Which AI visibility platforms ship an MCP server, what their tools can actually answer inside ChatGPT or Claude, and how to pick between read-only and read-write setups.
https://mcp.tryprofound.com/mcp, with OAuth for authentication.n nThe read-only framing is deliberate. Profound positions the server as a way to answer questions in chat, like u201cwhich bots crawled our site in the last five minutesu201d or u201cwhat's our visibility trend for ChatGPT this week,u201d rather than a way to execute changes. One caveat for buyers: Profound's published pricing does not clearly state which tier includes MCP access, so confirm before you commit. Plans run from Starter at $99/mo (ChatGPT only, 50 prompts) to Growth at $399/mo (three engines, 100 prompts) to custom Enterprise pricing.n n[tool:profound]n n### Otterly.AIn nOtterly's MCP server is the interesting middle option. It exposes 17 tools, split into 11 read and 6 write, grouped into four categories: workspaces and engines, brand reports, GEO audits, and write tools. The write tools handle prompt creation, prompt deletion, and tag management, and they require separate write permission plus confirmation before executing.n nThe server connects to Claude Code, Claude Desktop, Cursor, ChatGPT, n8n, Zapier, and Storyblok. MCP and API access unlock at the Standard tier, $189/mo with 100 prompts. Below that, the $29/mo Lite plan has no MCP. For teams that want read-write capability without Profound-level pricing, Otterly is probably the cheapest credible option.n n[tool:otterly-ai]n n### KIMEn nKIME's MCP server is available on every plan including the free trial, which is unusual. It connects Claude, Cursor, Windsurf, ChatGPT (Developer Mode), Claude Code, and VS Code via https://mcp.kime.ai/mcp with OAuth sign-in.n nThe toolset is mixed read-write. Most tools are read-only: visibility, share of voice, an eight-factor sentiment breakdown, sources and citations, competitor benchmarking, and topic keywords. A write subset lets the assistant create prompts in bulk and create, update, or comment on Actions, KIME's prioritized action plan. Write tools require explicit confirmation. Pricing starts at u20ac99/mo, with Core around u20ac399/mo and Enterprise custom.n n### LLM Pulse, SE Ranking, Ahrefs, Semrush, and Conductor n nLLM Pulse claims the largest AI-visibility-specific toolset, around 85 tools with both read and write capability, OAuth 2.1 with dynamic client registration, and standard MCP safety hints declared on every tool. The numbers come from LLM Pulse's own comparison, so treat the self-ranking with appropriate skepticism, but the methodology (unauthenticated tools/list calls plus documentation cross-checks, snapshot August 12, 2026) is transparent.n n[tool:llm-pulse]n nSE Ranking's MCP server exposes 217 tools, the largest raw count in the category, but that is the entire SEO suite. Only about 40 tools touch AI search and prompts. Ahrefs has roughly 22 brand-tracking MCP tools embedded in a larger backlinks and keywords suite. Semrush's tool count is unpublished, with OAuth 2.0 or API key auth, and Conductor publishes neither tool count nor auth model. If you already run one of these suites, their MCP servers are worth connecting. If you are buying specifically for AI visibility, the dedicated platforms above will feel less diluted.n n[tool:se-ranking] [tool:ahrefs-brand-radar] [tool:semrush]n n## Comparison tablen n| Platform | MCP tools | Read/write | Auth | MCP pricing gate | Entry price with MCP |n| --- | --- | --- | --- | --- | --- |n| Promptwatch | Not published | Read + write via Agent Chat and Unified Actions | OAuth | Essential plan | $95/mo |n| Profound | 17 | Read-only | OAuth | Not publicly stated | From $99/mo (confirm tier) |n| Otterly.AI | 17 (11 read, 6 write) | Read + write | OAuth 2.0 | Standard plan | $189/mo |n| KIME | Not published | Read + write (confirm required) | OAuth | All plans, incl. free trial | u20ac99/mo |n| LLM Pulse | ~85 | Read + write | OAuth 2.1 / API key | Not published | Not published |n| SE Ranking | 217 total, ~40 AI search | Read + write | OAuth 2.1 / API key | Not published | Not published |n| Ahrefs (Brand Radar) | ~22 | Read-oriented | OAuth | Not published | $199/mo (one platform) |n| Semrush | Not published | Not published | OAuth 2.0 / API key | Not published | $199/mo Starter SEO |n| Conductor | Not published | Not published | Not published | Not published | Custom |n n## What you can actually ask, and why the answers differ by platformn nThe questions below are the ones these servers can genuinely answer today, with notes on which platforms handle each best.n n### Visibility and share of voicen nu201cHow has our brand performed across AI search engines in the past 30 days?u201d Every server above can answer some version of this. The differences show up in engine coverage and granularity. Promptwatch tracks 11 engines plus AI coding assistants, and supports country, state, and city-level tracking on higher tiers. Profound covers up to 9 or 10 engines on Enterprise. Otterly covers six engines from the Standard tier. Ask the same question through different servers and you will get different pies, so know which engines your customers actually use before you compare numbers across platforms.n n### Citations and sourcesn nu201cWhat are our most cited URLs, and which competitors outrank us on citations?u201d This is where the underlying data quality matters most. Promptwatch classifies every citation into 22 content types and 20 source types, and tracks Reddit and YouTube citations in dedicated reports. Profound's citation data flows through the same server as its Agent Analytics. KIME and Otterly both expose top domains and share-of-citations tools.n nOne thing worth understanding when you interpret citation answers: the size of the pie varies a lot by engine. Promptwatch's data on average sources per response shows ChatGPT typically cites around 5 sources per web-search-enabled response, while Google AI Overviews and Perplexity both sit near 10. A brand's citation share in ChatGPT is competing for roughly half as many slots, so a u201cwhich sources got citedu201d answer from ChatGPT is a much more contested snapshot than the same answer from Perplexity.n n### Query fan-outs and prompt behavior n nu201cWhat sub-queries did ChatGPT generate from our target prompt?u201d Some servers, including Finseo's and LLM Pulse's, expose fan-out tools directly. This matters because of how ChatGPT actually searches. Promptwatch's query fan-out research found that average fan-outs per response fell from 2.15 in early December 2025 to 1.84 by early March 2026, and average query length collapsed from around 117 characters to the low 50s by April. ChatGPT search now behaves more like terse keyword search than conversational retrieval, which means your headings and titles should mimic short search queries, not full questions. A fan-out tool lets you verify that directly against your own prompts.n nThe retrieval landscape can also shift abruptly. On August 8, 2026, ChatGPT Search started using the site: operator at scale, jumping from about 0.4% to 17% of all fan-out queries overnight, with searches per response nearly doubling. If your MCP-sourced visibility numbers moved strangely in mid-August, that is likely why.n n### Crawler and bot activityn nu201cWhich AI bots crawled our site this week, and did any hit errors?u201d Profound's Agent Analytics and Promptwatch's AI crawler logs both answer this through their servers. Promptwatch logs 400+ crawlers including ChatGPTBot, ClaudeBot, PerplexityBot, and Meta's new Meta-WebIndexer, which went from roughly 2% to nearly 38% of all tracked AI crawler requests between mid-July and August 9, 2026. If you want to understand why visibility dropped, checking whether the crawler can actually reach and parse your pages is the fastest first question to ask.n n### Sentimentn nu201cHow positively or negatively do the models talk about us, and has it changed?u201d Promptwatch tracks sentiment over time per model. KIME's MCP exposes an eight-factor sentiment breakdown. Both are more useful than a simple positive/negative flag because you can see which dimension of perception is drifting.n n### Acting on the answersn nu201cList our open actions and create one to publish a comparison page against our closest competitor.u201d This is where read-only and read-write servers split. KIME publishes this exact example. Otterly's write tools can add and tag prompts mid-conversation. Promptwatch approaches the same problem differently: its Agent Chat and Unified Actions features work inside the platform itself, and its Content Agents can plan, write, and publish GEO-optimized content to Webflow, Framer, or WordPress. If your workflow is assistant-first, a read-write MCP server fits better. If your workflow is platform-first with the assistant as a query layer, read-only is fine.n n## How to choose between read-only and read-write n nA read-only server answers questions. That is enough for most reporting workflows: weekly digests, ad-hoc checks, competitor snapshots. The failure mode is subtle. You discover a gap, you are ready to fix it, and the assistant says it cannot add the prompt or update the plan, so you open a browser and the momentum dies. LLM Pulse's comparison puts it bluntly: read-only servers cannot finish the job.n nA read-write server closes that loop but introduces risk. Write tools can spend quota, create clutter in your tracking set, or, in the worst case, execute something you did not intend. The vendors that handle this well share three habits: they declare standard MCP safety hints on every tool (read-only, idempotent, destructive), they require OAuth rather than static keys pasted into config files, and they demand explicit in-chat confirmation before any write executes. KIME, Otterly, and LLM Pulse all advertise that confirmation step. Look for it in any vendor you evaluate.n n## Security notes worth five minutes of your attentionn nThe MCP spec requires OAuth 2.1, and the gap between that requirement and how servers actually implement auth is where most of the risk lives. WorkOS's analysis of MCP security rates prompt injection through tool results as critical severity: any external content a tool fetches, including cited pages, can carry hidden instructions the model treats as legitimate. Long-lived tokens held by an agent are rated high severity because a single injection can exfiltrate every credential the agent holds at once. OWASP's MCP Top 10 names token mismanagement, privilege escalation via scope creep, and intent flow subversion among its top risks, with a documented GitHub MCP attack where a public issue body hijacked agent reasoning.n nNone of this means you should avoid MCP servers. It means you should connect them deliberately: prefer OAuth over pasted keys, connect one server at a time, keep write tools disabled until you trust the workflow, and revoke tokens you no longer use. Five minutes of hygiene beats a week of incident response.n n## What to ask before you buyn nThree questions cut through most vendor noise. First, which tier actually includes the MCP server, and is API access bundled or separate? Promptwatch bundles both from $95/mo; Otterly gates it at $189/mo; Profound's gating is unclear in published pricing. Second, does the server expose the data you actually care about? Crawler logs, Reddit and YouTube citations, and prompt volumes are unevenly distributed across platforms, and a server that cannot answer your most important question is not a bargain at any price. Third, who owns the vendor? Adobe completed its acquisition of Semrush in April 2026 and Sitecore acquired Scrunch AI in June 2026, which matters if procurement independence or roadmap stability is part of your evaluation.n nIf you want to browse the full category rather than shortlist from this page, the GEO software directory at bestgeosoftware.com covers the broader landscape, and ai-rank-tools.com lists rank-tracking-focused options. For teams that would rather hand the whole problem to specialists, 1001 SEO Media, the agency that publishes this site, runs GEO and AI visibility programs end to end.n nThe honest summary is that MCP support is spreading fast but unevenly. Promptwatch offers the cheapest full-access entry point with the deepest data model, Profound is the safe read-only choice for enterprise reporting, Otterly is the value pick for read-write, and KIME is worth a trial if you want write tools from day one. Pick based on which questions you will actually ask, not on tool counts, because a 217-tool server you never query is worth less than a 17-tool server you talk to every morning.