ALL POSTS

Success With Adzviser GPT Is 99% in Your Prompts (I): To Receive Correct Reporting Data

I’ve received feedback from users saying, “The Adzviser GPT didn’t return the correct data.” Each time this happens, it shakes me a bit. First, I feel deeply sorry because I strive to offer the best experience possible. Second, I wish I could sit down and explain—quickly—how Adzviser GPT works and how to get the best results, so you don’t lose trust in our service and leave.

Before we dive into the three common reasons why the data might be incorrect, let me briefly explain how Adzviser GPT operates. (You can also choose to watch a more detailed explanation in this video.)

How Does Adzviser GPT Work?

Ever wonder how Adzviser GPT pulls reporting data from your connected accounts when you ask questions like, “How much did I spend on Google Ads this week?” The process is actually super simple.

When you type a question, Adzviser GPT sends a request to our server, which then returns a response. If you're curious, you can actually see this request by clicking on “Talked to copter.adzviser.com”, usually the fourth fourth in the log. The first two requests retrieve the latest information on available metrics and breakdowns for the data source you're querying. The third request retrieves the information about the information about your workspaces.

This request reveals the exact data sent to the Adzviser server. If the request includes incorrect information, it’s almost certain that the response you get won’t be accurate. Now, let’s look at the three most common issues.

Reason 1: Incorrect Date Range in the Request

One common problem is when marketers request data for “last week” or “last month”, but the GPT returns data from “last year”. This issue occurs because, unfortunately, GPT-4o doesn’t always perfectly follow the embedded instructions in our custom model. What instruction, you ask? It's simple: “Always run Python code to calculate the date range based on the user’s query.”

Ideally, every time you input your request, the Python code should automatically determine the exact date range you need. If this doesn’t happen, the result may reflect the wrong time frame. Here's an example.

To help prevent this, it’s a good idea to remind GPT of this instruction at the start of your conversation. Reinforcing this guideline ensures it follows the correct process for date range calculations.

Reason 2: Response Size Is Too Large, Leading to GPT-4o Hallucinations

If you're a Google Ads marketer, you might want to pull a search term report for the past three months. If you work with Meta Ads, you may often check comments on your posts and reels. LinkedIn Ads marketers might look at audience data to see which companies viewed their ads. Unfortunately, these types of requests return large volumes of data from your connected accounts, and this can overwhelm GPT-4o, leading to hallucinations—where the responses contain incorrect or nonsensical information.

To tackle this issue, I've implemented three effective solutions:

Google Sheets Add-on

If you know your query will result in a large dataset, use our Google Sheets add-on. It’s included in your subscription and can handle large data pulls—up to 50k+ rows—without any problems.

Truncate Responses to Fit GPT-4o’s Context window

As of September 2024, the context window for Adzviser GPT's free tier is limited to 8k tokens (about 32k characters). To prevent overwhelming GPT-4o, our server automatically truncates responses to fit within this limit, significantly reducing the chance of hallucinations.

Receive the Full Dataset via Email

If you're unsure whether the response from Adzviser GPT is accurate, no worries. A complete, untruncated version of the dataset is always sent to your inbox in spreadsheet form (screenshot). Since this data comes directly from our server, it won’t be affected by hallucinations and ensures accuracy.

Reason 3: Using Incorrect Metrics and Breakdowns

Another issue users may face is when GPT-4o makes up requests with metrics or breakdowns that aren’t provided by Adzviser’s connectors. This can happen when GPT-4o "hallucinates," meaning it makes up unsupported metrics, which leads to inaccurate responses.

For instance, if you ask Adzviser GPT, “How much did I spend on Facebook Ads last week?” and it incorrectly, though very rarely, compiles the request using the metric “Cost” instead of “Spend”, the result will be wrong. Adzviser only supports certain metrics and breakdowns that are well-documented in our help docs. If GPT-4o selects a metric that doesn’t exist or interprets your request inaccurately, it can lead to frustration 😔.

To avoid this, it’s best to be as specific as possible with your queries. Instead of asking, “What’s my ad spend?” try asking something like, “What are the Total Checkouts Initiated and Total Adds to Cart from Facebook Ads this week?” By clearly naming the supported metrics you want, you reduce the chances of GPT-4o misunderstanding your request or inventing unsupported metrics (chat example). You can check which metrics and breakdowns are supported by Adzviser for each platform in our documentation.

Conclusion

I strive to ensure that all marketers receive accurate reporting data from your connected accounts through Adzviser. As OpenAI’s LLMs become more intelligent and capable, I’m hopeful that these types of issues will become less common in the near future.