Steven Watterson

Logo

Computational Cardiovascular Research @ Ulster University

2024 December

Google Deep Research vs Perplexity Pro for scientific research

Google Deep Research versus Perplexity Pro

I work in science and anything that improves my efficiency is worth its weight in gold, so I’ve been looking at these tools.

Questions

I gave them 3 questions as prompts to see how well they covered the details of a research topic.

  1. What proportion of deaths occur from cardiovascular disease in each country of Europe?

  2. You are a biomedical researcher. Please provide an overview of polygenic risk scores for familial hypercholesterolemia.

  3. You are a scientific researcher working in biomedical sciences. Please provide a 1000 word description with references explaining the percentage of familial hypercholesterolemia cases that have been detected in each country of Europe.

Google Deep Research (GDR) is still experimental so it’s perhaps too early to compare it to Perplexity Pro (PP) which is much more polished. Watch the video to see how they got on in side by side comparisons. I’ve had to speed up the videos because GDR took so long.

Lessons Learned

  1. GDR is very slow. PP took roughly 90 seconds for each answer. GDR took 5-8 minutes for each answer.
  2. I tried this 8 or 9 times. Two times, GDR failed to provide an answer. Once it stated that it’s only a LLM and can’t answer (or words to that effect) and the other time it outputted what looked like a markup placeholder for a response.
  3. GDR did a poor job of keeping to word limits (see Question 3). PP returned text with 898 words. GDR returned text with 2591 words.
  4. As the lengths suggest, GDR’s answers were generally more detailed, but not necessarily about the focus of the question. Much of the extra text went into additional background and context.

Answers

  1. Both were broadly correct.

  2. Both broadly correct, with good detail. Not perfectly comprehensive, but what can you expect?

  3. This is harder information to scrape from papers. GDR didn’t really answer the question, but talked around the subject very knowledgably. PP produced a comprehensive table. Some of the numbers in the table are clearly wrong and not supported by the references (they’ve been mis-scraped), but some numbers are correct.

Conclusion

PP is still the winner for research. GDR is still experimental and it’s hard to imagine that it won’t improve hugely over time. That it will interact with your Google docs data sets has huge potential.