Key takeaways
- Running a prompt through an AI visibility MCP server once tells you almost nothing. SparkToro and Gumshoe.ai found less than a 1-in-100 chance ChatGPT or Google AI gives the same brand list twice for the same prompt, so plan to repeat prompts across the week, not just run them once.
- ChatGPT typically surfaces around 5 sources per web-search answer, Perplexity and Google AI Overviews cite roughly 10, and Microsoft Copilot swings wildly between under 2 and nearly 17, according to Promptwatch's data on average sources per response. That changes how hard each engine is to break into.
- The 10 prompts below cover brand discovery, category research, competitor comparisons, problem-solution queries, and source/citation checks, the same five buckets Aleyda Solis and other GEO practitioners recommend building a prompt library around.
- Reddit's share of ChatGPT citations collapsed from about 3.8% to under 1% in a single day (August 14, 2026), per Promptwatch's Reddit citation data, so if your mentions seem to vanish from a Reddit-sourced answer this week, check whether the engine changed before assuming you did anything wrong.
- MCP servers from Rank Prompt, Promptwatch, Otterly.AI and others let you run these prompts directly from Claude, Cursor, or an agent workflow instead of pasting them into a chat window one at a time.
What an AI visibility MCP server actually does
Model Context Protocol, open-sourced by Anthropic in late 2024, is basically a standard plug for connecting an AI assistant to outside data without writing a custom integration for every tool. An AI visibility MCP server sits on top of that standard and exposes your brand monitoring data, prompts, citations, competitor rankings, as tools an AI agent can call directly inside Claude Code, Cursor, or Claude Desktop.
That matters here because it means you don't have to manually paste ten prompts into ChatGPT, Perplexity, and Gemini one by one, screenshot the results, and paste them into a spreadsheet. You ask your agent, through the MCP connection, to run the prompt set, pull back structured results, and flag what changed since last week. Several vendors in this space, Rank Prompt, Finseo, Otterly.AI, Promptwatch, and others, now ship first-party MCP servers for exactly this.
Why one run through each prompt isn't enough
Before the list, a word on methodology, because this is where most people get it wrong. Evertune tested the prompt "top SUV brands" and found that 5 repetitions in ChatGPT gave one brand a 10% visibility score with a margin of error of plus or minus 27 points. The real number could be anywhere from 0% to 37%. Bump the repetitions to 12 and the margin tightens to about 12 points. At 100 repetitions it settles around 6 points.
The practical takeaway: don't run each of these 10 prompts once on Monday and call it a week. Run each one 3 to 5 times across different days, or let your MCP-connected agent schedule repeated runs, so you're looking at a pattern rather than a single noisy sample. ChatGPT also fans a single prompt out into multiple background searches, anywhere from 1 to 8 depending on the week, according to Promptwatch's query fan-out tracking, so the "one answer" you see is already an aggregate of several retrieval steps you can't fully control.
The 10 prompts to run this week
These cover the five areas that actually move the needle: whether you're named at all, how you're described, who you're losing to, where the AI is pulling its facts from, and whether anything happened on the engine side that has nothing to do with you.
1. The plain brand-name prompt
"What is [Brand]?"
This is your baseline. If the AI gets basic facts wrong, your website or Wikipedia-style pages probably aren't feeding it clean entity data. Run this one first every week, it's the quickest tell for drift in how the model describes you.
2. The category prompt, no brand named
"What are the best [category] tools in 2026?"
This is the one that actually matters for new customer discovery. You're not in the prompt, so inclusion is earned, not guaranteed. This is also the prompt most likely to produce wildly different brand lists run to run, which is exactly why it needs repeats.
3. The problem-first prompt
"How do I [solve the problem your product solves]?"
People don't search in product categories, they search in problems. If you only ever test category prompts you'll miss the much larger volume of problem-solution queries where competitors with better how-to content are quietly winning the citation.
4. The direct head-to-head
"[Brand] vs [Competitor]"
Run this for your top two or three competitors separately. Watch not just whether you're mentioned, but which one gets positioned as the better fit, and for what use case.
5. The "best for X persona" prompt
"What's the best [category] tool for [specific persona, e.g. small agencies / enterprise teams]?"
Generic category prompts hide persona-level differences. You might dominate the enterprise answer and disappear entirely from the small-business one, or vice versa.
6. The alternatives prompt
"What are alternatives to [Competitor]?"
This is one of the highest-intent prompts in the whole set. Someone asking this has already picked a lane and is actively shopping. If you're not in this answer, you're losing switchers before they ever hit your site.
7. The site-specific citation check
"According to [yourdomain.com], what does it say about [topic]?"
Since August 8, 2026, ChatGPT Search started using the site: operator at scale, jumping from roughly 0.4% to about 17% of all fan-out queries almost overnight, per Promptwatch's site-operator fanout data. Testing this prompt tells you whether ChatGPT can actually retrieve and summarize your own pages accurately, separate from whether it chooses to cite you in open competitive answers.
8. The sentiment and accuracy check
"Is [Brand] a good choice for [use case]? What are the downsides?"
Inclusion isn't the whole story. A mention paired with inaccurate or negative framing can be worse than no mention at all. Track tone here, not just presence.
9. The fill-in-the-blank association prompt
"___ is the best tool for [specific job]."
This format, borrowed from SiteSignal's brand association methodology, forces the model to commit to a single answer rather than hedging with a list. It's a sharper read on whether you're the default answer in a category, not just one of several names that get mentioned.
10. The source-audit prompt
"What sources did you use to answer that?" (asked as a follow-up to any of the above)
This is the prompt most people skip, and it's arguably the most useful one. ChatGPT citations in August 2026 were dominated by product pages (28.7%) and listicles (10.1%), with social posts collapsing from 4.4% to under 1% after August 14, per Promptwatch's ChatGPT citation types data for that month. If the model names a competitor's blog post or a review site as its source and not you, that's your actual to-do list for next week's content, not a vague "improve visibility" goal.
A simple weekly workflow
- Monday: run all 10 prompts through your MCP server, 3 to 5 repeats each, across at least two engines (ChatGPT and one of Perplexity, Gemini, or Claude).
- Midweek: spot-check prompt 10 (the source audit) against anything that looked off in prompts 2, 6, or 9.
- Friday: compare this week's results against last week's. A single bad run on one prompt is noise. A pattern across three or more repeats and two engines is signal.
- Log engine-side shifts separately from your own content changes. If Reddit mentions suddenly vanish from your results, check whether that's a platform-wide drop, like the one Promptwatch tracked on August 14, 2026, before assuming it's something you did.
Comparison: MCP-enabled AI visibility tools
| Tool | MCP access | Engines tracked | Entry price | Best for |
|---|---|---|---|---|
| Promptwatch | Yes, plus REST API v2, Agent Chat | ChatGPT, Gemini, Claude, Perplexity, Grok, DeepSeek, Copilot, Mistral, Llama, AI Overviews, AI Mode | Free tier (Explore), paid from $95/mo | Teams that want crawler logs, content generation, and action items, not just tracking |
| Rank Prompt | Yes, OAuth 2.1, no API key pasting | ChatGPT, Perplexity, Google AI Mode, Claude, Gemini, Grok | Free, 50 analyzed prompts | Quick manual checks and outreach-finding workflows |
| Otterly.AI | Yes, MCP plus public API and Claude Skill | 7 engines, 50+ countries/languages | $29/mo | Budget-conscious teams wanting broad engine coverage |
| Finseo | Yes, JSON-RPC 2.0, bring your own key | ChatGPT, Claude, Perplexity, Gemini, AI Overviews | Developer-focused, usage-based | Teams building custom agent workflows in Cursor or n8n |
Rank Prompt
Picking the right MCP server for this exercise
If you're already comparing dashboards, not every "AI visibility" vendor ships a real first-party MCP server, some only expose a REST API you'd have to wrap yourself. A useful filter from Foglift's 2026 comparison: does the vendor maintain the MCP server directly, is the data model built for actual agent call patterns like scan, compare, and citation lookup, and can the agent move from data to an action, like drafting a content brief, without you copying and pasting between tools. Promptwatch's MCP server, for instance, connects to Agent Chat and Unified Actions so the same conversation that surfaces a gap can also generate the fix. Rank Prompt and Otterly.AI lean more toward monitoring and reporting, which is fine if tracking is all you need this week.
If you want to browse more options before committing to one, the GEO software directory at bestgeosoftware.com lists a wider range of platforms by feature set, and ai-rank-tools.com focuses specifically on rank tracking across AI engines.
Final thoughts
Ten prompts, run once, on one engine, will tell you almost nothing reliable. Ten prompts, run several times each across two or three engines, logged weekly, and cross-checked against known platform shifts like the Reddit citation drop or the ChatGPT site-operator rollout, will actually tell you whether your AI visibility is moving and why. The MCP server is just the plumbing that makes running that routine fast enough to actually keep up with.