Key takeaways
- An MCP server connects Claude to your AI visibility data, so you can ask questions of live numbers instead of pasting CSV exports into a chat.
- A monthly review works because it is boring and repeatable: same prompts, same date window, same questions, every time.
- Social citation share and sources per response vary a lot by engine and drift month to month, so compare engine by engine and keep the window fixed.
- Write the review as a playbook (a skill, a saved prompt, or a scheduled chat) so Claude follows your standards instead of guessing.
- Finish every review with a short action list. A report nobody acts on is just a nicer-looking dashboard.
What an AI brand perception review actually is
A brand perception review answers a plain question: when someone asks an AI assistant about your category, what does it say about you, how often, and where did it get that from?
That breaks into four parts. Whether you are mentioned. How prominently and with what tone. Which claims the model repeats about you (pricing, features, support, reputation). Which sources it cites to back those claims up.
Doing this by hand means typing prompts into five chat windows and squinting at the answers. It works for a one-off curiosity check. It falls apart the second time, because you cannot remember exactly which prompts you ran or how the answers looked last month.
That is the gap Claude plus an MCP server fills. The server holds the data. Claude does the reading, comparing and summarizing.
Why MCP, and what skills add
The Model Context Protocol is an open standard for connecting AI applications to external systems. The common analogy is a USB-C port: one plug shape, many devices. A server you set up once works in Claude, and per the protocol's own site, other clients like ChatGPT, VS Code and Cursor support it too.
Anthropic's own explanation of how MCP and skills fit together is a good mental model for this workflow. MCP gives Claude access to your systems. Skills teach Claude what to do with that access. Their hardware store analogy: MCP is the aisles, a skill is the employee who knows which items you need.

For a monthly review, that split matters. The MCP server pulls visibility scores, citations, sentiment and crawler data. Your playbook tells Claude which charts to produce, which competitors to compare against, and what counts as a problem worth flagging. Without the playbook, you get a different style of report every month and no way to compare them.
Step 1: Pick an MCP server that actually has your data
There are two broad kinds of brand-monitoring MCP server, and the difference is important.
The first kind runs live checks. Open-source projects like ai-brand-monitor-mcp let you query how your brand appears in Perplexity, ChatGPT, Claude and Gemini from inside Claude, Cursor and other MCP clients. Newtation MCP is a remote server aimed at auditing how AI systems describe and cite a brand. These are good for quick spot checks and for teams that want to tinker.
The second kind sits on top of a platform that has been collecting data continuously. That is what you want for a monthly review, because month-over-month comparison needs history. A one-off live check cannot tell you whether your visibility went up.
Promptwatch ships a hosted MCP server with 110+ read and write tools, one-click OAuth or an API key, and works with Claude, Claude Code, ChatGPT, Cursor and other MCP clients. Because it sits on top of the platform's stored monitoring data, Claude can pull a full quarter in one request (the REST API v2 allows date ranges up to 92 days). The API and MCP docs list the tools.

Other platforms in the space are worth checking for their own MCP or API access before you commit, since support varies and changes quickly. Directories like bestgeosoftware.com are a decent place to compare options.
| Approach | History for month-over-month | Setup effort | Best for |
|---|---|---|---|
| Open-source live-check MCP | Only what you store yourself | Low to medium | Spot checks, tinkering |
| MCP on a monitoring platform (e.g. Promptwatch) | Built in | Low | Recurring reviews, teams, agencies |
| Manual prompts in chat windows | None | None | A one-time curiosity check |
Step 2: Connect the server to Claude
Two routes cover most people.
Claude on the web or desktop
Go to Settings, then Connectors, then add a custom connector and paste the server URL. Sign in with OAuth when prompted. The whole thing takes a few minutes. PitchKitchen's walkthrough of brand-voice servers describes the same path and puts it at about four minutes.
Claude Code
Remote HTTP servers are the modern default. The pattern is:
claude mcp add --transport http your-server-name https://your-server-url/mcp
Servers that need a token take a header, for example --header "Authorization: Bearer your-token". Servers that use OAuth authenticate in-session: run /mcp and follow the sign-in. Use the exact URL and auth method in your vendor's docs, not the placeholders above.
A few housekeeping commands help:
claude mcp listshows what is installed.claude mcp remove <name>uninstalls a server./mcpinside a session shows live status.- Add
-s projectto share the server with your team through a.mcp.jsonfile in the repo. Each teammate gets asked to approve it.

A note on safety. An MCP server can read your data and, depending on the tools exposed, change things. For a review workflow, start with read-only access wherever the vendor offers it, and keep write tools for when you actually want Claude to queue content or update monitors.
Step 3: Lock down the review inputs
This is the step people skip, and it is the one that makes or breaks the whole exercise.
If you change the prompts, models, countries or date window each month, you are not measuring your brand. You are measuring your own changes. Decide once and write it down.
- A fixed prompt set, grouped by topic and by intent: branded, competitor comparison, and category ("best X for Y").
- A fixed list of engines and countries.
- A fixed date window, such as the previous calendar month, compared with the one before.
- A fixed competitor list, maybe four to six names.
- A fixed definition of a mention, especially if your brand name is also a common word.
Promptwatch handles that last point with Brand Book analysis instructions, plain-language rules that decide what counts as a mention. Whatever tool you use, make sure there is an equivalent, or your numbers will have noise in them.
Why the window matters
Engines behave differently and they drift. Promptwatch Data's social media citations report for Oct 11 to Nov 9, 2026 found that social platforms make up 13.16% of Google AI Overviews citations, 1.94% of Perplexity's and 0.64% of ChatGPT's. That is roughly 20 times as often in AI Overviews as in ChatGPT. YouTube alone was 6.71% of AI Overviews citations, up from 4.28% in the October report. Reddit was back as ChatGPT's top social platform at 0.40%, up from 0.09%.
Two lessons for your review. Compare engine by engine rather than blending everything into one number. And keep the window identical, since the October report covered 7 days and the November one covered 30, which Promptwatch flags itself.
The same applies to citation counts. According to Promptwatch's average sources per response data, ChatGPT cites about 5 sources per web-search answer, Google AI Overviews about 10, and Perplexity almost exactly 10 every day. Copilot swung from under 2 to nearly 17 within a few weeks. If your mention count jumps in one engine, check whether the engine changed behavior before you congratulate yourself. Perplexity's stability makes it a handy test bench, and Copilot is better judged on monthly trends than weekly snapshots.
Step 4: Write the review playbook
Now turn the process into instructions Claude can follow every month. You can store this as a skill, a project instruction, a saved prompt or a scheduled chat. The container matters less than the content.
Here is a template to adapt:
You are running our monthly AI brand perception review.
Scope
- Brand: [name], competitors: [list]
- Window: previous calendar month vs the month before
- Engines: report each separately, never blended
Pull and compare
1. Visibility score, share of voice and average position by engine
2. Sentiment by topic and by competitor
3. The most repeated claims AI makes about us (pricing, features,
support, reputation) and which engines say each one
4. Top cited domains and URLs, and our self-citation rate
5. Prompts where a competitor appears and we do not
6. AI crawler errors on our key pages, if crawler data is available
Output
- One-paragraph summary, written plainly
- A table of metric changes by engine
- Top 3 things that got better, top 3 that got worse
- Flag any change that could be an engine behavior shift rather
than a real change in our standing
- A prioritized action list, max 7 items, each with evidence
The "flag engine behavior shifts" line is the one I would never drop. It stops Claude from writing a confident story about something that was really a retrieval change.
Reading the numbers together
Single metrics mislead. Read them as a set and they point at the problem:
- Low share of voice usually means a coverage problem: AI does not think of you for these prompts.
- Late position means a prominence problem: you appear, but behind competitors.
- Mentions without citations to your own domain suggest a content or crawlability problem.
- Weak sentiment is a reputation problem, often traceable to a specific claim.
Tell Claude to apply that logic in the playbook. Promptwatch's Academy has plain explanations of Visibility Score, Share of Voice and citations vs mentions if your team needs shared definitions first.
Step 5: Run the review
Open a fresh Claude conversation (or let a scheduled chat do it), confirm the server is connected with /mcp or the connectors list, and paste or invoke the playbook. A typical run looks like this.
- Claude pulls the visibility, sentiment and citation data for both windows.
- It builds the engine-by-engine comparison table.
- It extracts the claims AI repeats about your brand and groups them by theme.
- It cross-checks the claims against your own pages, where the tool supports it.
- It writes the summary and action list.
The claims step deserves attention. Brand Narrative in Promptwatch pulls individual claims out of tracked responses ("affordable for small teams", "its mobile app lacks features"), groups repeats, and scores themes such as pricing, features, support and reputation from 0 to 100, with competitor comparison. Fact Check then lets you mark a claim inaccurate and finds the pages on your site that state it or its opposite, so you know which page the model is learning the wrong thing from. Fact Check is on Business and above, so confirm what your plan includes at promptwatch.com/pricing.
If your server only does live checks and has no claim extraction, you can still ask Claude to read the raw responses and cluster the claims itself. It is slower and less consistent, but it works.
Step 6: Turn findings into actions
A review that ends with "visibility is down 3 points" has not done its job. Push for evidence-backed actions, and keep the list short.
Common findings and the usual response:
| Finding | Likely cause | Action |
|---|---|---|
| Competitor cited, you are not, on category prompts | Content gap | Publish a page that answers the prompt and its query fan-outs |
| Mentioned but never cited | Weak or hard-to-read source page | Refresh the page, check crawler access |
| Inaccurate claim repeated across engines | Outdated or ambiguous page on your site or a third party | Fix the source page, pursue corrections off-site |
| Visibility down in one engine only | Engine behavior change or lost citation source | Check source mix before changing strategy |
| Strong ChatGPT, weak AI Overviews | Little YouTube or social presence | Invest in YouTube for Google audiences |
The social data suggests a practical split. Promptwatch's advice in its November report: put YouTube first for Google audiences, focus on Reddit and LinkedIn for Perplexity, and do not rely on social posts to reach ChatGPT, where your own site and the listicles, guides and comparisons it cites matter more. Whether that holds next month is exactly what your review will tell you.
If you are using Promptwatch, the Action Board gives a prioritized GEO to-do list and Agent Chat can move items around for you. Other platforms have their own versions. What matters is that each action has an owner and a date.
Optional: automate it
Once the manual run feels stable, schedule it. Claude-based scheduled chats, Slack briefings, or an automation tool like n8n or Zapier can kick off the review on the first working day of the month and drop the summary into a channel. Keep a human in the loop for the action list. A model that confidently recommends the wrong fix is worse than no recommendation.
Pitfalls to avoid
- Changing the prompt set mid-stream. Add new prompts as a separate group so old trends stay clean.
- Blending all engines into one score. The engines differ too much.
- Treating a single month as a trend. Look at three months before you restructure anything.
- Letting Claude answer from memory. If it does not call the server, you get a plausible-sounding guess. Check that tool calls actually happened, and name the tools in your playbook.
- Skipping the write-up. Save each month's output in the same place so you can compare reports, not only dashboards.
A quick setup checklist
- Choose a server with stored history and API or MCP access.
- Connect it to Claude and confirm the connection.
- Freeze prompts, engines, countries, competitors and the mention definition.
- Save the playbook as a skill, project instruction or scheduled chat.
- Run the review in the first week of each month, same window logic every time.
- Produce at most seven actions, each with evidence and an owner.
- Archive the report and revisit last month's actions before starting the new one.
If you want a starting point without any setup, Promptwatch offers a free AI visibility report and a 7-day free trial on monthly billing. For teams that would rather hand the whole review process to specialists, 1001 SEO Media runs AI search and GEO programs for brands and uses Promptwatch for its own AI visibility work.
Start small: one engine, fifteen prompts, one competitor set. A modest review you run every month beats an elaborate one you run twice.


