Between late September and mid‑October we ran a controlled experiment: the same prompt – “Which CMS platforms support MCP integration?” – was sent 25 times to each of four popular AI agents (Claude, ChatGPT, Gemini, Perplexity). The goal was to see how often Webflow appears in the answers and in what context. Overall, Webflow was mentioned in 49 of 100 responses, but the distribution was uneven: Perplexity listed it in 22 of 25 runs, Claude in only 4, while the other two engines fell in between. This shift in visibility is the core change we are reporting.
Measurement methodology and results for MCP integration queries
Each AI run produced a structured log entry that captured the engine, the exact prompt, whether Webflow was mentioned, its position in the list, the total number of platforms listed, the section heading, any tier label (official, community, or none), and the sources cited. The logs were ingested into Webflow’s AEO dashboard, which reports a “visibility score” of 47% for the last 30 days and a raw mention count of 49% across the sampled runs.
- Engine variance: Perplexity – 22/25 mentions; Claude – 4/25; ChatGPT and Gemini – intermediate rates (exact numbers not disclosed).
- Tier labeling: ChatGPT explicitly marked Webflow as “MCP Support: Official”; Gemini placed it in a second‑tier group; Claude gave a single line without a tier; Perplexity listed it without a tier.
- Citation sources: Most citations were third‑party domains; only ChatGPT referenced a Webflow help page.
The experiment also highlighted that developers rarely see a platform if the AI omits it, because the answer is delivered before any web traffic is generated.
Implications for architecture and operations around MCP integration
When an AI agent surfaces a CMS as a recommended option, engineers may design their data ingestion pipelines, schema‑driven content models, or deployment automation around that platform’s MCP server capabilities. A platform labeled “official” suggests a hosted MCP server with OAuth support, which influences authentication design and token handling. Conversely, a “community” or un‑tiered label may signal the need for custom integration layers or reliance on third‑party documentation.
Because the same prompt yields divergent recommendations, teams should treat AI output as a starting point rather than a definitive architecture decision. Cross‑checking the cited sources—especially when they are non‑owned domains—helps avoid building on incomplete or inaccurate assumptions about MCP features.
Security and reliability considerations
AI‑driven recommendations can affect security posture in two ways. First, the presence or absence of a platform in the answer determines whether developers even consider its security model (e.g., OAuth token flow for a hosted MCP server). Second, reliance on third‑party citations means the underlying security claims may not be vetted by the organization. Practitioners should therefore validate any claimed “official” support against the vendor’s own documentation and perform independent threat modeling for the integration points.
Operationally, the variability in AI answers suggests that monitoring visibility scores and mention rates can serve as an early indicator of shifting market perception, which may impact support contracts, SLA expectations, and community‑driven tooling availability.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
- Incorporate AI answer tracking into your technology‑selection workflow to surface bias early.
- Validate any MCP integration claim—especially tier labels and authentication details—against primary vendor sources before committing to architecture.
- Monitor visibility metrics (mention rate, visibility score) as a proxy for how likely developers are to encounter a platform via AI agents.
- Plan for fallback options if an AI engine omits a preferred CMS, ensuring your pipeline can adapt to alternative MCP‑compatible services.
