ChatGPT as an ads manager: what it does alone, with connectors, and where it stops

Two different things are called a ChatGPT ads manager. The first is OpenAI Ads Manager Beta, OpenAI's own platform for buying ads inside ChatGPT: it creates, launches and monitors ChatGPT Ads campaigns, with impressions, clicks and spend in tables, charts and CSV exports, for advertisers in the countries OpenAI lists. The second is ChatGPT used to manage the ads you already run on Meta and Google: reading results, writing copy, deciding what to pause and what to scale.

This guide is about the second. It covers what ChatGPT for ads does on its own, what ChatGPT connectors and Developer mode add, where Meta's own connector fits, how the Adrails connector gives it your live accounts, and what happens when you ask it to change something. The answer to that last question is the one to hold onto: ChatGPT never touches a campaign by itself, and with Adrails a change is a proposal you apply in Adrails.

Upload an export from Ads Manager or Google Ads and ask. File uploads work on the Free and paid plans; OpenAI's limits are 512MB per file, about 50MB for a CSV or spreadsheet, 80 files every three hours, and three uploads a day on the Free plan. For data analysis ChatGPT writes and runs Python in a notebook environment, summarizes rows, trends and outliers, builds tables and bar, line, pie or scatter charts you can switch to an interactive view, and explains the method it used.

Web search is on every plan, Free included, and the answer carries citations you can open. For copy it is strong: give it the offer, the audience and the constraints, and ask for variants to test rather than one answer. Most of what people call ChatGPT marketing is this layer, files, copy and research, and for a solo advertiser it is already most of the week.

• Analyze an export and write the weekly summary
• Draft headline and primary text variants for a test
• Research competitors' offers on public pages, with citations
• Turn a brief into a test plan with a success metric

An export is a snapshot. It is stale the next morning, it holds only the columns you chose, and a comparison with last week needs a second file lined up by hand. OpenAI's data analysis page also says the Python environment cannot make external web requests or API calls, so ChatGPT cannot fetch yesterday's spend, your Shopify orders or a keyword's search volume from inside the notebook. What is not in the file does not exist.

It also cannot act. No upload lets ChatGPT pause an ad set, move a budget or launch a campaign. And a figure it states is only as good as the code it ran: OpenAI's own advice is to review the generated code and assumptions before relying on the result, and to ask for a specific method when the answer depends on one. Treat the model as the analyst, not the ledger.

Connected apps, under Settings, then Plugins, link ChatGPT to a service such as Google Drive or Slack. Every app is a remote MCP server, and a server that is not in OpenAI's directory is added through Developer mode as a custom MCP connector. OpenAI's developer guide lists Developer mode for Pro, Plus, Business, Enterprise and Education accounts on the web, for servers over SSE or streaming HTTP, with OAuth sign-in.

The help center page on Developer mode, updated on 3 October 2026, draws a finer line. Full MCP support, including write and modify actions, is rolling out in beta to Business, Enterprise and Edu workspaces; Pro users can connect servers with read and fetch permissions in Developer mode; custom MCP apps are web only. In a workspace an admin or owner turns Developer mode on and publishes an app, and ChatGPT then uses a frozen snapshot of the server's tools until an admin reviews an update. Check that page against your plan before you count on changing anything from ChatGPT.

OpenAI is explicit about the risk. Its developer guide tells you to watch for prompt injections, model mistakes on write actions that could destroy data, and malicious servers that try to steal information; write actions require confirmation by default, and the help center adds that some especially risky actions may be blocked rather than offered. App permissions range from Always ask to Allow all actions, which OpenAI describes as carrying elevated risk. Connect only servers you trust, and read what each one can write before the sign-in.

Meta Ads AI Connectors, launched as an open beta on 29 April 2026, give an AI agent Meta's own MCP server at mcp.facebook.com/ads. Meta's help center names ChatGPT, Claude, Claude Code and Perplexity as the currently supported agents, so in ChatGPT Meta's server is a Developer mode custom connector like any other. It reports, creates and edits campaigns, ad sets and ads, manages catalogs and custom audiences, reads signal diagnostics and runs A/B tests; mcp.directory counts 29 tools. Every ad it creates is paused until you set it live.

Since Meta's update of 16 July 2026, whoever has full control of a business portfolio can set rules per ad account in Meta Business Suite, under Settings, Integrations, Ads MCP server: block campaign creation, block budget edits, or block budgets above an amount you enter. What the connector leaves out is everything outside Meta: your store's revenue, your Google Ads account, and a review step beyond ChatGPT's confirmation. A budget change you confirm in ChatGPT is written to Meta, within the rules the portfolio owner set.

Adrails runs one MCP server at https://adrails.ai/api/mcp for Meta Ads and, once your workspace can connect it, Google Ads, with Shopify or Stripe revenue read next to spend. In ChatGPT, turn on Developer mode in Settings, open Plugins and add a custom MCP server named Adrails with that URL and OAuth as its authentication. Sign in to Adrails, choose the workspace and the permissions, choose Allow access, review the tools ChatGPT discovered and add Adrails to a new conversation. There is no API key or developer token to copy, for Meta or for Google.

The consent screen asks for one workspace. Read workspace data is required and covers every ad account of that workspace; Prepare advertising changes is a separate permission that only an owner or admin can grant. ChatGPT discovers only the tools the connection's permissions and your role allow, out of the 63 the server registers. performance_period reads one account for exact dates at the account, campaign, ad set or ad level, with a second range for the comparison; store_revenue returns gross, refunded and net revenue with the order count; budget_recommend says whether each budget should be kept, raised, cut or investigated and suggests no amount; creative_insights groups Meta ads by the asset they use and flags fatigue; google_keyword_planner returns monthly searches and the top-of-page bid range.

Every result comes back twice, as exact data for the model and as a readable table for you, and no AI model writes the text. A tool call uses no Adrails agent credit, because ChatGPT's model does the reasoning. Ask for the table first, then the reading, and when ChatGPT states a figure, ask which result it came from.

• Settings: turn on Developer mode
• Plugins: add a custom MCP server named Adrails
• URL: https://adrails.ai/api/mcp, authentication: OAuth
• Sign in, choose the workspace and the permissions, Allow access

Ask ChatGPT to pause an ad set, lower a budget by 20% or launch the campaign a launch_brief has completed, and two prompts follow. The first is ChatGPT's own confirmation of the write action, which OpenAI requires by default. The second is the one that matters: the tool returns a proposal with a link to the Review advertising change page in Adrails, with the message that nothing has been applied.

Open the link. The page names the AI client, the workspace and the ad account, and shows the change field by field; an owner or admin applies it or declines it there. On apply, Adrails reads the platform again and applies the exact change once, or blocks it if the entity changed in the meantime. A budget move is capped at 30% per change and an increase needs measured spend and results; everything a launch builds lands paused, and switching it on is a second proposal that lists every budget that starts spending; an applied change can be undone while the platform still matches what Adrails wrote. A proposal expires after 30 minutes, so an old suggestion cannot be applied by accident.

• ChatGPT's confirmation: OpenAI's setting, it applies nothing
• The proposal link: the Review advertising change page in Adrails
• Budget moves: at most 30% per change
• New campaigns: built paused, switched on as a second proposal
• Undo: while the platform still matches what Adrails wrote