Using AI & Analytics

All AI platforms have limits on the types of files they will accept and/or the amount of data allowed in those files. However, the biggest issue is determining how to phrase the question. 

You need to know what you want before you begin. There must be a specific task to perform, and the uploaded data must closely align with the request. 

Owens Community College Test

I wanted to learn the average lag time between a user’s first session on the Owens Community College website and the session when they complete the application form. My original upload included the standard columns found in any GA4 page report, but the AI became confused and was ‘broken,’ unable to complete the task. I edited the data down to just the date of first visit, the application date, and the count of users, and Gemini was able to handle it. 

What I Learned

My experience with ChatGPT, Gemini, and Perplexity: you cannot upload a large spreadsheet containing multiple columns of unrelated data.

If your question relates to users and conversions, avoid including columns for total pageviews and time on site. The AI processes everything in the spreadsheet simultaneously and then attempts to understand it before parsing the data to answer your question. This behavior may change over time, but this is the current situation. 

The prompt I used was not excessively long, but until I specified what each column reported and the data format in the column, the AI only returned error messages: 

• Neither free ChatGPT nor paid Gemini would accept CSV files, despite my paid Gemini account indicating that it does. 
• Gemini accepts PDF but not the type of PDF that Excel saves, while direct export to PDF from GA4 works in Gemini. 
• Perplexity accepted the CSV file, but its data conclusions were not quite the same as those from Gemini. 

While the results are similar, which conclusion should be trusted? In general, both ChatGPT and Gemini agree that half of application conversions occur in 25 days or less. However, the GPT response further segments the timeframe to “7-21 days,” providing a clearer picture for near-term marketing planning. On the other hand, Gemini seems to look at all the data, attempting to assess the “most average” behavior.

The prompts used were identical for both Gemini and ChatGPT, and I applied the exact same data upload file; therefore, the differences stem from the AI platforms themselves. I was able to ask follow-up questions on both platforms to achieve much closer agreement on the results. 

My Conclusion

I’ve determined that a user of an AI platform must understand how to communicate with it. Numerous large language model AI platforms exist, and just as there are differences in speech patterns between Brooklyn and Memphis, each AI speaks its own dialect. It’s not about one being smarter than another; it’s simply different.

Related reading: Want to Get Found in AI Search?