Brave announced a new AI-powered rich result on its search engine results pages that contains an AI summary of the content instead of the usual meta description.
In some cases, Brave Search Summarizer will display a summary derived from two sites together.
The summary is currently shown for about 17% of search queries and this number is expected to grow.
Brave Search Summarizer technology
The core technology of Summarizer is an internally developed large language model.
It uses existing technology developed to display knowledge graphs and featured snippets.
Brave’s announcement explained:
The Summarizer is not powered by ChatGPT or its backend systems; instead, it consists of three different LM2s trained on different tasks:
The first is QA (Question Answer): This form is used to try to extract a specific answer, if any, from text snippets.
Brave has been using LLMs for a while to improve search relevance, and this is an extension of what Brave Search was already in place to enhance the knowledge graph and featured snippets features.
The difference lies in the number and length of the analyzed text snippets.
After the QA extraction phase, the outcome candidates are further ranked with a set of zero-classifiers according to a variety of criteria (hate speech, vulgar writing, spam, etc.).
The final set of candidate text is finally processed by the summary/paraphrase model, which attempts to rewrite the input so that redundancy is removed and that language is kept uniform to improve readability. “
Brave summary misses
Brave up front that the system is not yet perfect and that it can make mistakes by combining unrelated snippets.
The advertisement stated:
Besides scalability, tremendous effort has been made to ensure the quality of abstracts generated.
However, since the paradigm is still in its early stage of development, there is the possibility of producing “hallucinations”, which blend unrelated snippets into a single result.
There is also the possibility of some wrong or offensive text, but our goal is to continue working on improving the models as our users’ feedback starts to flow in. “
This is something that can be seen in the search result for “Who is Danny Sullivan?“
Danny Sullivan is a very popular name, which may be why Braves Summarizer aggregates the details of several people into one summarization.
The summary humorously credits Danny Sullivan with winning the Indianapolis 500 and working as a chicken breeder.
This is Brave Search’s summary:
Danny Sullivan is an American technologist, journalist, and entrepreneur best known for founding Search Engine Watch in 1997 and launching Search Engine Strategies, one of the oldest commercial search marketing offerings.
He is also a former racing driver with 17 CART Indy Car World Series wins including the 1985 Indianapolis 500 and 1988 CART Championship win.
In addition, he attended the Kentucky Military Institute before dropping out to work as a janitor, New York cab driver, waiter, lumberjack, or chicken raiser.”
Here is a screenshot of the wrong summary result:
What’s notable about the summary in the screenshot above is that the data came from two locations.
Summarizer links to both source websites below, apparently using extracted favicon images.
Brave Research Summary links to sources
Summarizer snippets provide links to source Web sites.
Links are highlighted in red in the following example:
The link to Cars.USNews.com in the image above looks faded, with a gray circle to the left of the field displayed.
The gray circle is where the image icon from the website should be.
Below is a comparison of links to three websites.
What is notable is that the link to the Cars.USNews.com web page contains no code, causing the link to lose some visibility:
The Cars.USNews.com webpage does not have a favicon. a Favorite icon A banner image that is displayed in a browser tab and also used in other ways.
The Crutchfield and BestBuy.com web pages both have a favicon.
USNews.com does not have a favicon on this specific webpage.
They obviously omitted to link to a proper favicon on that webpage (even though they are on other pages in the same subdomain).
A new way to display search results
Bing is taking a different approach to incorporating AI into search.
It uses a system called Orchestrator that does the sites between the chat AI and the normal search index.
Brave is taking a more conservative approach because they don’t think the chat method is viable at this time.
They explained why they used the abstract method:
“Although the industry is generating a lot of hype around AI, we at Brave are not yet convinced that LLMs can replace research as we know it.
However, if used correctly, these new forms can help the user navigate results, which is the approach we take with Abstract.
Chat-like interfaces and oracle-based search are still unproven, and even today, we remain skeptical that they will be useful for all search tasks.”
How do you feel about the summary?
Many publishers and SEOs are not happy when Google replaces their meta descriptions with content extracted from a web page.
What Brave Summarizer does goes beyond what Google snippets often rewrite by generating an entirely new summary or mixing and matching content from two websites.
Considering one way, Brave Summarizer creates a context for the link to the source web page, which can lead to better traffic.
Then there is the situation where two websites are attributed to one summary. Who will get clicked?
Read the Brave Search announcement:
Brave Search introduces Abstract, an artificial intelligence tool for relevant composite results
