Skip to main content
Once your assistant is connected, it can read everything AeoNut has collected for your workspace and think it through with you. That covers which prompts you win and lose, who gets cited instead of you, what competitors publish that you don’t, what you’ve already written, and how your brand is defined in Deep IQ. This page walks through how to ask so you get useful answers back.

​
What it’s good at

  • Explaining a number. If you ask “Why did our citation rate drop on Perplexity this week?”, it will go through the prompts and answers behind the drop with you.
  • Comparing you to a competitor. It can show the rates and coverage side by side, along with the exact pages of theirs that AI answers cite.
  • Finding gaps. It can find topics competitors cover that you don’t, tell you which buyer stage they sit in, and check whether you already have an idea or article for them.
  • Reading the answers themselves. The assistant can open the actual AI responses for a prompt and quote what they say about you.
  • Grounding new content. Before you write, it can pull your personas, voice and product context from Deep IQ, plus the pages that already earn citations.
It can’t change anything in your workspace, because every tool only reads data and nothing spends credits.

​
How it works

Your assistant sees a set of tools, one for each kind of data, and a few prompts, which are ready-made analyses. When you ask a question, it picks which tools to call, calls them with your key, and writes an answer from what comes back. You’ll usually see the calls it makes, and that’s useful because you can check where the answer came from. Most answers depend on these three things:

​
Asking well

Name the time window. “This month vs last month” and “the last 7 days” will give you different answers, and if you don’t say, it uses 30 days. Name the thing you mean. Something like “Rival’s cited pages”, “the prompt about pricing” or “our decision-stage topics” works a lot better than a general question, because a vague question leads to vague tool calls. Ask for evidence. If you say “quote the answers” or “which pages did it cite”, the assistant will open get_prompt_answers or get_competitor_pages and won’t just summarise from the rates. Ask for a format. You can ask for “a table”, “the top five” or “one line per platform”. The tools return a lot of data, and the assistant is good at trimming it down when you tell it how. Ask in steps. The most useful questions usually have two or three steps, for example find the weakest prompt, read its answers, then check whether we have a page on that topic. You can ask for all of it at once or go one step at a time.

​
Questions to try, by area

AI visibility
  • How visible is our brand in AI answers this month, and what moved since last month?
  • Which platform ignores us most, and on which prompts?
  • Which categories do we win, and which do we lose?
  • List our ten weakest prompts by mention rate.
Competitors
  • Who gets cited more than us, and by how much?
  • For Rival, which of their pages earn the most citations, and for which prompts?
  • Where does Rival beat us, by platform, category and prompt?
Pages and sources
  • Which of our pages earn citations, and which prompts cite them?
  • Which domains show up most in the answers for our prompts? Are any of them ours?
  • Why isn’t our pricing page cited? Look at the answers for the pricing prompt.
Content map
  • Which topics do competitors cover that we don’t, and at what buyer stage?
  • Show me Rival’s decision-stage pages on the “pricing” topic.
  • How big is each competitor’s site compared to ours?
Content
  • What have we published in the last 30 days?
  • Do we already have an idea or article about topic X?
  • Read our article on Y and tell me what it doesn’t cover that the cited competitor pages do.
Deep IQ
  • Summarise our brand voice and the personas we target.
  • Which product does the “Ops manager” persona map to?

​
Ready-made prompts

Prompts are complete analyses. Each one tells the assistant which tools to combine and what to produce, so you get the same kind of output every time you run it. You can pick them from your client’s prompt menu. In Claude Code, type / and look for mcp__aeonut__…. A prompt only shows up if your key has every scope that prompt needs.

​
Workflows that work

1

Monday review (5 minutes)

Run Weekly visibility briefing, and then ask “Of the three prompts you flagged, which one should I fix first, and why?”
2

Competitor deep-dive (15 minutes)

Ask “Who gained the most on us this month?” and run Competitor gap analysis on the top name. After that you can ask “Which of those recommendations do we already have an idea for?”
3

Planning content (20 minutes)

Run Content gap plan, for one buyer stage if you like. For each topic that looks good, ask “Show me the competitor pages that cover this and what they include,” and write your briefs from there.
4

Before you write

Ask it to “Pull our brand voice, the persona this is for, and the three pages we already have on this topic that earn citations.” That gives whoever is writing, a person or an agent, the same context AeoNut’s own agents use.
5

Auditing what AI says (monthly)

Run Answer audit and ask it to list factual errors first, because fixing those is usually where a small wording change helps the most.

​
Getting the best results

  • Use one key per assistant. Create a separate key for Claude Code, Cursor and each script. That way you can revoke one without touching the others, and the activity log shows which client made which call.
  • Use the narrowest scopes that still answer your questions. A key with only visibility:read gives the assistant a smaller, cleaner set of tools and keeps your content and IQ out of its reach.
  • Let it page through results. Prompt lists come in pages of up to 100, and content lists come in pages of 15. On a large workspace, “show all prompts” takes several calls, but “the top 20 by citation rate” only takes one.
  • Keep answer pulls small. get_prompt_answers returns up to 200 answers for one prompt. With full=true each one is about 2 KB, so a full pull can reach 400 KB. If you only need part of it, ask for one platform or one day.
  • Watch out for the big reads. get_brand_assets with no filter returns the full content of every asset, which can be tens of KB. It’s better to pass type or search, or list them first and then open one at a time with get_brand_asset.
  • Use the prompts for work you repeat. Free-form questions are best when you’re exploring, and the ready-made prompts are best when you want the same report every week.
  • Check its evidence. When a claim matters, ask “which answers say that?” and the assistant can quote them for you.

​
Limits

  • Read-only. The assistant can’t add prompts, write articles or change settings, so you’ll do those in the app.
  • Rate limit. Each workspace gets 120 requests a minute, shared by all its keys. A busy chat session stays well under that.
  • No bulk export through MCP. If you need a full dataset, use the API directly.
  • ChatGPT connectors aren’t supported yet, because they need an OAuth sign-in flow. Every other client on the connect page works with the API key.

​
What the assistant can see

The assistant can only read the workspace the key belongs to, and only the areas its scopes allow. Everything it reads is logged. You can open Settings → API keys, pick the key, and look at the Requests tab to see every call with its status, timing and the data it returned. You can revoke a key at any time, and anything using it stops on the next request.