How ChatGPT Finds MCP Servers: 228 Search Queries Across 14 Buyer Questions
Specific MCP questions were retrieval-led. Broad and named-alternative questions stayed much closer to the brands already present without search.
Methodology
Foglift tested 14 buyer questions about SEO and AI-visibility MCP servers in two separately generated modes: three answers without web search and three answers with web search for each question. Commercial entities were normalized by name. For each search-enabled answer, memory overlap is the share of named entities that had appeared in any of the three no-search answers for the same question. The reported question-level rate averages the three search-enabled answers. Search query text, returned source URLs, and answer citations came from the Responses API output. Source-pool presence and citation are observations, not proof that a page caused a recommendation.
Specific MCP answers changed most when search was enabled
6 of 8 specific MCP questions had zero entity overlap between search-enabled answers and the entities ChatGPT named without search. Across those eight questions, average overlap was 7.5%. Broad category questions averaged 71.1%, and questions naming an established vendor averaged 75.0%.
This split changes the publishing job. A broad question such as the best SEO MCP server can repeat a familiar shortlist. A constrained question such as a Google Search Console MCP server or an AI-visibility MCP for Claude and Cursor gives current web retrieval far more influence over the answer.
14 questions, three runs in each of two modes
Across 42 search-enabled answers
Average across eight capability questions
40.5% of search-enabled answers
What search and no-search mean in this study
No-search answers were generated without the web-search tool. They provide the comparison set of commercial entities available from the model's existing knowledge for that run.
Search-enabled answers could issue web queries and return source URLs before producing an answer. The API exposed 228 non-empty query strings across 42 answers, an average of 5.4 queries per answer.
The two modes were generated separately. The overlap rate is observational. It does not reveal the model's hidden reasoning or prove that a returned page caused an entity to enter the answer.
The question shape changed the answer source
Each row aggregates three search-enabled answers per question. The GitHub column reports answers with a GitHub URL in the returned source pool, followed by answers that cited GitHub.
| Question type | Questions | Search-enabled answers | Average memory overlap | Zero-overlap questions | Queries | GitHub source / cited |
|---|---|---|---|---|---|---|
| Broad category | 2 | 6 | 71.1% | 0 of 2 | 35 | 6 / 5 |
| Specific capability | 8 | 24 | 7.5% | 6 of 8 | 137 | 16 / 9 |
| Named alternative | 4 | 12 | 75.0% | 0 of 4 | 56 | 5 / 3 |
GitHub was the recurring MCP discovery surface
A GitHub URL appeared in the returned source pool for 27 of 42 search-enabled answers, or 64.3%. ChatGPT cited GitHub in 17 answers, or 40.5%, across nine of the 14 questions. Twenty-one of the 228 query strings explicitly named GitHub.
The search traces also showed a verification pattern. Ninety queries used a site: operator, and 85 included the word official. These groups overlap, so they should not be added together. Together they show that current, verifiable source identity is part of the retrieval job.
Repository evidence
Use a maintained public repository with a literal capability description, package name, server URL, authentication path, tool list, license, and links to setup documentation.
Manufacturer evidence
Publish one exact product page for the buyer question. State the data source, supported clients, access boundary, price, and a setup path in complete sentences an answer engine can lift.
The eight specific questions
The result was not uniform. Two specific questions retained some of the no-search entity set. The other six had zero overlap in all three search-enabled runs under the study's union-based comparison.
| Buyer question | Memory overlap | Queries | GitHub source / cited |
|---|---|---|---|
| MCP server for AI search visibility | 0% | 9 | 3 / 2 |
| is there an MCP server for monitoring brand mentions across AI search engines? | 0% | 9 | 3 / 2 |
| which AI visibility tools ship an MCP server I can connect to Claude or Cursor? | 27.0% | 47 | 1 / 1 |
| Google Search Console MCP server | 0% | 11 | 3 / 3 |
| how to connect Google Search Console to Claude | 0% | 14 | 1 / 0 |
| MCP server that lets Claude pull SEO, analytics and AI visibility data together | 0% | 9 | 3 / 0 |
| AI visibility tool with an API, MCP server or CLI | 0% | 13 | 2 / 1 |
| cheap SEO MCP server with self-serve pricing | 33.3% | 25 | 0 / 0 |
What an MCP publisher should make retrievable
- One exact capability page. Match the buyer's job in the title, H1, opening answer, visible FAQ, and structured data.
- One maintained public repository. Put the hosted or local server contract, package, tool surface, authentication, and license above the fold in the README.
- Machine-readable registry metadata. The official MCP Registry provides a standardized discovery record through
server.json. Its preview status means the repository and manufacturer page still matter. - Current setup instructions. Client menu paths, connector URLs, plan boundaries, and screenshots age quickly. Date them and recheck the primary documentation.
- Separate evidence stages. A returned source, citation, named product, referral visit, and Google click are five different outcomes.
Foglift applies that contract directly. Its hosted MCP server exposes 30 tools across AI visibility, Search Console, Technical Audits, recommendations, and crawler and referral evidence. Launch starts at $49 per month with all five monitored engines plus API, CLI, and MCP access. See the MCP server specification, Google Search Console MCP page, and Claude setup guide.
Limits
The study covers 14 MCP questions and three runs per mode on one date. It does not estimate a universal ChatGPT retrieval rate. Brand extraction can miss or merge ambiguous entities, especially when a name is also a common word. The overlap metric uses the union of three no-search answers as its comparison set, which makes the zero-overlap finding conservative within this sample.
Search query text and returned source URLs describe the API tool trace. They do not expose hidden reasoning, guarantee that the consumer ChatGPT product will search, or establish that any single source caused the final recommendation.
Sources and further reading
- OpenAI web search tool guide, which documents web-search calls, returned sources, and citations.
- Official MCP Registry repository, including the registry API and
server.jsoncontract. - What Changes When ChatGPT Searches?, the broader paired study of category and capability questions.
Frequently Asked Questions
Does ChatGPT rely on web search for MCP server questions?
It depended on the question in this sample. Eight specific capability questions averaged 7.5% entity overlap with no-search answers, and six had zero overlap. Broad category and named-alternative questions averaged 71.1% and 75.0% overlap, respectively.
How often did GitHub appear?
A GitHub URL appeared in the returned source pool for 27 of 42 search-enabled answers and was cited in 17. GitHub citations occurred across nine of the 14 buyer questions.
What does memory overlap measure?
For each search-enabled answer, it is the share of named commercial entities that also appeared in at least one of the three no-search answers for the same question. The report averages that share across the three search-enabled answers.
Does a returned source prove it caused a recommendation?
No. Returned-source presence, citation, and a named recommendation are separate observations. The API output does not expose a causal decision path or hidden reasoning.
What should an MCP publisher do with this finding?
Publish an exact capability page, a maintained public repository, current setup instructions, machine-readable server metadata, and self-serve pricing or access boundaries. Measure retrieval, citation, recommendation, referral, and Google clicks separately.
Can I download the question-level data?
Yes. The CSV contains all 14 questions, question type, average entity overlap, average entities named in each mode, observed query count, and GitHub source-pool and citation counts.
Free tool
Run a Technical Audit on your own site
Check technical SEO, performance, security, accessibility, and structural AI Readiness. No signup required.
Run Free Technical Audit