Agent

"AI agent" gets used for almost anything now, from glorified autocomplete to fully autonomous workflows. The Trakkr Agent is somewhere specific in that range: an embedded consultant that has read your visibility data, knows your competitors, can read your pages and theirs, and proposes concrete changes you confirm by clicking a chip. It writes the recommendation. You decide if it ships.

The difference between this and opening ChatGPT in another tab is what it can see. The Agent starts every turn with a pre-loaded snapshot of your brand: which prompts you track, where you rank on each one, who's beating you, which pages get cited, which don't, what your perception data says, what crawlers visited your site this week. When you ask "why did my Perplexity score drop", it doesn't guess. It looks.

What it knows about your brand

The Agent opens every conversation with a structured brief of your account. Most of this comes from Trakkr's daily report. Some of it comes from you, by way of the Understanding panel.

Data it has by default:

• Visibility scores and trends across all eight AI models, per prompt and overall
• Your tracked prompts, their rankings, the competitors that appeared alongside you
• Citation sources, citation gaps, and which of your pages get cited (or don't)
• Perception scores and the narratives AI tells about you and each competitor
• Crawler activity: which bots are hitting your site, which pages they're reading
• Visitor traffic from AI sources, if you've connected analytics
• Your action queue, your connected CMS, your audit findings
• The brand profile you set up: industry, description, declared competitors, market

Context you can teach it:

• Open the Understanding panel (brain icon, top right) and add facts about your brand: "we only sell B2B", "we don't support Windows", "our biggest competitor is Hoka, not Nike". The Agent treats this as binding. If a recommendation would contradict something you taught it, it won't make the recommendation.

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 is built for a different job: 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 run it, or dismiss to throw it away. Nothing happens to your workspace until you click.

The kinds of action it can propose:

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

The whole system is built around one rule: the Agent never mutates your account without your click. That rule is what makes the trust model work, so it's worth understanding the edges of it.

A few specifics worth knowing:

• Page fixes preview before they ship. When the Agent proposes a CMS edit, 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 the Agent ever claims it did something, treat that as a bug worth flagging. The pattern is always: chip first, your click, then change.

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.
• A long-running 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.

What's running under the hood

For the curious, a few specifics. The Agent runs on DeepSeek v4 Flash through a streaming tool loop. Every turn starts with a fresh snapshot of your brand, then the 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.

Two depth modes:

• Quick is the default. Fast lookups, short answers, ideal for "what's my score" and "show me X" questions.
• Deep runs a critique pass after the first draft, where the Agent re-reads its own answer with a skeptical-editor lens and rewrites anything that's vague or unsupported. Better for "why" and "audit" questions. Slower.

The mode is picked automatically based on the shape of your question. You don't have to choose.

Common questions

Is the Agent the same as ChatGPT or Claude?

No. The Agent is a Trakkr-specific system built on top of an open model. It has tools, your brand snapshot, the approval flow, and the Understanding panel, none of which exist in generic chatbots. The model itself is just the language layer.

Can the Agent edit my site or my account without asking?

No. Every mutation, from tracking a prompt to applying a page fix, requires you to click confirm on a chip in the conversation. There is no autopilot. If something changed without your click, that's a bug.

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.