Graphical Representation: Graphics We Live By (Part VII: Reading a Conversion Rates Chart with ChatGPT and Copilot)

Graphical Representation Series

One of the areas where ChatGPT, Copilot and other similar AI-based chatbots can help is in summarizing a chart saved as image. Ideally, the chatbots should be able also to approximate the points from the chart as well (an image is made of pixels and thus areas should be easy to delimit). So, I was wondering how far the chatbots can be used for these purposes. I used first an image copied from the web, though I realized that not all visual elements could be read (e.g. Copilot had issues retrieving the values for some months) and I had no basis data for comparisons to identify how big the deviations are. 

So, I created a chart in Power BI based on the below chart (see original data):

Conversion Rates Dual Axes Chart

Here's the output based on Copilot over several attempts:

	Original data			First attempt		Second attempt		Third attempt		Fourth attempt
Sorting	Month	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate
1	Jul	8	4	10	1	10	1	8	4	8	4
2	Aug	280	16	275	15	275	15	275	18	275	18
3	Sep	100	13	225	12	225	10	225	12	225	12
4	Oct	280	14	275	12	275	11	275	11	275	11
5	Nov	90	4	75	5	75	6	75	6	75	6
6	Dec	85	3.5	100	5	100	5	100	5	100	5
7	Jan	70	4.5	50	3	50	3	50	4	50	4
8	Feb	30	1.5	50	3	25	2	50	2.5	50	2.5
9	Mar	70	4	25	1	50	2.5	25	1.5	25	1.5
10	Apr	185	11	200	10	200	10	200	10	200	10
11	May	25	3.5	50	4	50	4	50	3.5	50	3.5
12	Jun	195	4	225	10	225	10	225	11	195	4

Please note that same values were repeated to identify whether the chatbot is able to identify the resemblance correctly. 

Based on the data several tests were made with ChatGPT4o, where the following prompts were given:

1. generate the data behind the chart as excel list (see 'First attempt' in the table)
2. generate the data behind the chart as excel list where blue is represented on the left axis and orange on the right axis (see 'Second attempt');
3. reevaluate the values from the previous output considering that the values for Jul are 8, respectively 4% (see 'Third attempt');
4. reevaluate the values from the previous output considering that the values for Jun are 195, respectively 4% (see 'Fourth attempt');
5. generate the data behind the dual axis chart as list from the last conversation;
6. summarize the chart by revealing trends and other statistics.

From attempt to attempt, the values get closer to the original data, corrections being made, though there are points in the final output that diverge considerably. ChatGPT does generate a table which can be easily copied to MS Excel, and even gives some instructions on how to generate a dual axis chart. 

It's great that one can use data between sessions and improve maybe the output. ChatGPT's summary is better than expected and it leverages the latest corrections. However, the estimations need to improve so that the inferences can be correct. 

Here's the output based on the same prompts in Copilot:

	Original data			First attempt		Second attempt		Third attempt		Fourth attempt
Sorting	Month	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate	Conv.	Conv. Rate
1	Jul	8	4	300	18	300	18	8	4	8	4
2	Aug	280	16	250	16	250	16	250	16	250	16
3	Sep	100	13	200	14	200	14	200	14	200	14
4	Oct	280	14	150	12	150	12	150	12	150	12
5	Nov	90	4	100	10	100	10	100	10	100	10
6	Dec	85	3.5	50	8	50	8	50	8	50	8
7	Jan	70	4.5	300	18	300	18	300	18	300	18
8	Feb	30	1.5	250	16	250	16	250	16	250	16
9	Mar	70	4	200	14	200	14	200	14	200	14
10	Apr	185	11	150	12	150	12	150	12	150	12
11	May	25	3.5	100	10	100	10	100	10	100	10
12	Jun	195	4	50	8	50	8	50	8	195	4

Copilot's estimations are higher than the ones made by ChatGPT and deviate more from the original data. No reevaluations are done between prompts for the other values. The summary provides information that can be used to complement ChatGPT's output. 

Overall, ChatGPT seems to perform better than Copilot, at least for this example (though we might talk here about different "generations"). Unfortunately, given that the estimations provided by both chatbots deviate considerably from the expectation, the output needs to be revised and corrected, which decreases the usability of such chatbots. In fact, one can use them to generate an initial set of data and correct then the deviations.

The outputs of other chatbots like Google's Gemini or Claude-3-Haiku (via Poe) can't be compared with the ones from ChatGPT or Copilot yet. Claude-3-Haiku does provide estimated values (even with comma), though they deviate considerably from the original data. 

It would be interesting to test how other charts and plots are processed by chatbots, respectively whether the various visual elements (e.g. gridlines, ticks, markers) make a difference.

Previous Post <<||>> Next Post