What makes Agent Chat different
Agent Chat is a project-scoped analyst: it runs read-only research agents over your visibility, citations, sentiment, crawler, and content data, then reasons over what came back. It can also act. The main agent carries create, update, and delete tools, so “mark the sitemap action as done” or “create a report for last month” are things it does, not things it explains. Two design choices matter for how you ask:- Research is delegated and read-only: the agent consolidates your question into a few broad research tasks, so one well-scoped question beats five fragments sent one at a time.
- Every change runs in front of you: mutations happen in the conversation, and unattended scheduled runs get no write tools, so you never come back to a silently modified project.
How to ask well
- Give a date range: data is stored in UTC days, and the agent resolves ranges against them. “Last 7 days” works; “compare the last 7 days against the 7 before” works better because it forces a baseline.
- Name the cut you care about: “visibility on Perplexity for my comparison prompts” gets a focused answer. “How am I doing” gets a broad summary you probably didn’t need.
- One goal per message: “find where competitor X outranks me, then draft outreach targets” is one goal with steps. “Also, what’s my sentiment, and can you fix my sitemap” is three chats’ worth in one message.
- Ask for evidence: add “show the responses behind this” or “cite the sources”, so you can check the answer before you act. When a tool returns nothing, the agent is instructed to say so instead of filling the gap.
- Ask for a chart when trends matter: the agent can render line, bar, and area charts inline, which beats a paragraph of numbers.
- Confirm before mutations: phrase state changes explicitly (“create these three prompts on the US monitor”). The agent performs writes itself and reports what changed.
Skills, tools, and model settings
- Agent Skills: the prompt cards on a new chat (grouped Popular, Visibility, Citations, Content, Research) are full, tested asks. Use them as-is or as templates for phrasing: notice how each one pins a date range, names the data to use, and caps the output.
- External tools: Search Console, Web search, News search, Google Trends, DataForSEO, and Webscrape are opt-in toggles in the input bar. Enable them when the question leaves your project data (“compare my Google rankings with my AI visibility” needs Search Console). Selecting a skill enables the tools it requires automatically.
- Model tiers: Standard is the default and fits everyday questions. Switch to Advanced or Max for genuinely hard analysis, and pick a reasoning effort (Low, Medium, High). Heavier settings consume more agent credits per question, so match the tier to the difficulty.
Schedules: recurring asks
Scheduled in the chat sidebar turns any ask into a recurring task. The agent runs it on your cadence and records each result. When the answer changed, you get a bell notification, and you can also post that result to a project Slack channel. Scheduled runs are read-only by design. Use them for standing questions like “weekly: did any model change how it ranks my comparison prompts?”