An AI agent as your team's strategist
Built-in assistant and MCP: visibility analysis, competitor and prompt management, brand facts, accuracy checks, and saving tasks with drafts.
The built-in agent: chat on top of your data
Open Agent in the sidebar: the assistant uses the selected brand’s data without a separate connection. It reads AI answers, compares metrics and competitors, examines sources, and helps choose next steps. In the same conversation you can approve changes to prompts, clusters and competitors, prepare company facts, or save a task with a draft in Command Center. Use MCP to work with this data from your own AI client.
- Zero setup: the «Dashboard / Agent» switch in the top-left corner of the sidebar, and the chat is open. Data access is exactly the same as yours in the UI
- Presets help you review visibility, perception, reservations and answer accuracy. The competitors section includes presets to find irrelevant competitors and duplicate records.
- Data at hand: reach and share of voice, competitors, cited pages, brand perception and presence quality, prompts with the full text of AI responses, command center tasks, and content plans
- The agent can change data: it first presents proposed changes and waits for confirmation. An explicit request to save a task or prepared draft allows it to save directly. Available operations depend on your permissions.
- Answers link to the relevant dashboard pages: read the conclusion, follow the link, verify the numbers yourself
- Chats are saved in the sidebar and grouped by day, so you can return to an earlier breakdown. The daily answer limit is counted per brand and shared by the whole team; its size depends on your plan
Start with a specific task: “Compare the last 7 days with the previous 7, show answer examples, and separate observations from hypotheses.” Check dates, sample scope and links. A metric change alone does not establish its cause; the assistant should explain when evidence is insufficient.
Review and curate competitors
Ask the assistant to check the registry against your brand profile, websites, aliases and real answer examples. This works in the built-in chat and through MCP with write permission.
- “Find irrelevant competitors and duplicates. Show evidence and proposed changes without changing anything yet.” The assistant should state how many records it inspected.
- After approval, you can enable or disable competitors, correct or clear their official website, and merge duplicates while transferring mentions.
- If a record represents your own brand, the owner can transfer its historical mentions into the brand. This deletes the competitor record and is separate from an ordinary duplicate merge.
- Few mentions or a missing website do not establish irrelevance. Similar names and a shared domain alone do not establish that records should be merged.
Changing the competitor registry or transferring mentions may change historical metrics. This corrects attribution; it is not visibility growth in new AI answers.
Save a task and its draft
Save an analysis outcome in Command Center with an idea, article, brief or proposed edits. This is available in the built-in assistant and through MCP with write permission.
- “Create a task to correct this inaccuracy. Include the page, disputed quotation and suggested wording.”
- “Write an article for this gap and save its draft in a Command Center task.” The assistant writes the text; saving prepared text does not start a separate generator or charge credits.
- “Save this brief to the existing task.” The assistant can attach or update a draft that you can open on the task card.
- After saving, check the task description, evidence links and draft. Saving a task does not publish content or edit website pages.
Explicitly ask to save. An analysis request alone does not authorize creating tasks. Saving to Content Studio and starting its generator are unavailable in this chat.
Prepare facts and check answer accuracy
The assistant and MCP help configure Accuracy: prepare company facts, select existing prompts and investigate contradictions.
- “Prepare company facts from these URLs. Show the statements and sources before saving.” Review each statement; checks can use up to 50 active facts.
- After approval, save the facts and select existing monitoring prompts. Checks use collected answers; a separate branded cluster is not required.
- For a paid run, first request a preview of the facts, prompts and answers to check, plus the credit cost. Approve that specific quote in a subsequent message.
- Ask for findings, the disputed quotation and the original answer. You can review an individual claim, correct its verdict or cancel an eligible check job.
Availability and launch rules depend on your plan and permissions. See the accuracy guide.
Your own AI client as strategist: why MCP
The built-in agent lives in the dashboard; MCP brings the same data into your AI client — Codex, Claude Desktop, Cursor, Claude Code. The gain is context the dashboard doesn't have: your repository and landing pages, article drafts, tracker issues, other connected services, all right there. The strategist reads full AI responses, breaks down every cited source, compares with competitors, iterates on hypotheses — and turns the result straight into a brief, a commit, or a ticket. Ask "why does AI answer this way", "what should the article look like to land in this prompt", "give me a quarterly content roadmap" — and get a full analytical breakdown, not a rehash of dashboard numbers. Technically, MCP (Model Context Protocol) is the standard that gives the agent this access.
- Strategist clients: Codex, Claude Desktop, Cursor, Claude Code, or any other MCP-compatible client. ChatGPT is only available through ChatGPT Apps / developer mode when your account has access
- Connect through OAuth or a personal access token (PAT). Access belongs to your account; after connecting, ask the assistant to list available brands and select one.
- Connection permissions control access: analytics reads and, with write permission, work with prompts, clusters, competitors, facts and tasks. Some actions require brand ownership or a partner account.
- The token belongs to the account, not to a single brand: the agent switches the active brand itself — handy when you run several brands or work as an agency
- The strategist has access to everything in the dashboard plus the full text of AI responses: metrics, sources, competitors, prompts, citations, actions, content plans, and monitoring settings
An AI strategist isn't a lightweight alternative to the dashboard. It's a layer of deep work on top of the data: the work a senior marketer or external consultant typically spends days on gets done in one evening of conversation. The dashboard shows cuts, the command center sets priorities, and the agent does the analytics and strategy — the built-in one on GEO Scout data, the external one over MCP with the context of your files and tasks as well.
Codex
Codex supports streamable HTTP MCP servers in the CLI, IDE extension, and app. MCP settings are shared: add the server from the Codex UI or through `~/.codex/config.toml`.
- Open Codex → Settings → MCP servers → Add server (as in the screenshot), or edit `~/.codex/config.toml`
- Add the `geoscout` server with URL `https://geoscout.pro/api/mcp`
- For OAuth, start authorization from the UI or run `codex mcp login geoscout`
- On the GEO Scout consent screen, pick a brand and approve access
- For PAT access, use `bearer_token_env_var = "GEOSCOUT_MCP_TOKEN"` and set `GEOSCOUT_MCP_TOKEN=gs_<your-token>` before starting Codex
Claude Desktop
Claude Desktop speaks the stdio MCP protocol, so it connects to the HTTP server through the mcp-remote bridge (an npm package). npx installs it automatically — you only need Node.js 18+ on your machine.
- Open `claude_desktop_config.json` (macOS: `~/Library/Application Support/Claude/`, Windows: `%APPDATA%\Claude\`). If the file doesn't exist, create it
- Add an `mcpServers` block with the command `npx -y mcp-remote https://geoscout.pro/api/mcp` (the exact JSON snippet is on the /mcp page in the dashboard, in the Claude Desktop section)
- Save the file and fully restart Claude Desktop. On first launch npx downloads mcp-remote and opens a browser for OAuth authorization
- Approve account access on the GEO Scout consent screen. After connecting, ask the agent to list available brands and select one; the active brand is saved for this connection.
- An MCP server indicator appears at the bottom of the Claude Desktop chat window — that's the live connection
For PAT access, add `--header "Authorization: Bearer gs_<your-token>"` to args. On Windows, there's a known mcp-remote quirk with space escaping — set the token via an `AUTH_HEADER` env var (value `Bearer gs_...`) and reference it in args as `Authorization:$AUTH_HEADER`. Exact snippets are on the /mcp page.
ChatGPT Apps / developer mode
The regular ChatGPT Connectors screen is not a reliable universal path for adding arbitrary external MCP servers. The current ChatGPT path is ChatGPT Apps SDK: the app provides an MCP server and testing runs through developer mode. Use this option only if you have access to ChatGPT Apps development.
- For day-to-day work, use Codex, Claude Desktop, Claude Code, or Cursor — these clients directly support the GEO Scout MCP connection
- If ChatGPT Apps developer mode is enabled for your account, register GEO Scout as the app's MCP backend following the OpenAI Apps SDK docs
- GEO Scout OAuth consent still happens in GEO Scout: the user picks a brand and approves the scope
- Do not treat this section as instructions for the regular ChatGPT Connectors UI: availability and setup depend on account access and developer mode
Cursor
Cursor connects MCP servers through the "Connect to a custom MCP" UI form — no JSON files to edit. Works as a global connection or project-only.
- Open Cursor → Settings → MCP servers → "+ Add server"
- Fill in the dialog: Name = `geoscout`, Transport = `Streamable HTTP`, URL = `https://geoscout.pro/api/mcp`
- Save — Cursor opens a browser for OAuth and you pick a brand on the GEO Scout consent screen
- For PAT instead of OAuth: under Headers click "+ Add header", Key = `Authorization`, Value = `Bearer gs_<your-token>`
- After saving, the server appears in the MCP servers list — the tools icon is available in Composer and Chat
To connect the server for one project only, drop a `.cursor/mcp.json` file with the same config in the repo root — Cursor picks it up automatically and the connection won't leak into other projects.
Claude Code CLI
Claude Code uses a one-line `claude mcp add` connect. OAuth runs in the browser, and the config is saved in the project.
- From the project root: `claude mcp add --transport http geoscout https://geoscout.pro/api/mcp`
- Claude Code opens a browser for OAuth automatically. The `--transport http` flag is required for streamable HTTP MCP
- Approve account access on the GEO Scout consent screen. After connecting, ask the agent to list available brands and select one; the active brand is saved for this connection.
- The connection is saved in the project's `.claude/` folder. For PAT add `--header "Authorization: Bearer gs_<token>"`
Browser agent (WebMCP)
The same strategist, with zero config. If your browser supports WebMCP (a W3C draft built on the navigator.modelContext API), GEO Scout exposes its read-only tools to the AI agent running right in the tab. No token, no OAuth — the agent acts as your session and sees the current brand's data. It's the shortest path to analysis: open the dashboard and ask «why did we drop out of this prompt» in the same place you read the numbers.
- Zero setup — the current brand's tools are exposed to the browser agent automatically when you open the dashboard
- Read-only: visibility, Share of Voice, citations, competitors, sources, content plans. Writes (prompts, clusters) stay with OAuth/PAT
- Authorized by the browser session — no separate token, access to exactly the brands you see in the UI
- Available today in Chrome 146+ Canary behind the «WebMCP for testing» flag; the standard is early and browser support will grow
WebMCP complements OAuth/PAT rather than replacing them. For writes, scripts, and external clients (Claude Desktop, Cursor) use a token connection; WebMCP is the quick read-only strategist in the same window where you already work.
Start here: where your brand is already visible
Ask the agent: “Show the current brand’s visibility over the last 7 days, separately for each AI platform. Include dates, response counts, coverage and Share of Voice. Explain where we are recommended and where we are only mentioned.”
- Check the project first: the agent should name the active brand, analysis period and latest data date. If there is no monitoring data, collect responses before drawing conclusions.
- Request coverage with its denominator: how many responses mention the brand and how many responses were collected in total. This distinguishes missing mentions from missing data.
- Review platforms separately. An overall figure can hide strong presence on one platform and near absence on another.
- Useful output: a short platform breakdown and specific responses with dates, prompts and links to their details. Distinguish mentions, recommendations and reservations about the brand.
Coverage describes how often the brand is named. Share of Voice describes its share of all brand mentions. These metrics can move in different directions.
Where AI chooses competitors instead of your brand
Ask the agent: “Find 5 monitored questions from the last 30 days where competitors are mentioned and our brand is absent. For each, show the question, platform, response date, named brands and cited URLs.”
- Start with real monitored customer questions about choosing a product or supplier. Do not substitute questions that already name your brand.
- Pick one gap and open full responses. Ask: “Which competitor offering is being recommended: the product, price, delivery, industry experience or a specific capability?”
- Separate brands named in the answer from pages used as sources. A competitor mention does not mean the platform cited that competitor’s website.
- Request short supporting quotations from responses and available saved pages. If evidence is insufficient, the agent should label the explanation as a hypothesis.
- Useful output: specific questions where your brand misses out, response examples and a next step for each: verify facts, improve a page or examine an external publication.
Prioritize scenarios that matter to the business. Competitor mention counts alone do not establish search demand or sales value.
Which pages are worth pursuing for brand mentions
Ask the agent: “For the identified questions where our brand is absent, show the most frequently cited pages. Separate our site, competitors’ sites and independent publications. Identify pages with verified brand mentions and pages that still need checking.”
- Review individual URLs and the questions they support. A domain may host both a relevant category review and unrelated material.
- Look for a relevant format: a solution comparison, supplier shortlist, practical guide or industry case study. Assess placement opportunities from the page content.
- If a saved copy is available, ask the agent to read it: which brands are named, which criteria are compared, and which facts and links are included. Check the capture date.
- Do not treat every cited page as a brand mention. Verification may confirm presence, find no mention or remain inconclusive; an unchecked URL needs further checking.
- Useful output: a shortlist explaining which question each page answers, which brands AI currently takes from it and what your agency or PR team should verify before approaching the publisher.
A cited publication is a research candidate rather than a guarantee of AI visibility. Verify a page’s brand mention and the platform’s later citation of that page separately.
Analyze reservations about the brand
Ask the agent: “Analyze reservations about our brand over the last 30 days: what prevents AI from recommending us confidently, when it prefers competitors, and which responses support these findings.”
- Collect the reservation wording and open full responses with dates, platforms and original prompts. Reservations can appear inside positive answers; negative sentiment alone is insufficient.
- Group concerns by meaning: product limitations, purchase terms, missing capabilities, unclear positioning or competitor advantages. Support each group with observed examples.
- Ask: “Who does AI prefer in these answers, and why? Show exact quotations.” Separate genuine product limitations from outdated or unverified claims.
- For important concerns, examine available cited pages for supporting passages. A URL being cited does not establish that it caused the reservation.
- Useful output: reservation → response example → preferred competitor → what to verify or change. Actions may include clarifying terms, confirming a fact, updating a capability description or acknowledging an unsuitable use case.
If reservations have not been extracted or the sample is small, state that explicitly. An empty list does not always mean AI has no doubts about the brand.
Which competitor pages are cited while ours are not
Ask the agent: “Compare competitor and brand pages cited for the same questions over the last 30 days. Review our brand sitemap with available HTML titles and show where we have a matching page and which material appears to be missing.”
- Start with exact competitor and brand URLs. Align the period, platforms and questions: a homepage, product page and how-to guide address different needs.
- Request paginated sitemap inventory with available HTML titles. Include the inventory date, completeness and URLs whose titles could not be fetched.
- Find a brand page on the same topic for each relevant competitor page. Titles and URLs help select candidates, but read the content before confirming a match.
- Distinguish our pages that are cited, surfaced as sources but were unused, exist in the sitemap without observed citations, or have no match in the available inventory. The last case does not prove the website lacks such a page.
- Useful output: competitor URL → how it answers the question → matching brand URL and citation status → verified differences → action. Improve matching existing pages first; propose new material after checking for an existing match.
No citations in the sample does not mean a page is unindexed or unknown to AI. If the sitemap or page content is incomplete, identify what remains unchecked.
What data the agent can access
Through MCP the agent gets read tools for analytics and scoped write tools for selected operations. You don't need to memorize them — the agent picks the right one for each question.
- Metrics: coverage, Share of Voice, Citation Share, URL mention rate — by period, provider, and prompt cluster, with period-over-period deltas
- Sources: domains and URLs, categories, competitor comparisons and saved page content when available. A saved version helps examine context but may differ from the current page.
- Competitors: active and inactive records, websites, aliases, comparisons and competitive gaps. Mention totals in the registry cover all time rather than a selected period.
- Prompts and responses: top gap prompts, per-prompt details with full AI responses, sentiment, scores, citations, mentioned competitors
- Actions and content plans: read tasks and plans, create Command Center tasks, and save assistant-written ideas, articles or edits on a task card.
- Answer accuracy: company facts, prompts selected for checking, check results, disputed claims and source links.
- Brand sitemap: URLs, actual HTML titles, freshness and discovery completeness. Large inventories are paginated; unavailable titles remain unknown rather than inferred.
The full list of available tools with descriptions is on the /mcp page in the dashboard. Need a tool or data cut that isn't there yet? Message support — the MCP toolset is actively expanding.