Revamps Google SEO Model For LLM Era

ChatGPT’s release on 30 Nov. 2022 was met with much fanfare and plenty of pushback. It quickly became clear people wanted to ask AI the same questions they asked Google—and ChatGPT often wasn’t capable of an answer.

The problems were numerous. ChatGPT’s replies were out of date, didn’t cite sources, and frequently hallucinated new and inaccurate details. Emily Bender, director of The University of Washington’s Computational Linguistics Laboratory, was quoted at the time as saying that AI search was “The Star Trek fantasy, where you have this all-knowing computer that you can ask questions.”

Perplexity initially hoped to build an AI-powered Text-to-SQL tool. But something different started brewing in the company’s Slack channels.

Founded in August of 2022, Perplexity the start-up stumbled into—and then raced towards—building an AI-powered search engine that’s updated daily and responds to queries by citing multiple sources. It now has over 10 million monthly users and recently received an investment from Jeff Bezos.

“I think Google is one of the most complicated systems humanity has ever built. In terms of complexity, it’s probably even beyond flying to the moon,” says co-founder and CTO Denis Yarats.

In the beginning, it was a Slack bot

Perplexity initially hoped to build an AI-powered Text-to-SQL tool, Yarats says, to let developers query and code for SQL in natural language. But something different started brewing in the company’s Slack channels—a chatbot that combined search with OpenAI’s large language models (LLMs).

Then, in late November of 2022, ChatGPT went public and became the fastest-growing consumer application in history, hitting 100 million users within two months. People were asking ChatGPT all sorts of questions, many of which it couldn’t answer. But Yarats says Perplexity’s Slack bot could.

“Literally in two days, we created a simple website and hooked it up to our Slack bot’s backend infrastructure, and just released it as a fun demo,” says Yarats. “Honestly, it didn’t work super well. But given how many people liked it, we realized there’s something there.”

For a time, Perplexity continued to work on its Text-to-SQL tool. It also created a Twitter search tool, BirdSQL, that let users find hyper-specific tweets, like “Elon Musk’s tweets to Jeff Bezos.” But the AI-powered search engine stood out and, within a couple months, became the company’s new—and daunting—mission.

How is AI-powered search possible?

This begs an obvious question. How did Perplexity, a company founded by four people (it has since grown to roughly 40) less than two years ago, cut through the problems that seemingly made AI terrible for search?

Two decades of failed Google competitors have proven “decent” isn’t good enough. That’s where AI offers a shortcut.

Retrieval-augmented generation, or (RAG), is one pillar of the company’s efforts. Invented by researchers at Meta, the University of London, and New York University, RAG pairs generative AI with a “retriever” that can find and then reference specific data from a vector database, which is passed to the “generator” to produce a response.

“I do agree RAG [is useful for search],” says Bob van Luijt, Co-founder and CEO of AI infrastructure company Weaviate. “What [RAG] did was allow normal developers, not just people working at Google, to just build these kinds of AI native applications without too much hassle.” He points out that the resources for implementing RAG are freely available on AI developer resource HuggingFace.

That’s led to widespread adoption. Weaviate uses RAG to help its clients ground the knowledge of AI agents on proprietary data. Nvidia uses RAG to reduce errors in ChipNeMo, an AI model built to aid chip designers. Latimer uses it to combat racial bias and amplify minority voices. And Perplexity turns RAG towards search.

But for RAG to be any of use at all, a model must have something to retrieve, and here adopts more traditional search techniques. The company uses a web crawler of its own design, known as PerplexityBot, to index the Internet.

“When trying to excel in up-to-date information, like news… we won’t be able to retrain a model every day, or every hour,” says Yarats. But crawling the web at Google’s scale also isn’t practical; Perplexity lacks the tech giant’s resources and infrastructure. To manage the load, Perplexity splits results into “domains” which are updated with more or less urgency. News sites are updated more than once every hour. Sites that are unlikely to change quickly, on the other hand, are updated once every few days.

Perplexity also taps Bidirectional Encoder Representations from Transformers (BERT), an NLP model created by researchers at Google in 2018, which was in turn used to better understand web pages. Google took BERT open-source, offering companies like Perplexity the opportunity to build on it. “It lets you get a simple ranking. It’s not going to be as good as Google, but it’s decent,” says Yarats.

Keeping Google at bay

But two decades of failed Google competitors have proven “decent” isn’t good enough. That’s where AI offers a shortcut.

“For Google, there’s a lot of constraints. The biggest is ads. The real estate of the main page is very optimized.” —Denis Yarats, CTO,

LLMs are excellent at parsing text to find relevant information—indeed, finding patterns is kind of their whole thing. That allows an LLM to produce convincing text in response to a prompt, but it can also be used to efficiently parse and then present information an LLM examines. You can try this yourself by uploading a PDF to ChatGPT, Google Gemini, or The LLM can ingest the documents within seconds, then answer questions about the document.

Perplexity essentially does the same for the web and, in so doing, it fundamentally alters how search works. It doesn’t attempt to rank web pages to place the best page at the top of a list queries, but instead analyzes the information available from an index of well-ranked pages to find what’s most relevant and generate an answer. That’s the secret sauce.

“You can think of it like the LLM does the final ranking task,” says Yarats. “[LLMs] don’t care about an [SEO] score. They just care about semantics and information. It’s more unbiased, because it’s based on the actual information gain rather than the signals Google engineers optimize for whatever reasons.”

Of course, this begs the question: can’t Google do this, too?

Yarats says Perplexity is aware of the difficulty of facing down Google and, for that reason, is focused on “the head of the distribution” for search. Perplexity doesn’t offer image search, cache old web pages, let users narrow down results to a specific date or time, or include shopping results, to mention just a few Google features that are easy to take for granted. He also believes Google will face problems linked not to its technical execution but its existing, and highly profitable, ad business.

“For Google, there’s a lot of constraints,” he says. “The biggest is ads. The real estate of the main page is very optimized. You can’t just say, let’s remove this ad, and I’m going to show an answer instead. We don’t have that. We can experiment.”

Source: IEEE Spectrum Computing