In Sextas Ímpares #142, I showed different ways to work with campaign data using AI. One approach was exporting a file for analysis, another was connecting the ad account to an external connector so the model could read the information directly.
Confirm what the exported file actually contains
The most common mistake is exporting only the view that's open in the ads manager and assuming the file contains everything. In the lesson I flagged this exact issue: the CSV exported the view that was selected, nothing more. If you're in the campaign view, you don't get ad set or individual ad data.
If you want to understand which ad is losing momentum, you need to open the ad-level view before exporting. If you need daily data, you need to select that breakdown before downloading, otherwise the file only brings aggregated totals for the period. This seems obvious, but it's the most frequent cause of wrong conclusions when asking AI to analyze a CSV.
Before sending the document, open it in a spreadsheet and check what's actually there. Confirm whether the date range matches what you want to study, whether the currency is correct, and whether any important column, like cost per result, wasn't cut off or shifted by a formatting issue.
Remove grand totals and rows that aren't campaigns
Exported tables often include a "total" row at the bottom, and non-campaign rows too. Left uncleaned, a model may treat these as regular campaigns, inflating spend totals and skewing calculated averages. This won't happen with every model, but it's a real, observed risk worth checking before sending the file.
Before uploading the CSV, delete any summary or account total row. If there are archived campaigns with no impressions or spend in the period, it's also worth removing them, just to keep the file focused on what matters. A cleaner file reduces the chance of the model adding things it shouldn't.
Also check the column headers. If they come with odd characters or truncated names, rename them to clear terms like "Spend," "Impressions," "Link Clicks," or "Purchases." This helps the model correctly match each column to the right metric, and it helps you too when reviewing the file before sending it.
Choose the right export level for the question you have
The file's structure needs to match the depth of the decision you want to make. If the question is about budget distribution between campaigns, campaign-level data is enough. If the question is about which creative is fatigued, you need ad-level data with the creative identified.
Imagine, as a hypothetical example, that you want to decide whether to pause an entire ad set or just one creative within it. If you only export at campaign level, the AI has no way to tell you which specific ad is driving costs up, because that row never made it into the file. In these cases, it makes sense to export several files, one per level, or to include columns with the campaign, ad set, and ad identifiers in the same file, so you don't mix levels without a clear indication.
If you want to compare placements, like feed versus stories, you also need to request that detail in the export, because by default it might not come broken down. For anyone structuring this kind of work more broadly, it can be useful to look at the article on using AI in your business beyond chatbot questions.
Explain what you want to decide before asking for conclusions
A question like "what do you think of these campaigns?" leaves too much room for a generic, unhelpful answer. It's more productive to state exactly what decision you're facing and what criteria you use to make it, such as your target cost per acquisition or your available budget limit.
In the lesson I showed analysis suggestions generated from campaign data produced with that kind of context. A suggestion from AI is a starting point, not a ready-made decision. Before acting on it, confirm whether the numbers it used match what you see at the source, because models have already gotten simple math wrong before.
Instead of asking for a general review, ask for something narrower: list the ads with cost per result above average, or compare click-through rate between static and video format in a specific period. The more precise the question, the more actionable the answer, and the less room there is for the AI to invent vague reasoning just to fill out a response.
Relate platform data to actual sales, carefully
The result an ad platform shows isn't always the same as what actually came into the business. A "purchase" event logged in the manager might include zero-value transactions, internal tests, or carts that were never paid. Lead generation campaigns still require later commercial validation to know whether those contacts became customers.
If you want AI to assess a campaign's real profitability, you need to combine the platform cost with confirmed sales in your billing system, over the same date range. Without that cross-check, the risk is that the model recommends scaling an ad set that generates lots of leads but few actual sales. This balance between lead generation and final conversion is also the central theme of how to turn more leads into customers in a services business.
Use aggregated data whenever it's sufficient for the analysis. Usually, to compare investment with sales, you don't need to include customer names, emails, or phone numbers in the request to AI. A total of sales by campaign and by period is already enough for the comparison.
Protect sensitive data before sending any file
Campaign analysis doesn't need customers' personal data to work. Names, phone contacts, emails, or notes from sales meetings don't help calculate acquisition cost or conversion rate, and exposing that data to external tools creates a risk that's easy to avoid.
Before exporting a file crossed with your commercial system, remove columns with personal identifiers. If you need to keep some reference, replace it with an order number or sequential code. To calculate averages and closing rates, AI only needs aggregated totals and generic identifiers.
Checklist before sending the file: 1. Names, emails, and phone numbers removed. 2. Total or summary rows deleted. 3. Date range consistent across all columns. 4. Currency and value scale confirmed. 5. Conversion criteria explained in the initial request.A cleaner file also tends to be faster to process and less prone to reading errors caused by odd formatting.
Verify the calculations before changing anything in the campaign
Any recommendation from AI should be treated as a hypothesis to confirm, not as a fact. Language models have gotten simple math wrong before, like percentage changes between periods, and that doesn't change with a better-prepared file. Data preparation reduces errors, it doesn't eliminate them.
Pick one of the conclusions you receive and redo the calculation at the source, in the spreadsheet. If the AI says cost per lead dropped 30%, confirm whether the two time windows compared have the same number of days and the same calculation basis. If the numbers don't match, adjust the request and ask for the analysis again.
Also note what the file can't answer. A file might be sufficient for one question and insufficient for another, like knowing display frequency or landing page load time. Recognizing that limit avoids rushed decisions about variables that simply weren't in the data.
The most efficient approach today: connect your data to AI through MCP
Today, for recurring campaign analysis, I recommend an MCP connection as the most efficient and effective approach when a reliable connector provides the metrics you need. It reduces file exports and lets you explore follow-up questions in the same conversation by querying the data available at the source.
MCP stands for Model Context Protocol. It is a protocol that connects an AI application to tools and data sources through a server. In this case, a compatible connector can provide access to your ad account and return the requested data within the permissions you have authorised. That connection needs to exist and be configured; typing a platform's name into a conversation does not give it access to your account.
The advantage becomes clear when a second question comes up. After comparing campaigns, you can ask for individual ad data, change the period or investigate an increase in cost per result. If the connector supports those queries, AI can retrieve the additional information without you having to prepare and upload another CSV. You save time gathering data and can give more attention to the decision.
A specific request to start with would be:
Query the last 14 complete days for this account and compare them with the previous 14, using the same time zone. Show spend, attributed purchases and cost per purchase by campaign. Identify the ads that contributed most to the changes and show the calculations. State the attribution window, any missing data and how up to date the available data is. Limit this to analysis, without changing campaigns or budgets.
The criteria in this article still apply: the period, currency, level of detail and definition of a conversion need to be clear. MCP makes access easier, but data freshness and coverage depend on the platform and connector. Comparing ads with confirmed sales also requires access to the sales system and a valid match between the records.
For this work, choose a trusted connector and configure read-only permissions wherever available. Check the totals in the ads manager before acting. A CSV remains useful for a one-off analysis, to keep a snapshot of a period or when a suitable connection is unavailable; for regular monitoring, I would start with MCP.
If you want to go deeper into this kind of work with data and AI in a more structured way, check out my training in AI and online business.