本字幕由TME AI技术生成
Npr a few years ago
Muti vegmen got a new laptop
Matti is working on his phd at bahh's university
Vimamarin
Jermany
And he wanted ted new new mesensenger bag
So he could walk around with it mattty sesearch for bags on gggle
But he wasn't happy
Other results
One of the pages i stumbed on
They would got ted lots of backpicks
And they were descriing them
But i've had very dissatisfied a discriforand after visitting them
These were the kinds of spami websites
That had a ton of ads and links
But not really any insights
Somewhat new infrustrative experience
Because we were used to just going on gggle kicking the first link
And it would look
And there got mati asking and google getting worse who provocative tive questions
Yes
Of course
There is certainly a lot of conluder
Now a time of sponsored links
That new ai generated box for some sesechch
Which sometimes work sometimes dozens said hmselphone sometimes wrong always confident them
That's how i approach ed lifeteran
I often find myself to scrolling pass that stuff to get to the what i consider like the real search result to the bottle
And we too want to know about the real stuff
The google search results
This is the indicator from plenty of money
I'm daring woods
And i'm waie in wang
Today on the show testing google search
There are many anedotal complaints about google
Not being what it used to be
Today we test that claim
And we bring it to google
Itself one way of figuring out the useleiness of gggle
Over time is to look at product reviews
When you search for a reviews for a new message of bag
You could capteorize the types of websites
You might get and on one side
There are a websites dedicated to careful reviews
Maybe they have have a bag specialist going through the prosing coons on the other side is a category that has less
You saw types
Wewesites
I content farm you could cit not there are useful content
And also not very trust words
If you looked ked this into detadetathat
Know know tytypes
Valthen Oh
Yes I do
And on that content farm we had some spam mi get and and yeknow our messenger
Baackshop
Marti
Vegman is aphd candidate in computer science
And for his research
He decided to investigate muttian
His colleagues collected the top twenty search results for more than seven of product review searches
And tracked these every two weeks for over a year between twenty
Twenty and and tracred they did p for google
But also for other search engines
Like being ying and ductducgo
So we ended up with hundreds of thousands of search pages
And then measured properties of these pages
How many words are on the page
How many images are
Are on this page
How many links
How long other links and so on how many headlines out
They out now google
Favor
Muttii found that out of the commercial search engines
They looked at gggle
Performed the best at filtering out spammi websites without much content
That said
Muttii found that certain legacy magazine websites have sections that turn out a lot of product review pages
Every day just to get advertizing and blink commission money
And these websites tend to get reked highly on gggle
They put out so much content over such a broad area
And all of them is what you would consider reasonably ality
So the very strongly signed
So the gggle ranks them highly
对对对
They are also very shadow
And they are very very satisfying
If you really want to have an get an informed decision
Martin is colleagues found that gggle would issue periotic updates to their search algorithms
And that would push low quality review websites off the results page
But then as these websites learn the new ways to game the search engine
They would creep back up again
This sounds would repeat somewhere or another
It sounds like a constant game of cat and mass se
And this is exactly what we call it in alfaapor too it seems there is an ongoing gggle between the content publishers that try to get rank burhhighly and google themselves who try to update to such engine
So what does gggle itself say about this cat and mouse game punting niic is the chief scientist for search at gggle
There's certainly uh an adversarial component do it so yes
Paniu confirms
There is this constant updating needed to keep search ch sults relevant
This has been sort of challenging since the the beginning of search and google started they introduced of speciges of algorithm called page ranks about uh using links as sort of votes of confidence in a particular site
And then aggregating it in a particular way
That give a very nice signal of reliability
And google getting worse over time
Um actually all our measurements uh h that that not get geworworse over time
We uh launch you a thousands of uses
Ers to ensure that the changes ges we making it actually
Tually the benefit of users since ce theh sppaper
They says google has added all kinds of additional measures of whether a pages higher low quality
We complement that chanof of algthomc chh
Agggimic and policies that allows us to take action against sites
That are trying to manipuate us
So this sort of this comprehensive effort
It's a fairly significant effort
That allows us to do these things
Google had a big updating march aimed and improving search quality by the way
If you're super interested in the interner working of search in nal documents that google actually leak recently
And this is been one piece in all this scrutiny that gggle has been under and among all this
That ve actually read matti's paper and puning points out that mati's paper is just focused on product of views
Not all types of search
There's just this tremendous uh idea of qualities we get argggers just seof them
Pundo says
Everyday fifteen percent of search queries are something googles never seen before
And so gggle is always updating its search tools even using artificial intelligence most noticeable to use these
These days is google 's io overview
It's this box at the top of some searches
Which generates
Answers to uses queries using a large language model kind of like chetgy bt now like all large language models
It's not always accurate
And this is occasionally sponse some unintentionally hilarious answers
Yeah
My favorite was an ai overview suggegesting users d glue to a pizza to stop the cheese from falling off
How do you know
Till youve try ed it
It
It probably works
It may have some side of it s now to be feared to gggle
This ai overview box is marked as experimental
But it hasn't been great pr pond
You says the algorithm of google core search engine has been using ai like deep learning
Since at least
Twenty fifteen
The doctor is generative ai makes it easier to create a kind of junk website
Well i mean
I think uh um that certainly true for any technology that it can have uh good and bad uses
Right now it's possible that with generavi the scale of the problem might uh uh might go up uh in in the future
Hh once again
I think we're sort of ready and willing to uh engage with that problem
It's like ai has heightened cat and mass problem
Yeah
Now it's a jaguwar versus copy bar problem big cats the world's biggest roughtly
But that said muti vvigman the computer scientist is feeling optimistic about the future of search
If only because of all this extra attention from academics and governments into how big tech operates plus for all his scrueny of ggggthe search ine did eventually help and find his messenger bag
My i phfound a very interesting in decide just about his pages down on on gggle that actually had like a lots of lots of of videos for all that facus he ended up
Settling on a plane black messenger bag
It was the world's greatest deep dive
Just to buy backpack
You know i respect that research
That to buy a backbag
We should mention is a sponsor
Npr
The sepseab was produced by julia
Ricchie
With engineering by new rage was facctacc by cy of quadis caken
Catton edits the show and the indicators of production of npr
