Check the pricing-page automation now" |
| Generate an article | "Draft an article on 'best running shoes for flat feet'" |
| Update outreach records | "Mark this G2 pitch as contacted" |
| Organize tags | "Tag your 12 French prompts as FR" |
| Build a report or export | "Export the prompts we lost this month" |
| Open something in Trakkr | "Open the prompt detail for 'best CRM for startups'" |
Most kinds support undo. If you confirm a track and change your mind a minute later, the undo button on the chip reverses it cleanly. A few actions can't be undone (article generation spends a credit, outreach sends an email, audits cost compute), and the chip says so before you confirm.
Approvals and safety
Chat and standing automations have separate trust rules.
Standing automations are separate. You turn each one on deliberately. They are off by default, permission is granted per work area, site changes have a weekly limit, and every change records which automation and permission allowed it. Site changes are limited to fix types Trakkr can snapshot and revert.
The safe default remains Suggest only. Higher rungs are rolling out gradually and may not be available for your brand yet. Read Agent permissions for the full contract, including permanent limits on outreach and community replies.
A few specifics worth knowing:
- Chat Page fixes preview before they ship. When the Agent proposes a CMS edit in a conversation, the chip shows you the exact change first: the old text, the new text, the URL. You see it, you decide. After confirming, the change applies and Trakkr verifies it landed.
- Brand memory is binding. If you teach the Agent something via the Understanding panel ("we don't sell to consumers"), it won't propose actions that contradict it. If you ever want it to ignore something, delete the memory item from the panel.
- Off-topic questions get redirected, not answered. Ask the Agent for stock picks or to write a novel and you'll get one polite sentence pointing back to what it can help with. This is intentional: a sharp tool that knows its lane beats a fuzzy one that tries to be everything.
- Account and billing questions get a link, not an action. The Agent can tell you what plan you're on, but it can't upgrade you or rotate your API key. It'll surface the right Settings tab as a navigation chip instead.
If chat claims it changed something before you confirmed a chip, flag it. A standing automation reports through its own check history and Action record, not by pretending a chat message was your approval.
The Understanding panel
The Understanding panel is where you teach the Agent things it couldn't otherwise know. It opens from the brain icon in the top right of the workspace.
What lives there:
- Identity facts: "We're a B2B SaaS in the sales-engagement space."
- Preferences: "Don't recommend SEO actions, we have a separate team for that."
- Topics you care about: "Focus recommendations on enterprise buyers, not SMB."
- Things to ignore: "Hoka isn't a real competitor for us, we sell to runners not casual buyers."
Each item is a single sentence. The Agent reads all of them at the top of every conversation. The panel also shows an understanding score, a rough gauge of how well-briefed the Agent is on your brand. Higher score, sharper answers.
You can also add memory mid-conversation by just telling the Agent. "Remember we don't ship to Canada." A chip appears, you confirm, and from then on every conversation knows that.
Threads and history
Every conversation is a thread, saved to your account. You can rename them, pin the ones you'll come back to, and reopen any of them to keep the conversation going with full context preserved.
A few patterns that work well:
- One thread per investigation. "Why did Perplexity drop in May" is a thread you can return to as you ship fixes and re-check.
- An ongoing competitive thread. Pin a thread for your main competitor and keep adding "what did they ship this week" turns over time.
- Fresh thread for a fresh question. When the topic shifts hard, start a new thread. It's cheaper for the Agent to start clean than to keep dragging old context into a new investigation.
Hit New in the top bar (or click the Briefing chip in the header) to start fresh. The thread sidebar holds everything you've ever asked, scoped to the brand you're in.
Automations from the Agent workspace
The Agent workspace does not embed a second automation manager. On a fresh conversation it shows one quiet line below the composer, such as 3 automations running, with a Manage link to the Automations register. If none exist, Set one up opens the same register.
As standing automations reach your brand, you can also ask in chat to create, pause, resume, or check one. The Agent responds with a suggestion chip, so chat still needs your click. Ongoing permission, delivery, check history, and pause controls live on the automation record.
How it works under the hood
For the curious, the main tool loop uses an OpenAI reasoning model. It starts with the primary model and can move to a lower-cost OpenAI fallback after a rolling spend threshold or when the spend check is unavailable. A separate DeepSeek-compatible model handles auxiliary routing and simple triage; it does not write the main investigated answer.
Every turn starts with a fresh snapshot of your brand, then the main model decides which tools to call: snapshot slicers for fast data lookups, Page investigation tools to fetch real URLs, infrastructure tools to check crawler and delivery health, and web search for anything outside your account. Tool calls stream into the conversation as they happen, so you can watch the investigation unfold.
The current service uses one full-budget pass. The frontend still sends an old depth setting for compatibility, but it does not create separate Quick and Deep answer modes.
Common questions
Is the Agent the same as ChatGPT or Claude?
No. The Agent is a Trakkr-specific system built around language models, tools, your brand snapshot, the approval flow, and the Understanding panel. The model is the language and reasoning layer; the Trakkr data, tools, permissions, and records make it the Agent.
Can the Agent edit my site or my account without asking?
Not from chat. Every chat mutation requires you to confirm a suggestion chip.
A standing automation can make a supported site change only if an editor separately granted that work area Do it, then tell me or Handle it. Those grants are off by default, capped weekly, attributed, limited to reversible fixes, and reported after every change. See Agent permissions. The Agent cannot use any setting to change billing, your plan, your keys, or another brand.
Why does the Agent sometimes refuse to investigate my question?
It's scoped to your brand's AI visibility. Questions outside that lane (general SEO advice, code help, off-topic conversation) get a one-line redirect instead of a wasted investigation. This is intentional, not a limitation you'd want to remove.
Why does it sometimes give a generic answer?
Two usual causes. One, the question itself was too generic to ground in your data ("how can I improve"). Two, the snapshot is thin because your brand profile or memory is sparse. Filling out the Understanding panel and giving the Agent a real anchor in your question both raise answer quality fast.
Does the Agent remember things across conversations?
The brand memory in the Understanding panel persists across every conversation in that brand. Conversation context itself stays within a thread: when you start a new thread, the Agent starts fresh on the conversation but still reads your brand memory.
What if I taught the Agent something wrong and want to fix it?
Open the Understanding panel, find the item, click the trash icon. Or just tell the Agent in a conversation: "forget that we're B2B only, we just launched a consumer tier." A chip appears, you confirm, the memory updates.
Can I see what tools the Agent used to answer me?
Yes. Tool calls stream into the message as they happen, with a small timeline above the answer showing which tools ran. Click any tool row to expand the raw output. This is also how you check whether a claim about your data is grounded in a real lookup.
Does the Agent count against my prompt limit?
No. Conversations with the Agent are separate from your tracked prompts. The only Agent action that consumes a paid resource is generating an article (one article credit) or generating an agency report (PDF credit on agency plans).
Why did the Agent suggest the same thing twice?
Usually because the suggestion was dismissed without action and the situation that triggered it hasn't changed. If a suggestion isn't useful, dismiss it and the Agent stops surfacing it for a while. If you'd rather it never came up again, teach the Understanding panel why.