AI for Paid Media
How to Use ChatGPT or Claude to Help Manage Ad Campaigns
Use ChatGPT or Claude to analyze search term exports, mine negative keywords, draft ad copy variants, write creative briefs, and turn weekly numbers into a readable report. Neither assistant can open your ad account on its own; that needs a connector like MCP, and a person should still approve anything that spends money.
Key takeaways
- ChatGPT and Claude work best on paid media when you hand them an export or brief; neither has native access to your ad accounts without a connector.
- Five jobs they do reliably: search term analysis, negative keyword mining, ad copy variants, creative brief writing, and weekly performance narratives.
- Prompts work better with four parts: role, real data, constraints, and output format; vague prompts produce generic output.
- Assistants can hallucinate benchmarks, cite outdated platform rules, and invent numbers in reports when you don't paste the real figures, so always verify against the platform.
- MCP now lets both assistants connect to live ad data on platforms like Google Ads, Google Analytics and Meta, but a connected chat still has no memory of your pipeline between sessions the way a purpose-built agent does.
ChatGPT and Claude are genuinely useful for paid media, but only for the parts of a campaign that live in text and numbers: analyzing an export, drafting copy, writing a brief, turning a spend report into a narrative someone will actually read. Neither assistant opens your Google Ads account or LinkedIn Campaign Manager on its own. That changes once you connect one to your ad accounts through a protocol like MCP, and even then a person still has to approve anything that spends money.
What ChatGPT and Claude can and can't do with your ad account today
In a default browser chat, neither assistant has any connection to your ad platforms. What they can do is work with files you give them. OpenAI's help center describes ChatGPT's data analysis feature, which lets you upload a CSV, Excel, PDF or JSON export and have the assistant write and run Python against it, up to 10 files per conversation, up to 512 MB per file, with CSV and spreadsheet files capped around 50 MB. Claude does the same job through its code execution and file creation capability, which Anthropic's support center confirms is on by default for every plan from Free through Enterprise, turning an export into a working Excel model, a slide deck or a PDF. That's reading a file you hand it, not reading your account live, and further still from changing anything in that account.
Live access requires a connector. Google now publishes free, read-only MCP servers for Google Ads and Google Analytics, and Meta opened its own Ads AI connectors into beta in April 2026, which can create and edit campaigns, not just report on them. Third-party connector tools plug the same assistants into LinkedIn, Meta and Google through aggregated data pipes. Even where a connector can write to a live account, the sane setup is read-only reporting first, with any change that touches real budget routed through a person before it goes live.
Quick gut check: if an assistant's answer about your account can't be re-traced to a number you can see in the platform itself, don't act on it. Ask it to show the exact rows or math it used.
Five jobs a general-purpose assistant does well
These jobs play to what large language models are actually good at: pattern-finding in text, language variation, and turning structured data into prose. All five work from an exported file, no live connection required.
| Job | What you feed it | What a human still checks |
|---|---|---|
| Search term analysis | Search terms report (Google Ads or Microsoft Ads export) | Whether a "wasted spend" call actually ignores a branded or high-intent term |
| Negative keyword mining | Same search terms report, plus your product/service list | Terms that look irrelevant but convert, before they get added to a negative list |
| Ad copy variants | A written brief: offer, audience, proof points, character limits | Claims, numbers and compliance language, every time |
| Creative brief writing | Campaign goal, audience, past winning creative, brand notes | Whether the brief actually reflects this quarter's positioning |
| Weekly performance narrative | This week's and last week's real numbers, pasted in or uploaded | That every number in the write-up traces back to the platform, not the model's memory |
Search term analysis
Paste or upload the search terms report from Google Ads or Microsoft Ads and ask the assistant to group terms by intent, flag ones that don't match your offer, and call out spend concentrated in low-converting terms. A larger context window helps here: you can paste a full multi-thousand-row export in one pass instead of chunking it, which keeps groupings consistent across the whole file.
Negative keyword mining
This follows directly from search term analysis: once terms are grouped, ask for a negative keyword list split by match type and by ad group versus campaign level, with a one-line reason for each. Have someone who knows the account scan the list first. A term that looks irrelevant in isolation sometimes converts fine once checked against actual conversion data.
Ad copy variants
This is the highest-volume use case, because copy testing needs volume, not two headlines and a prayer. Give the assistant a brief (offer, audience, proof points, exact character limits) and ask for variants by angle: urgency, social proof, specific benefit, direct offer. Generate more than you need, then cut hard based on what the brief supports.
Creative brief writing
Feed it the campaign goal, audience segment, what's worked before, and brand voice notes, and ask for a one-page brief a designer could work from without more context. Briefs are structured documents, exactly the kind of output a language model produces reliably when the input is specific.
Weekly performance narrative
Paste this week's actual numbers (spend, leads, CPL, pipeline if you track it) next to last week's, and ask for a short narrative: what moved, a plausible reason why, and what to watch next week. This saves real time turning a spreadsheet into something a CMO will read. The failure mode is skipping the real numbers; more on that below.
Give it the right context before you prompt
Output quality is almost entirely a function of input quality. Four things are worth preparing once and reusing every week:
- Clean export files. Keep column headers consistent (campaign, ad group, spend, clicks, conversions, CPA) so you're not re-explaining your schema every session.
- Naming conventions. If campaigns and ad groups follow a consistent pattern (channel, segment, funnel stage, offer), the assistant can parse hierarchy from the name alone, which makes grouping and summarizing far more accurate.
- ICP and offer notes. A half-page doc on who you sell to, their pain points and your proof points turns generic copy into copy that sounds like your company. Paste it once per project, not once per prompt.
- Brand voice rules. A short list of tone rules and banned words keeps 50 generated headlines from needing 50 manual edits.
A prompt pattern that produces usable output
The pattern that works consistently has four parts: role, data, constraints, output format. Skip any one of them and you get generic filler.
- Role: tell it what kind of expert to act as and for what channel.
- Data: paste or upload the actual export, brief, or numbers, not a description of them.
- Constraints: character limits, what to avoid, what must be true (compliance, claims you can actually back up).
- Output format: exactly how you want the answer structured, so you can act on it without reformatting.
Worked example, for negative keyword mining on a Google Ads account:
You are a B2B paid search manager reviewing a Google Ads search terms report for a project management software company selling to mid-market operations teams. Here is the search terms export for the last 30 days: [paste CSV data]. Our product is for teams of 50 to 500 people; we do not serve consumers or enterprise (5,000+ employees). Group the search terms into: (1) clear negatives to add now, (2) terms to watch before adding as negatives, (3) terms indicating new keyword opportunities. For each term in group 1, give the match type to add as a negative and a one-sentence reason. Output as three separate tables.
The role sets expertise and channel, the data is the actual export rather than a description of it, the constraint rules out false positives on legitimate B2B queries, and the output format means group 1 can be copied straight into Google Ads without re-sorting it.
Where this goes wrong
Four failure modes come up often enough to name directly.
Hallucinated benchmarks. Ask "what's a good CTR for LinkedIn sponsored content" and you'll get a confident number. Unless the assistant names a specific, dated source, treat it as a guess dressed as a fact, and verify it yourself.
Outdated platform rules. Models train on data with a cutoff, and platforms change character limits, match type behavior and ad formats often. An assistant describing a limit or targeting rule may be describing last year's version. Check the platform's own current help page first.
Invented numbers in reports. Ask for a weekly narrative without pasting real numbers, and a model will sometimes produce plausible-looking figures that aren't real, especially mid-conversation when it's staying consistent with something said earlier. Always paste the actual export for the period, and spot-check two or three numbers against the platform before the report goes anywhere.
CRM and pipeline data privacy. Consumer ChatGPT and Claude accounts are not the place to paste raw Salesforce or HubSpot exports with named contacts and deal values. For that level of data, use your organization's Team or Enterprise plan, where admins control network access, and confirm data retention and training settings first.
Connecting assistants to live ad data with MCP, and when you need more than a chat window
MCP, the Model Context Protocol, is the piece changing what "connect my ad account" means. Anthropic open-sourced MCP in November 2024 as a standard for two-way connections between AI assistants and the systems where data lives: a server exposes an app's data and actions, and a client, the assistant, connects to it. The protocol spread well beyond Anthropic's own products; a one-year review of MCP's adoption named ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code among the platforms that had integrated it, alongside more than 10,000 active public MCP servers, around the time Anthropic handed the protocol's governance to the Linux Foundation's Agentic AI Foundation. In mid-2026 the protocol's maintainers shipped a significant spec revision, MCP 2026-07-28, with Anthropic rolling out support for it across Claude products, a sign the standard is still evolving. Practically, a connected assistant can pull live spend and account structure into a chat instead of waiting on a CSV export, and in Meta's case, propose changes directly. For a deeper look at what MCP means for ad platforms specifically, see Tofu's explainer on MCP.
That capability is also where the honest limits of a chat assistant show up. A connector gives an assistant read or write access to data and actions; it does not give it a model of your pipeline or which segments actually close. Ask ChatGPT or Claude to "build me an audience of our best-fit accounts" and it has no way to know which accounts in your CRM became customers or how long they took to close, unless you've typed that into the conversation. A chat assistant, even a connected one, reasons over whatever you put in front of it that session. It isn't maintaining an ongoing model of your business.
That's the practical dividing line. If the job is "turn this export into something readable" or "give me 40 headline variants," a ChatGPT or Claude session handles it well, connected or not. If the job is "read our closed-won deals and build the audience, plan and creative from that, then stage it into the ad account," that's a different tool category. Tofu Ads Agent reads pipeline data from Salesforce or HubSpot, turns it into an ICP and funnel plan, builds and sizes audiences, and stages campaigns into your own LinkedIn, Google or Meta ad account switched off, for a person to turn on. The real difference from a chat assistant is that the agent works from your actual CRM and ad-account connections between sessions, not from whatever you happen to paste into that day's conversation. For the fuller comparison, see AI ads agent vs. ad automation software.
Either way, the control model should match: recommendations and drafts come from the AI, and a person decides what goes live, whether that's a search terms export pasted into Claude on a Tuesday or a connected agent running weekly. For more on where that line should sit on your team, see automate vs. human in B2B paid media.
Frequently asked questions
Can ChatGPT or Claude actually log into my Google Ads or LinkedIn account?
Not by default. In a standard chat, both assistants only see what you upload or paste, like a CSV export. Live access requires a connector built on MCP, such as Google's free read-only MCP servers for Google Ads and Google Analytics or Meta's Ads AI connectors, and even then, changes to a live account should go through a person before they launch.
Which is better for paid media work, ChatGPT or Claude?
Neither wins outright. Practitioners comparing both report ChatGPT ahead for fast, high-volume ad copy and file-based data analysis, and Claude ahead for strategy, longer documents and anything in a regulated vertical. For pure PPC mechanics like search term mining and negative keyword lists, both handle the task well once you give them the actual export.
What is MCP and do I need it to use AI for ad campaigns?
MCP, the Model Context Protocol, is an open standard Anthropic released in 2024 for connecting AI assistants directly to external data sources and tools instead of relying on copy-pasted exports. You don't need it to get value from ChatGPT or Claude on paid media tasks like copy and analysis, but you do need a connector built on it if you want the assistant to read your ad account live rather than a file you uploaded.
Is it safe to paste CRM or customer data into ChatGPT or Claude?
Not into a personal, consumer-tier account with named contacts, deal values or other sensitive pipeline data. Use your organization's Team or Enterprise plan instead, where admins can control settings like network access and confirm data retention and training policies before anyone uploads customer information.
How do I stop an AI assistant from inventing numbers in a performance report?
Always paste or upload the real export for the exact period you're reporting on, rather than asking the assistant to recall or estimate figures from earlier in the conversation. Before sending the report anywhere, spot-check two or three of the numbers in the output against the platform itself.
When should I use an AI ads agent instead of a chat assistant?
Use a chat assistant for self-contained text jobs: analyzing an export, writing copy, drafting a brief. Use a purpose-built agent when the job requires an ongoing model of your business, like reading closed-won deals from your CRM to build an audience and a funnel plan, then staging campaigns into the ad account. Tofu Ads Agent is built for that second category; see how it compares to rule-based automation.
Sources
- Data analysis with ChatGPT · OpenAI Help Center
- Create and edit files with Claude · Claude Help Center (Anthropic)
- Introducing the Model Context Protocol · Anthropic
- What Is MCP (Model Context Protocol)? Complete Guide 2026 · Atlan
- Bringing MCP 2026-07-28 to Claude · Claude (Anthropic)
- Claude vs. ChatGPT for Marketing: What I Actually Use and Why · Orr Consulting


