Introduction
With the development of large language models (LLMs) like ChatGPT, searching the internet for information has become more efficient and easier. Large Language Models are trained on vast text corpora with billions of parameters, enabling them to understand and interpret human language. Traditionally, search engines relied on simple methods like keyword matching and rank ordering to find the most relevant web pages or links when a user asks a question. However, search engines like Open AI Search or Perplexity are integrated with LLMs, which can accurately identify user intent and generate a customized response to the user’s query. AI search engines can have numerous applications in industries like healthcare, e-commerce, and entertainment, where multimodal query support can enhance customer experience. For example, a user can upload a picture of a product and search for similar items or dictate their requirements quickly with voice commands.
In this article, I’ll introduce you to a few major AI search engines and discuss their internal workings, key features, and challenges.
1. OpenAI Search (GPT-Driven)
Open AI search is a widely used AI search engine that can analyze various search results and generate human-like responses to user’s queries. The conversational nature of querying makes it a user-friendly interface. It is powered by GPT (Generative Pre-trained transformers) models such as the latest version, GPT-4, or the more cost-effective GPT-3.5, depending upon the complexity of the task.
Key Capabilities of Open AI Search:
Contextual Understanding: GPT models have a transformer-based architecture that uses a self-attention mechanism, which allows it to store and process the user's previous inputs and search for contextually correct answers.
Personalized responses: Rather than providing a bunch of links like traditional search engines, Open AI uses GPT models to scan all the sources, extract relevant information, and summarize them for the user in the format they prefer. For example, users can prompt GPT to generate responses in concise bullet points, detailed paragraphs, statistical reports, and much more.
Code and Technical Support: GPT models are also trained on understanding different programming languages like Python, C++, and Javascript and can help users debug or optimize their code.
Accurate Data Retrieval using Embeddings: Apart from web search, users can also upload documents or datasets and use OpenAI to query and find the required information. OpenAI has APIs that convert user queries into vector embeddings, which are matched against the vector dataset to find the most similar data points. 2. Google AI Overview
Current Challenges
While the GPT models are updated regularly, the search does not happen in real-time and may miss out on the latest advancements.
Generalization bias is a common challenge with LLMs like GPT that are trained on very broad datasets, and the response can lack the technical depth or nuance needed. This is commonly faced in fields like healthcare and law, where a high domain knowledge is essential.
Open AI search does not provide citation links or direct sources of the information and users may need to perform additional fact-checking before using the results.
GPT models can unintentionally generate biased responses towards gender or communities if the underlying training data has a bias.
2. Google AI Overview
Google’s AI search engine can generate accurate and personalized responses to user’s queries by surfing the web in real-time. The search engine is integrated with Google Gemini, a multimodal LLM that is designed to process and generate various data types including text, images, videos, and audio. The LLM has a transformer-based architecture that is optimized for retaining context during a conversation. The latest version of Gemini has advanced features like an enhanced context window that enables it to handle complex queries like reasoning questions or generating detailed reports. Apart from Gemini, the search engine also uses models like LaMDA (Language Model for Dialogue Applications) that were designed specifically for conversational AI tasks.
Key Capabilities:
ML-based Ranking Algorithms: Google AI used machine learning algorithms like RankNet and LambdaRank to rank the search results, prioritizing the quality of information. These algorithms measure a webpage’s content quality through the relevance of keywords, backlinks, user behavior, and trustworthiness.
Improved Reasoning and Logic: Google’s search engine also shows significant improvement in logical reasoning, and can handle complex queries.. It uses PaLM (Pathways Language Model), which was trained on mathematical datasets.
Real-time Image & Video Processing: The extensive multimodal support is what sets Google AI apart from other search engines. Users can upload an image or video, and query the search engine. For example, one can upload a picture of a plant, and prompt Google “What is this plant and what fertilizers should I use for it?” It can also be used for facial recognition, object detection or generating live transcripts from a video during a webinar or panel discussion.
Extended Contextual Understanding: As the latest Gemini model is developed with longer context windows, it can process large-scale queries better. For example, users can upload a 300-page market trends document and prompt “Summarize how the market trends were affected by seasonality?”
Translation to over 100+ Languages: Google’s LLMs are trained in over a hundred regional languages like Chinese, Hindi, and more. It can be used to translate news articles, transcripts, or even books accurately.
Current Challenges
Poor transparency and explainability: Google AI relies on multiple complex models, and it is difficult to explain why a certain response or decision was made.
Google collects and processes vast amounts of user data to provide personalized search results. This raises concerns about user data privacy and security.
It is optimized for speed and can miss out on detailed information while generating results
3. Bing Search
Bing search was developed by Microsoft, it combines traditional web search techniques with AI models to enhance search relevance. Bing Chat is integrated with Open AI’s GPT models and BERT (Bidirectional Encoder Representations from Transformers). It uses a mix of traditional ranking techniques like web crawling, and indexing with advanced ML models - finding the right balance. Bing Chat can be used for a personalized shopping experience too, as it filters results based on user history, location, etc
Key Capabilities
Comprehensive Web Search: Bing used automated bots like web crawlers to browse through billions of web pages, extracting their metadata, page structure, and media. Post crawling, the data flows through Bing’s search index and ranking to shortlist the most relevant web pages for the user.
Integration with Microsoft Eco System: Bing Search is integrated with many Microsoft products such as MS Word, Powerpoint, Microsoft Teams, and Onedrive. Users can use Bing AI to draft PowerPoint presentations by using prompts like “Create a slide with 5 points on climate change initiatives”. It can be used in Outlook to draft and schedule emails.
Supports Visual Search & Video Previews: Bing also supports multimodal search. For example, users can upload an image of shoes from a magazine, and ask the prompt “Where can I find these shoes online?”. Bing chat also offers interactive video previews while generating responses, users can hover over the icon and get a glimpse without going to the actual link.
Personalized search results: Bing search performs excellent when it comes to finding the nearest restaurants, or events happening in the region. It uses the customer's geographical location and finds the most relevant matches to the search. Bing Maps is also available with Bing search, making it easier to visualize.
Apart from these, Bing Search also has a Microsoft rewards program through which users can earn points for searching and redeem them later.
Current Challenges :
Bing AI search suffers from limited context understanding in domain-specific queries, and cannot perform at par in advanced reasoning compared to Open AI or Google AI search. It can hallucinate when faced with complex queries.
As Bing search relies primarily on Bing search indexing, the results may not be accurate when there is a delay in indexing new or updated web pages.
It has limited integration with other products like Apple or Google Ecosystem
4. Perplexity
Perplexity AI is a search engine that provides contextually relevant answers along with source citations to users’ queries. The platform is integrated with a variety of LLMs including Open AI’s GPT, Claude, Mistral, and its own custom models. Perplexity is widely used by researchers, as it focuses on providing verifiable answers along with sources. It caters to a wide range of queries by using a diverse set of LLMs and aggregating their responses.
Key Capabilities:
Provides Source citations: Perplexity’s focus on transparency of responses makes it stand out from its competitors. It provides direct, concise answers to user queries along with links to citations/sources.
Dynamic LLM-Model selection: As Perplexity is integrated with a mix of different LLMs, it chooses the most appropriate model for a given query based on the complexity, domain, or depth required. For example, it can leverage the Claude model for safety-sensitive conversations, or GPT-4 for ambiguous, complex queries.
Freemium Model: The platform has a free version that can be used by students or researchers and a paid version for large-scale use by businesses.
It also has real-time web integration and can provide the latest updated responses.
Current Challenges :
The responses can be relatively delayed during high usage, as it queries multiple models in the backend and aggregates the final response.
The personalized search capabilities are limited for free tier users.
There are data privacy concerns, as it uses third-party LLMs and APIs that could store user data.
5. Arc Search
Arc Search is an AI-powered mobile browser that generates quick responses while prioritizing user’s data privacy. It is available across both IoS and Android devices, offering flexibility to users. It can be used by consultants, marketing teams, and researchers to quickly study market trends, social media patterns, etc.
Key Capabilities:
“Browse for Me” Feature: It has a browse for me functionality that inputs user queries, searches relevant information, and creates a customized web page. The user can simply read through the tailored web page created for them rather than going through multiple links or sites.
Data privacy: Arc Search focuses on user data privacy, and has very limited data trackers compared to search engines like Google or Bing.
User-Centric Interface: Arc Search is designed with a simple user-friendly interface, that allows users to change to ‘Reader Mode’ for a clean reading experience. It also supports customizable widgets for users.
Ad-blocks: It has a built-in feature to block advertisements, and pop-ups and allows user to focus without distractions.
Current Challenges :
While Arch search is optimized for speed and simplicity, the responses may lack sufficient information when faced with complex queries.
It cannot function offline, as it relies on internet connectivity for real-time data search
It lacks features for large-scale usage in enterprises and is limited to a particular niche.
Building Your Custom AI Search with RAG
RAG (Retrieval-Augmented Generation) is a technique where LLM models are combined with retrieval techniques to generate context-relevant responses to user queries.
A RAG-based architecture has two major components:
Retriever Component: This part takes care of retrieving the relevant information from the database for a user query. To build this, the first step would be to collect your dataset and convert it into vector embedding using APIs/embedding models. These vector data points can be stored in a vector database like Milvus that supports indexing, and fast retrieval speeds. Next, the user query is also converted into a vector and we can use similarity search to find the closest vector points.
Generative Component: The data extracted from the retriever can be passed to a Large Language Model like GPT or Mistral. The LLM can be prompted to generate a simple, summarized response from the extracted data.
The RAG system needs to be continuously evaluated through metrics like search relevance, response time, etc. If you are interested in building your own RAG, you can check out our blog.
Recap Summary
AI search engines have transformed how we search, shop, and research. All the 5 AI search engines discussed in the above blog can generate quick, personalized responses to queries and save users a lot of time and effort. Enterprises or individuals can choose their AI search based on the complexity of their queries, their need for privacy, and other factors.
Open AI is best suited for ambiguous or complex queries or large-scale use in enterprises.
Google AI is the go-to choice for real-time search and multi-model queries (to search through images and videos)
Bing search can provide the best localized search results while balancing both AI and traditional search.
Perplexity is the best solution for research purposes where transparency is crucial.
Arc Search is excellent for individual content creators or students who prefer speed and privacy.
You can also develop your RAG-based AI search engine for domain-specific queries by finetuning LLMs on your dataset.
Comparison Table:
| Category | OpenAI Search | Google AI Search | Bing Search | Perplexity AI | Arc Search |
| Strengths | Contextual, deep insights | Real-time breadth | AI + traditional combo | Transparency, sources | Privacy, design-focus |
| Features | Summaries, APIs | Real-time, multimedia | Conversational AI | Citations, lightweight | Visual, privacy-first |
| Limitations | Real-time data limited | Lacks contextual depth | Smaller index | Niche, small-scale | Limited indexing |
| Best Use-Cases | Custom AI apps, research | General, real-time | MS ecosystem, shopping | Fact-checking | Creative, private use |
| Scalability | Enterprise use | Global infrastructure | Enterprise use | Small-to-medium | Limited |
Top comments (0)