|Head of Google Search Quality denies brand bias in Google Instant|
In a short interview posted on Fast Company, Amit Singhal, the head of Google's Search Quality Team, denies brand bias in Google Instant.
Top Google Engineer: Google Instant Has No Brand Bias [fastcompany.com]
Mon Dec 13, 2010
|"What we do at Google and what we've done for years is to not inject any subjectivity into these algorithms," says Amit Singhal, Google Fellow and head of the company's search quality, ranking, and algorithm team. "We didn't want to introduce any bias into the mathematical modeling--our modeling is predicting, given a letter, what's the probability of completion." |
"Completion" here is referring to the Autocomplete feature of Google Instant, which Singhal says is completely mathematical.
The report doesn't discuss the algorithmic factors which might cause brands to come up as frequently as they do.
Thanks to Barry Schwartz for bringing the blog post to our attention in Search Engine Land [searchengineland.com].
Their Doubleclick team is also denying the existence of the drive-by malware attacks taking place through their own network, so I'd say this is par for the course with Google these days. Deny, deny, deny.
Sure, we didn't mean to collect that wifi data either, a bit of accidental but sophisticated coding just happened to appear with our image capture stuff.
And the camera cars accidentally sprouted antennas to collect the accidental wifi data?
Actually, I tend to believe him. In google.com.au, the first letter typed brings up different results and not all are commercial brands, e.g.
A = AFL (one of our sports sites), next ANZ (a bank), ATO (our IRS), then Amazon.
B = BOM (Meteorology bureau), Bunnings (like Home Depot), Bigpond (ISP), Big W (Woolworths)
I can't raise Google Instant in google.co.nz but when I try the UK site, I get the UK predictions, e.g. Argos before Amazon.
"The report doesn't discuss the algorithmic factors which might cause brands to come up as frequently as they do."
That Robert - is because there are none. Dont you know that? :)
"whats the probability of completion"
Well - probably something to do with how many searches are performed for any given phrase which may result given the letters so far.
So - possibly - the more popular something is - the more hits it will get. The more it gets - the more popular it is - so the more hits it gets and so on.
Algo wise - this - unfortunately for small business - could favour the big brands.
So without bias - they get what they want anyway no?
Hmmm - if we copy the words from the auto complete list into the keyword tool - exact match - sort by number of searches (global - but may be US) - then maybe we should see the data they are using.
|"What we do at Google and what we've done for years is to not inject any subjectivity into these algorithms," |
|"We didn't want to introduce any bias into the mathematical modeling--our modeling is predicting, given a letter, what's the probability of completion." |
I've been writing legal jargon for years. This guy is expert at it.
I believe that they did not tweak the algorithm to seek out brands ... no mention was made of the data used by the algorithm ...
They can say that their organic search algorithm is not bias and they do not inject subjectivity ... they still remove websites manually, set up penalties that catch specific website types etc.
It is impossible for them to successfully deny that they do nothing to influence the output of the mathematical programs they create. They're fiddling with the data they give these mathematical programs all the time.
|So - possibly - the more popular something is - the more hits it will get. The more it gets - the more popular it is - so the more hits it gets and so on. |
My thoughts exactly. Show people only what is already popular and it therefore becomes more popular. A snowball rolling down the hill.
I use an analogy to songs played on the radio. Play the same old song often and long enough, and that's what people request. How does the new artist get on the air?
I had to shop on-line recently for a specialized tool for my company. For fun I tried Bing. Wow - found some great companies I'd never seen before on Google and they had great prices and delivery.
I don't need to see Amazon in the top 5 for every product I search for. I can go there directly and search.
I hope the EU investigation into Google turns up some facts.
|Show people only what is already popular and it therefore becomes more popular |
+1 to that and Mr Fewkes.
Google have always said that they are shaped by intent.
I have two concerns with Instant - first that it has the possibility to shape and narrow intent, and to direct rather than reflect. There are half the suggestions there used to be - that's a huge amount of diversity being lost there.
Second, there's a fundamental difference to my mind between typing a search term and typing a few letters and selecting one of five choices. Once you start selecting you've started to become a browser, NOT a searcher.
The only statistically significant data I've seen about organic traffic changes following Google Instant shows an increase in video click-throughs. That's all.
The long-tail/short-tail comparison studies show only minor differences in some verticals only, and those differences are not always in the same direction from test to test! Some studies say long tail searches went up a tad, others say they went down a smidge.
I conclude that the world as we know it has not ended. It's certain that my smaller business clients noticed nothing.
One issue I've seen with Instant is that it doesn't always get things quite right.
Eleven or twelve letter search term, depending on the singular or plural version. Instant seems to have an issue with the first four letters which also happen to be the first four letters of a clothing company catering to young women.
I don't know how many times I've entered the full 11/12 letter string, only to be provided with results based on the first four letters which obviously aren't what I'm looking for.
Do I type fast and hit "enter" or a search button? Probably more than systems are designed for. Collateral damage, it appears I have to emulate "normal" users to get intended results.