Bumping this with additional evidence, as this bug may be harder to reproduce in a generic setup — it only affects MCP tools that have no UI widget associated with their response.
I ran the exact same query on three different clients connected to the same MCP server:
1. ChatGPT — broken 
2. Claude — working 
3. MCP Inspector — working
{
"total_count": 2,
"results_returned": 2,
"results": [
{
"id": "00673379",
"name": "Leroy Merlin Rennes Sud - Chantepie",
"address": "Parc D' Activites Rocade Sud 16 all Guerlédan 35135 Chantepie",
"main_activity": "bricolage, outillage",
"permanently_closed": false
},
{
"id": "09436522",
"name": "Leroy Merlin",
"address": "Rocade Nord Rd 29 Zac Pluvignon 35830 Betton",
"main_activity": "bricolage, outillage",
"permanently_closed": false
}
]
}
All three clients hit the same MCP server, same tool, same query. The server returns identical JSON in all cases — the multimodal_text wrapping is introduced exclusively by ChatGPT.
The key differentiator between the tools on this server: search_businesses has a UI widget and works correctly in ChatGPT. search_businessesId_by_name has no UI widget and triggers the broken behavior. This strongly suggests ChatGPT’s pipeline falls back to citation/grounding mode when no UI definition is found in the tool response, instead of passing the JSON through as-is.
On a side note — between the previous bug and this one, we’ve been struggling to properly showcase our app for close to a month now. We’re the team behind the pagesjaunes MCP integration, which is the main business directory for the French market, so this has been a bit of a tough stretch for us. No pressure at all, but if anyone could share a quick status on whether this is being looked into, it would genuinely help us plan on our end.