Perplexity Offers a New Conversational Search Experience Challenging Google and ChatGPT
It does a pretty good job of maintaining context and the source links all work, but...
Perplexity AI introduced a new conversational search engine Thursday that is worth testing out. It is based on a large language model (LLM), and the company says the product “is an answer engine that delivers accurate answers to complex questions.”
Like ChatGPT, Perplexity also offers some disclaimers in its footer, saying that “Accuracy is limited by search results and AI capabilities. May generate offensive or dangerous content. Perplexity is not liable for content generated. Do not enter personal information.”
What is Conversational Search?
While Google says it already has conversational search, that claim really is a misnomer. You cannot really be conversational and refine your search with Google as if you are chatting with a human. Instead, it attempts to provide search results based on the context of the queries in a particular session.
ChatGPT, on the other hand, provides a conversational search experience as if you are speaking with a human expert on any topic. It will maintain context from your chat session, and search results will take that into account as you continue your topic search or discussion. However, it does tend to hallucinate by offering incorrect answers on occasion and does not provide the source of its information. That makes it harder to verify whether or not ChatGPT is telling the truth.
YouChat from You.com also provides a conversational search experience. It also has the hallucination and false information issues but maybe not quite as frequently as ChatGPT and does have up-to-date internet access. YouChat will also provide source links. It occasionally does this of its own volition and generally will if you ask to do so in your query. Sometimes the links actually take you to live URLs, but very often, they are 404 “Page Not Found” links that likely never existed.
Perplexity does maintain conversational context and always provides source links both in the response as footnote references and conveniently below the answer as buttons. It also provides three “Related” search results similar to Google’s “People also ask” section.
My testing so far indicates that the search links are always legitimate. I have yet to come across a 404 error. Granted, you will find 404 errors through the web due to deleted pages. Perplexity seems to be avoiding that issue and definitely does not seem to be hallucinating URL links like we found in several instances from YouChat searches.
However, Perplexity’s answers are not immune from hallucinations. I asked YouChat in a December search who the CEO of Twitter was, and it answered “Jack Dorsey.” That was incorrect at the time, as Elon Musk was CEO. I asked the same question to Perplexity twice today, and the first time it responded with Parag Agrawal, who is a former CEO of Twitter. The second time it offered the correct answer with some minor inaccuracies in the full response. Interestingly both responses included a source link to Wikipedia.
It is worth noting that the hallucinations seem to be rare in my Perplexity searches to date. The search responses don’t always answer the question precisely and sometimes lift entire sentences from a source, but they are generally grounded in some easily verifiable information from the provided links.
The error above may have something to do with the position I asked the question in the conversation. It received the wrong answer in a thread where the previous question was about Microsoft’s acquisition of Nuance. When that event took place, Agrawal was indeed the CEO of Twitter. When I started a “New Thread,” Perplexity answered the question correctly — or mostly correctly.
How Perplexity Differs from ChatGPT
ChatGPT is not a search engine. OpenAI executives have stressed this point. It is not connected to the internet and does hallucinate, which the executives also proactively stated. However, many people are enthusiastic about ChatGPT for search because many of its answers are so good.
One of the first questions I asked ChatGPT was, “What is the most proficient voice assistant, and why?” So, I repeated that question for Perplexity and Google to offer some contrast.
We don’t need to discuss the Google result as everyone should know that experience very well by now. ChatGPT and Perplexity both provide suitable answers. Perplexity provided a better answer to the first query and offered a very good answer for the second.
Granted, the second answer did say Sonos is “one of the most well-known voice assistants,” which is incorrect. That gives ChatGPT the edge for the second answer, but Perplexity is superior overall because it is accessing real-time internet data and provides source links which makes verification easier.
Through a number of queries, ChatGPT provided better answers to some and Perplexity to others. Much of this is subjective. For example, ChatGPT is often more erudite in its written responses, while Perplexity is more concise and fact-based. Both do a good job with clarity. The main difference, however, is Perplexity is designed specifically for the search use case and is superior if that is your task.
How Perplexity Works
Perplexity is driven by a large language model, and in order to optimize the results, it appears that additional information is appended to each submitted query. This is also true of ChatGPT. A Twitter post from December 7, 2022 showed what was said to be appended to each query entered into the search engine.
The comprehensive and informative answer that is no more than 80 words explains why the answer lengths tend to be uniform. It also shows how Perplexity is instructing the language model to identify citations and what type of latitude it has in crafting language for the response. The unbiased journalistic tone is consistent with the output even today.
Perplexity’s user interface and language model were updated since this example was published, but the main elements of the appended instruction are likely very similar today. With ChatGPT, the tone and length of response are terms that you can add to your prompt to influence the response. You can get a longer response if you ask for it, but the options to change tone and style appear limited or non-existent.
This is not a criticism of Perplexity. It’s a search engine that offers answer engine features. Crafting a response in Old English, rhymes, or code snippets is not its purpose. It is designed as an information search tool and not a creative outlet.
The State of Conversational Search
I would like Perplexity’s response to “What is conversational search?” to include a reference to features such as maintaining context and enabling users to more easily refine their searches. However, the answer it provides is correct if not complete. More importantly, I can easily dig deeper into the topic by clicking on the sources, which I cannot do with ChatGPT.
The Google search result for this query provides an answer box which an almost identical response. This is because the primary source is the same. Perplexity offers an additional idea in its response, but the result is similar. This will likely be viewed positively by many search users. They already trust Google. If Perplexity is providing similar link sources to Google and similar responses when there is an answer box, then maybe it can be trusted too!
Google does not provide an answer box response to every search, so Perplexity may be a way to ensure you get that type of answer as a response. More importantly, Google doesn’t enable easy search refinement through conversation. Perplexity does. That may be the key differentiator as users choose which service to employ for their next search.
The idea that you might consider a search engine other than Google for your search was almost unthinkable to most people just a few months ago. However, ChatGPT demonstrated to many that there is an alternative approach to search.
ChatGPT is conversational but does not meet many of the traditional requirements of search, such as high accuracy and source links. Perplexity has these features and the conversational context-maintenance of ChatGPT. Microsoft is also expected to introduce conversational search based on GPT-3.5 technology to its Bing search engine in 2023. This means there is a new style of search emerging as an alternative, and it is being led by companies other than Google. No wonder Google issued a “code red” last month.
Let me know what you think about Perplexity AI and whether you plan to use conversational search.
Thanks Bret. I tried out purplexity and I have this slight fear that it sources the wikipedia for too many queries. It seems as if wikipedia is the go to source for this tool. But as you said, it did not give too many hazy or outrightly wrong answers. In fact, I asked a few questions about Sanskrit verse forms and it gave right answers. I somehow do not see purplexity as conversational to warrant a comparison with ChatGPT. ChatGPT is extremely conversational. As in, when I asked if he can help me with search, the bot said he is not connected to the internet and listed out all his limitations. It sounded like a conversation. Whereas purplexity is yet to give such a cohesive answer, although I must admit that I have not tested it extensively as I write this. Thanks for leading me to this wonderful tool.