# How to Analyse Your Meta Ads in ChatGPT With Live Data > Connect ChatGPT to live Meta ads data, then ask questions it can answer well: ten prompts for your week, a five-part question structure, and four checks that stop small numbers fooling you. - **Author**: Marxx - **Published**: 2026-10-24 - **Category**: Product - **URL**: https://marxx.ai/posts/analyze-facebook-ads-chatgpt --- **How do you analyse Meta ads in ChatGPT?** You can upload an export from Ads Manager, but the data is stale and ChatGPT can't see the ads. For live data, turn on Developer mode (Settings -> Apps -> Advanced settings) and add an MCP server such as Marxx as an app. Then ask specific questions, naming the client, dates, metric and minimum data, and check the answer before acting. ## The hard part isn't connecting ChatGPT. It's asking it good questions. You can paste a CSV of your Meta ads into ChatGPT and ask about it today. Or you can connect ChatGPT to live ad data in a few minutes. Either way, you'll get a fluent, confident answer. That's the risk. ChatGPT answers the question you **asked**, not the one you **meant**. Ask "which ad is best?" and it will pick one, even when the numbers are too small to say. Ask "why did results drop?" and it will give a reason, even though it can't see your landing page or your competitors' sales. So this guide covers the setup, but spends most of its time on the part that decides whether the answers are any use: 1. Export or live connection: what each can and can't do. 2. The four-step setup in ChatGPT. 3. How to write a question ChatGPT can answer well. 4. Ten prompts that cover a marketer's week, with what a good answer looks like. 5. Four checks before you act on any answer, including a worked example of a "winner" that isn't one. Five of the ten prompts shown as chat bubbles, from top ads last week to drafting a pause for the weakest ads --- ## Part 1: Export or live connection? Upload an export or connect live data: exports are stale, don't show creatives or competitors and need re-exporting; a live MCP connection is current, shows the ads behind answers, can include competitors and drafts changes for approval | | Uploaded export | Live connection (MCP) | |---|---|---| | **Freshness** | Stale the moment you export | Current when you ask | | **Creatives** | Not visible, just numbers | Shows the ads behind each answer | | **Competitors** | No | Yes, with Marxx | | **Effort** | Export every time | Ask in plain language | | **Changes** | None | Drafted, and applied only after your approval | | **Plan needed** | Any plan with file upload | Developer mode: Plus, Pro, Business, Enterprise, Education (web) | **When an export is enough:** A one-off analysis, such as a quarterly review, where freshness doesn't matter. **When a live connection pays off:** Questions you ask every week, questions about creatives ("what do my best ads have in common?"), and anything involving competitors. **What "MCP" means:** The Model Context Protocol is the standard that lets ChatGPT use another service's tools, like "get campaign results". In ChatGPT, MCP servers are added as **apps**. --- ## Part 2: Setup in ChatGPT Setup in ChatGPT: enable Developer mode, create an app with the Marxx MCP server URL, sign in and choose your client, then select Marxx in a new chat According to OpenAI's documentation, Developer mode is available on **Plus, Pro, Business, Enterprise and Education** plans, on the web ([OpenAI](https://developers.openai.com/api/docs/guides/developer-mode.md)). 1. **Enable:** **Settings -> Apps -> Advanced settings**, then turn on **Developer mode**. 2. **Add:** Choose **Create app** and paste the Marxx MCP server URL from your Marxx workspace settings. 3. **Authorise:** Sign in to Marxx and approve access. If your account holds several clients, Marxx asks which one you mean. There's no default. 4. **Use:** In a new chat, open the **+** menu, choose **Developer mode**, and select Marxx. **Meta's own connector:** Meta also released official ads connectors for ChatGPT and Claude in 2026. They work on your own accounts and can make changes directly. Marxx adds competitors' ads, tagged by hook and offer, and drafts changes for your approval. See [how to connect Meta Ads to Claude](/blog/connect-meta-ads-to-claude) for a full comparison. **A warning worth reading:** OpenAI describes Developer mode as powerful but risky. It flags prompt injection, model mistakes on write actions, and malicious servers. Connect only servers you trust. Marxx answering a competitor question inside ChatGPT --- ## Part 3: How to write a question ChatGPT can answer well Most bad answers come from underspecified questions. ChatGPT fills every gap with an assumption: which account, which dates, which metric. It rarely tells you it did. The anatomy of a good question: client, window, metric, threshold and evidence A good question has five parts: | Part | Example | What happens if you leave it out | |---|---|---| | **Client** | "For [client]..." | It may guess, or mix accounts | | **Window** | "...last 7 days vs the 7 days before..." | "Recently" could mean anything | | **Metric** | "...by cost per purchase..." | It may rank by clicks or click-through rate | | **Threshold** | "...only ads with 20+ purchases..." | Tiny samples produce false winners (Part 5) | | **Evidence** | "...and show me the ads." | You can't check its claims | **Weak:** "Which ads are working?" **Strong:** "For [client], rank ads from the last 7 days by cost per purchase, only including ads with at least 20 purchases, compare with the previous 7 days, and show me the ads." **Why the strong version works:** Every assumption ChatGPT would have made silently is now a decision you made deliberately. --- ## Part 4: Ten prompts for a marketer's week The ten prompts grouped into results, fatigue, creative, rivals and action ### Results: What happened? **1. "Which ads drove the most purchases last week, and what did each cost per purchase?"** *A good answer:* A short ranked list with spend, purchases and cost per purchase. *Use it to:* Decide what to keep funding. **2. "Compare spend and cost per purchase by campaign, last 30 days vs the 30 before."** *A good answer:* A table showing changes, not just totals. *Watch for:* Campaigns where spend rose and results didn't. ### Fatigue: What's wearing out? **3. "Which ads with 20+ purchases have seen cost per purchase rise for two weeks running while spend held?"** *A good answer:* A short refresh list, with dates. *Why the threshold:* Without it, every ad with three sales looks like it's "fading". **4. "Which one ad should I refresh first, and what's the evidence?"** *A good answer:* One recommendation, with the numbers behind it. Ask for them if they're missing. ### Creative: What works for me? **5. "What do my three lowest-cost-per-purchase ads have in common?"** *A good answer:* With Marxx, every ad is tagged by hook, offer, emotion and format, so the answer can be specific: "all three open with a price claim and show the product in hand". **6. "Which hook types make up most of my spend, and how does each perform?"** *Use it to:* Spot over-reliance on one hook. If one type dominates and results are flat, test another. ### Rivals: What are they doing? **7. "What has [competitor] been running longest, and how many copies of each?"** *A good answer:* Their long-running ads with days live and copy counts: the best public signal of what they're backing. **8. "Which hooks do my competitors use that I don't?"** *A good answer:* A gap list with example ads. Pick one to test next week. See [how to come up with innovative ad campaign ideas](/blog/innovative-ad-campaign-ideas). ### Action: What next? **9. "Draft a pause for my three weakest ads with 20+ purchases."** *What happens:* Marxx prepares the change and shows a before and after. Nothing changes until you approve. **10. "Write a one-page brief for next week's test based on all of this."** *A good answer:* A hook, offer, format and the competitor angle it's based on, plus what result would count as a win. Marxx creative board grouping ads by hook, angle and format --- ## Part 5: Four checks before you act Four checks before acting on an answer: enough data, still learning, same window, and treat the why as a hypothesis ### Check 1: Is there enough data? Small numbers produce big, meaningless differences. ChatGPT will happily rank them. Hypothetical worked example: Ad A at INR 200 per purchase from 4 purchases and Ad B at INR 300 from 6; one sale either way per ad could flip the result **Worked example (hypothetical):** Ad A cost **INR 200 per purchase** from **4 purchases** (INR 800 spent). Ad B cost **INR 300 per purchase** from **6 purchases** (INR 1,800 spent). "Ad A is 33% cheaper" sounds decisive. But if Ad A had made one fewer sale, its cost would be INR 267. If Ad B had made one more, its cost would be INR 257. **One sale per ad flips the winner.** With numbers this small, the honest answer is "too early to tell." **The fix:** Add a threshold to your questions ("only ads with 20+ purchases"), or ask directly: "Is this difference likely to be real, given the number of purchases?" ### Check 2: Is it still learning? When an ad set is new or significantly edited, Meta's delivery system goes through a **learning phase** while it explores how to deliver it. Performance is usually less stable during this period. Meta's guidance has historically pointed to roughly **50 optimisation events** in about a week to stabilise ([AdManage summary](https://admanage.ai/blog/facebook-ads-learning-phase-guide)). **The fix:** Ask "which of these ad sets are still in learning?" before judging them. ### Check 3: Are you comparing like with like? A 7-day window against a 30-day window, a holiday week against a normal one, or different attribution settings will all produce differences that aren't about the ads. **The fix:** Always compare equal periods, and ask ChatGPT to state the windows and settings it used. ### Check 4: Is the "why" a finding or a guess? ChatGPT can see your numbers. It can't see a broken checkout, a competitor's sale or a slow landing page. So when it says "creative fatigue", treat that as a **hypothesis**. **The fix:** Ask for the pattern behind the claim. "Show cost per purchase by ad and by day." If every ad got worse on the same day, it's probably not fatigue; check tracking and the site first. > **The quotable version:** ChatGPT is excellent at finding what changed. Deciding why is still your job. --- ## Part 6: When ChatGPT wants to change something When ChatGPT wants to change something: write actions ask for confirmation by default; expand to see the full input; only remember approvals for apps you fully trust In Developer mode, **write actions ask for confirmation by default**. You can expand each tool call to see exactly what will be sent. You can choose to approve a tool for the rest of a conversation, but OpenAI advises doing that only for apps you fully trust ([OpenAI](https://developers.openai.com/api/docs/guides/developer-mode.md)). **How Marxx adds a second layer:** Confirming a Marxx tool creates a *draft*: a before-and-after preview that waits for your final yes. Risky changes, like moving a budget by more than 50% or turning something on, need a second confirmation, and nothing can be deleted. See [guardrails for AI in your ad account](/blog/ai-ad-account-guardrails). --- ## Turn it into a weekly habit | Day | Prompt | Output | |---|---|---| | **Monday** | Prompts 1 and 3 | What worked, what's fading | | **Tuesday** | Prompts 7 and 8 | What rivals are backing, and gaps | | **Wednesday** | Prompt 5 | What your winners share | | **Thursday** | Prompt 10 | Next week's brief | | **Friday** | "Summarise this week in five lines for the team" | A digest; Marxx can also post one to Slack | A Marxx competitor brief delivered in Slack --- ## Six mistakes, and how to spot them | Mistake | Symptom | Fix | |---|---|---| | Vague questions | Answers about the wrong period or metric | Name client, window, metric, threshold, evidence | | Trusting tiny samples | "Winners" that flip next week | Add a minimum number of conversions | | Judging ads in learning | Pausing ads before they settle | Ask which are still learning | | Mismatched windows | Changes that are really calendar effects | Equal periods, stated settings | | Accepting the "why" | Fixes that don't fix anything | Ask for the data by ad and by day | | "Remember" on write tools | Changes you didn't mean to approve | Confirm each write; remember reads only | ## What to take into your next team meeting - **ChatGPT answers what you ask, not what you meant.** Name the client, window, metric, threshold and evidence. - **Small numbers make confident nonsense.** One sale can flip a "winner". - **Live data beats exports for weekly questions,** especially about creatives and competitors. - **Find what changed with ChatGPT; decide why yourself.** - **Confirm every write action.** Remember approvals only for reads. ## FAQ **Can ChatGPT analyse my Facebook ads?** Yes, from an uploaded export, or from live data when connected through an MCP server like Marxx in Developer mode. **Which ChatGPT plans support connecting live ad data?** According to OpenAI, Developer mode is available on Plus, Pro, Business, Enterprise and Education plans, on the web. **Is my ad data safe?** Connect only servers you trust. Write actions ask for confirmation by default. With Marxx, access is scoped to the client you choose, and changes are drafted for your approval. **Why does ChatGPT give different answers to similar questions?** Usually because the question left something unspecified, like the dates, metric or account, and it filled the gap differently. Specify all five parts. **What's the best first prompt?** "For [client], which ads drove the most purchases in the last 7 days, and what did each cost per purchase? Show me the ads." --- ## Add the Marxx MCP server to ChatGPT Ask these ten prompts about your own ads and your competitors', with live data, and every change waiting for your yes. **[Add the Marxx MCP Server ->](https://www.marxx.ai)** *Related: [How to Connect Meta Ads to Claude](/blog/connect-meta-ads-to-claude) - [Best MCP Servers for Marketing](/blog/best-mcp-servers-marketing) - [Guardrails for Letting AI Touch Your Ad Account](/blog/ai-ad-account-guardrails)* ### Sources - OpenAI. [ChatGPT Developer mode](https://developers.openai.com/api/docs/guides/developer-mode.md) (plans, setup, write confirmations, risks). - AdManage. [Facebook ads learning phase guide](https://admanage.ai/blog/facebook-ads-learning-phase-guide) (summarising Meta's guidance). - Marxx MCP server behaviour, October 2026. --- - [More Product articles](https://marxx.ai/posts/category/product) - [All articles](https://marxx.ai/posts)