A number of years ago for my MBA dissertation, I conducted a survey to understand the business impact of providing free Wi-Fi to visitors of shopping malls in South Africa, specifically looking at how it influenced venue and destination loyalty.
The report I produced remains one of the most visited pages on my site: "The effects on consumers driven brand equity with within-venue Wi-Fi as a tool" .
What readers might not realise is just how long it took to crunch that data in Excel. Manipulating and interrogating the 10,700 responses to extract meaningful insights took weeks. I had to upskill on the fly and elicit the assistance of a colleague just to tackle the complex pivot tables required.
This got me thinking: how quickly could modern AI models crunch this identical dataset and arrive at deep, meaningful insights of their own?
While I usually find Claude exceptionally well-suited for heavy analytical workflows, I decided to test my personal paid subscription to Google Gemini first.
Unfortunately, Gemini repeatedly ran into context handling and processing walls with a dataset of this size. Even attempting to break the workload into linear, bite-sized steps resulted in consistent silent failures. Instead of throwing an explicit error, the UI would simply hang indefinitely when prompted for deeper thematic synthesis. When it did complete, it only spat out surface-level summaries directly mirrored from the prompt itself.
Hoping to keep that thread progressing in the background, I pivot-tested a free-tier Claude account using the exact same robust analytical prompt I started with in Gemini:
"I want you to look at this survey data and produce a report for me with insights into it. I want your summary conclusions, tables to explain the data and spider graphs representing it.
Also to do cross comparisons such as: 'Survey responses received, by province & gender with national totals', 'Responses by Grouping into Categories' (Shoppers, Followers, Entertainers, Workers), 'Indexed landing page preferences of Shoppers compared to the universe', '% distribution by gender of Workers compared to the universe', etc.
Other topics to think about: Psychographic Behavioral Segmentation Profiles, Cross-Segment Digital Sentiment Alignment, Advanced Indexed & Behavioral Cross-Comparisons, Gender behavior comparisons, etc. Run these as the whole universe and also sub-categories.
These are just examples to give you a starting point. You need to apply critical thinking to the data and extrapolate insights and cues. I want you to apply your mind to generating a report with insights that are deep. I want you to come up with ways to cross-analyse between categories.
I also want you to write an executive summary, introduction, content with each section and a conclusion at the end. Accompanying paragraphs with each graphic and table to explain them also need to be added. Produce a visually rich structured report in HTML."
Claude’s initial pass was incredibly accurate and hit the analytical parameters perfectly. However, it hit a technical snag: several visual placeholders it created for charts and spider graphs rendered entirely as blank white space.
|
| Claude's first pass layout containing empty chart containers |
When prompted to resolve the layout gaps, Claude acknowledged the error, but right at that moment, I slammed into the free-tier daily message limit. Having a near-infinite token allowance inside my regular enterprise workplace environment made hitting a consumer cap a unique experience!
Meanwhile, Gemini was still churning away in its chunked-workload attempt, but it ultimately failed to assemble the structured output file correctly.
|
| Waiting on Google's Gemini workspace loop |
Once my Claude limits refreshed, I fed screenshots of the broken layout elements back into the chat window to clarify exactly where the rendering engine dropped the ball. Interestingly, during the compile phase, Claude dynamically loaded third-party, open-source script components into its scratchpad environments to construct the visual assets.
Then, Claude hit an absolute turning point in self-correction:
What stood out was Claude’s programmatic diligence. Once the rendering flaws were identified, it completely re-validated its HTML structures prior to presenting them, offering a complete, transparent root-cause analysis explaining exactly why the initial charting objects broke down.
The Showdown: Model Outputs Compared
After checking three manual iterations with Gemini versus a single streamlined correction loop with Claude, the final outputs revealed a massive performance gulf. It's fascinating considering that Gemini runs on a paid tier, whereas Claude pulled this off entirely within its free-tier ecosystem.
Click each platform logo to explore and compare the full interactive report outputs directly:
Final thoughts:
While Claude clearly did a superior job understanding and materialising the analytical scope of the prompt, the real takeaway here centers on efficiency and baseline capability. Evaluating the absolute time saved against output quality makes one thing certain—re-engineering this dataset manually inside Excel in a few minutes would have been totally impossible.
