Google Image Search

0:09

Hi everybody. My name is Peter Linsley, I'm a product manager at Google, working on image

0:14

search. Now what we thought we'd do today was to run over some slides that I presented

0:19

at SMX West in February 2009. Just a couple of weeks ago. And the slides were very, um,

0:27

sort of a high level introduction to image search. So first of all I thought we'd run

0:31

through the presentation that I gave at SMX West and then afterwards I'll run through

0:35

some of the questions that came up that seemed like topics of interest to the webmasters.

0:39

Okay, so I'll start the presentation.

0:42

So first of all, our mission with Google Image Search is to organize the world's images.

0:49

We put a lot of focus on satisfying the end users. So when they come with a query, and

0:54

they have an image that they're looking for, our goal is to provide relevant and useful

0:58

images for that query. And of course the theory here being that if they find what they're

1:02

looking for and they enjoy their experience they'll come back and use us again. What I

1:06

wanted to get out of this talk as well was to start to engage a little bit more with

1:11

the webmaster community. If we look at what has come out of various conferences like this.

1:17

Where web search representatives from different companies have gone out and had a conversations

1:22

with the webmasters and found out what their pain points were and we found this sort of

1:26

ad hoc consortium came together and came up with things like the Sitemaps standard or

1:30

they came up with rel nofollow and well they came up with robots wild cards and things

1:34

along those lines. So one of our hopes in Image Search is that we can try and start

1:40

this dialogue and find out what sort of pain points you guys might have as webmasters.

1:45

Where we, you know, the likes of Google and also other search engine companies can try

1:49

and come together and try and enhance the end user experience by finding easier ways

1:54

for you guys to get your images both indexed and ranked.

1:59

So I'm just going to move on to the first slide that I had. I wanted to paint a little

2:03

bit of a picture of Image Searches. What they do and why they might be slightly different

2:07

to the kind of audience you might be used to with Web Search. First of all Image Search

2:13

appears in a lot of places beyond images.google.com. You've probably seen images appear in Universal

2:22

Search, so whenever you do a query like "pictures of San Francisco" there might well be a portion

2:27

of the results page that's dedicated to showing images for that result. And the theory here

2:31

is very much in line with our goal in Image Search which is that we're going to show you

2:36

those results when we believe that they are very useful and informative and relevant to

2:39

the query. Images also appear in other places like on Maps. You might have seen a little

2:46

row of images in Maps which come from our Panoramio.com property, which is a really

2:50

cool product if you haven't seen it before. So images appear everywhere across all of

2:55

our properties and we're really just trying to align it with when they match the user

3:00

intent or they enhance the user experience. Image searches also have a very unique search

3:05

behavior. They are very different animals to web searches. If you think about the paradigm

3:11

when they do a query, it's not so much about what's the first result. We don't really have

3:15

this sort of I'm feeling lucky paradigm. It's more about saying here's a query, well here's

3:19

20 images that you might like. And users can consume those images in a heartbeat. And if

3:24

the image they happen to like is at the bottom left hand corner or the bottom right hand

3:28

corner than so be it. They'll see that, there'll be something about that image that attracts

3:32

them and they'll click through.

3:34

The other thing that they do is that they search a lot of images. So there's a lot of

3:38

next paging going on, they'll go very deep looking for the image that they like. One

3:42

of the reasons why this happens is that a lot of queries we see are very subjective

3:46

in nature. So if you see a query like "waterfalls" then the waterfall that you like and the waterfall

3:52

that I like might be on two very different pages and there's no way as a search engine

3:55

we can figure out what you're looking for. So there's a lot of next-paging, users can

4:00

consume results very quickly and it's just interesting to think about what that might

4:04

mean for you guys as marketers. That it's not all about being in first position on the

4:08

first page.

4:10

There's also a lot of novel use cases in Image Search which might not be apparent. Users

4:16

use image search for inspiration. They want to get a haircut or a tattoo and they're looking

4:20

for ideas. So "Tattoo ideas" and then they go through the pages looking for some inspiration.

4:26

Quite often they'll refine their query, there's a lot of sort of exploring and this sort of

4:30

browse with intent. Users also use image search for shopping. They use it for research, health

4:37

queries, or sometimes they use it just to kill time, just for the fun of it. Another

4:42

really interesting use case that we've seen is using Image Search as a visual dictionary.

4:47

So there's a Googler in Germany who's learning German and he, if he hears a noun or a word

4:54

that he's not too sure what it is, he'll type it in. And he knows exactly what the word

4:57

means. Even though he doesn't look it up in a text dictionary.

5:01

Okay, so this is the slide on How Does Image Search Works. Simply put, as a webmaster you'll

5:08

see Googlebot come along and download the HTML as normal. Then what happens is, we pass

5:15

through your page and we look for references to images. And typically references to images

5:19

can come in one of two forms. It's either an href where you're linking to an image directly,

5:24

or its an inline image, in an image source tag. Then what happens is we come along

5:30

and crawl the images and then go through this process of classifying it. Now what we're

5:35

trying to do here is to figure out how to bucketize this image correctly and one classification

5:42

we do is to work out if its a photograph or not. Another one might be "does it contain

5:47

a face?" Other buckets might be things like, is it line art? Is it black and white? Or

5:53

is it an unsavory image that we want to put under Safe Search , only show when Safe Search

5:59

is disabled. So this sort of classification goes on and the reason we do that is we found

6:05

that image searchers really like to slice and dice their results. They like to do a

6:10

query and look at it and say "Well, these images are sort of nice, but I really wanted

6:13

to see just images with faces in them. " So if you've seen across the top of the results

6:17

page there's a blue bar which contains some drop-downs where you can actually filter the

6:23

results down to just photographs or just faces or just line art and so on and so forth. And

6:30

these filters tend to get used quite heavily, so we like to try and bucketize things off

6:36

so that they're shown in a more relevant context.

6:40

Finally of course the images are indexed. And that's where we squirrel them away and

6:44

we have an index of the image with all of the text that we associate with it, with that

6:48

particular image. Another part of this process is about identifying the duplicates. So if

6:54

you think about the way images are typically deployed online, you might put an image up

6:59

and a particular page will refer to it, another page might refer to it, you might have other

7:04

pages on your site that refer to it. Every now and again an image will get copied. And

7:09

maybe it gets copied as is or maybe it gets transformed ever so slightly. But as far as

7:13

the user is concerned it's still very much the same image. So the next process we go

7:17

through is one of trying to cluster all of the very similar or identical images and try

7:22

and treat them as one. And this is very much the same as the way things are done in web,

7:27

when web pages are analyzed for duplicates and then one sort of canonical winner is picked

7:34

out of that entire group. So the same thing happens with image search. We try and identify

7:38

all of the duplicates, and again, the main reason for us doing this is that when somebody

7:42

comes in and types blue widgets we really don't want to be showing them exactly the

7:45

same blue widget twenty times. We want to try and cluster those together and say, well,

7:48

here's one interpretation and here's another one. So there are multiple images, our goal

7:55

is to try and cluster these and figure out which is the best one. And at the same time

7:58

we have multiple pages that are including that image. And another task is at run time,

8:03

at query time, to try and figure out which one of these referrers makes the most sense

8:07

for this particular image that we've chosen. And the answer to this is we try and choose

8:10

the best one. We try to choose the best image that meets the user's intent more accurately

8:18

and maybe it's about size or maybe it's about quality or something like that. And the referring

8:25

page that includes that image is selected based on how good it is essentially. And that

8:30

could be one of many things such as its relevance to the actual query itself.

8:37

And finally, ranking is performed on a whole lot of signals. And typically we don't go

8:42

into the details of the signals but its very much like web, there's more than one signal

8:48

that we use to try and figure out what the most relevant image would be. So the next

8:53

slide, is on best practices. You're sitting there thinking, "That sounds great, I've got

9:00

good images that I think are going to be useful for your users. What can I do about it?" Probably

9:05

the best bit of advice we can give is to really focus on the user. Now you might be thinking

9:10

what exactly does that mean? What can I go out and do tomorrow to focus on the user?

9:15

The answer is pretty simple. If you think of a user that comes to Google Image Search

9:20

and what they might be looking for, and if we take one use case such as coloring pages.

9:24

Maybe they're looking for a site that has a lot of coloring pages and they choose to

9:27

use Image Search to get there. Then the first thing they're going to do is come along and

9:32

type in "coloring pages" and they're going to look at the results. And maybe they see

9:37

something they like, maybe they don't, they might hit next page a few times. And all of

9:42

a sudden one image will catch their eye and they like it for some reason. Maybe its just

9:46

the quality of the image itself, or maybe its the snippet, maybe there's something about

9:50

the size or host name that sort of draws their attention. Maybe its coloringpages.com "I

9:56

know that site! I'm going to click through, I trust it". Then they land up on your page

10:01

and the question is: What sort of experience are you immersing them into? What sort of

10:05

experience are they getting now that they've come to your page, given that they were looking

10:09

for coloring pages. Do they see the coloring page they just clicked on above the fold?

10:15

Is it large enough? It's one thing to send people to a coloring page page where you show

10:21

them very small thumbnails and its another thing to say this is what you just clicked

10:24

on, here it is, here's some descriptive text, here is some related pictures, here is some

10:33

comments from the users, ratings and all sorts of things. It's really about immersing the

10:37

user into a very image-centric experience. These are the kinds of landing pages and the

10:43

kinds of images we've observed that our users tend to like. And again, our intent is to

10:49

try and match up the intent of the end-user. So focus on the user, and high quality images

10:55

are always good. If you're taking photographs to put on your site go and buy a digital SLR

11:00

and learn how to use it, get a good lens,take really nice high quality pictures. You don't

11:06

necessarily have to show, take up the whole screen with the photograph or the image, but

11:10

just you know large enough is usuallly what the users like. Above the fold and plenty

11:16

of descriptive text. And fundamentally the impetus to all of Image Search is a text query

11:24

and the extent to which you have a lot of descriptive text that is on topic and talks

11:28

about what is in the image, maybe you want to expose EXIF data, maybe you want

11:33

to talk about when the image was taken, maybe it has a nice title across the top. All those

11:37

sorts of things are really good clues for us to figure out when an image is relevant

11:41

or not. But more importantly its useful for the end user. They can read the description,

11:47

read the caption and learn a lot more about the image.

11:51

The last slide I talk about resources. We have Webmaster Help Centers where you can

11:55

go and read a lot more about Image Search. We also have forums where you can post questions

12:00

about Image Search. We really really encourage webmasters to come to these forums and post

12:04

all of their questions or concerns. We monitor these very closely, we pick up these concerns,

12:11

and we'll take a look at them. There's also Web Search Help and Forum for end users so

12:20

if you're an end user of image search and you have questions you can leave them there.

12:24

The other thing is to monitor the Google official blog because that's where we typically put

12:28

our announcements of new features and changes and news and what have you, specifically around

12:33

image search. So that was the end of my presentation that I gave at SMX in a nutshell.

12:40

So at the end of the presentation we had Q&A and I wanted to pick up on a couple of questions

12:45

that came up during that time. The first question was "Hey, you guys mentioned large images

12:52

are a good best practice but I have a concern with that because I don't want to load up

12:56

the really really large version of the image that I have because it takes the page a whole

13:00

lot of time to load up. How do I manage the trade off there?" So I think the answer to

13:07

that question is to show an image that is large enough. Typically two-thirds of a screen

13:12

is one sort of rule of thumb. The point here being that users tend to like to actually

13:17

be able to see the image as opposed to it being a very small thumbnail. So a good way

13:22

to get around this, to allow users to see the larger version if they wanted to is to

13:28

turn your image into a link to either the large image itself or another HTML page that

13:33

includes a larger version of the image. Ultimately, nobody really wants to see an image that is

13:37

larger than the browser size. So, another question that came up was about Analytics.

13:44

Somebody was saying "Hey, can I get Analytics information from the traffic that's coming

13:49

from Image Search?" The answer is "Absolutely". In the referrer string that we send across,

13:55

well, that the browser sends across, both the query that ranked the image plus the image

14:03

itself is sent in that string. So one slight difference with image search, of course, is

14:07

that we're not necessarily sending people to that page as much as we're sending people

14:11

to your image on the page and there could of course be more than one image. So by passing

14:16

it by the referrer string you should get all the analytics you need to know. What query

14:20

sent the user there and what image sent the user there.

14:24

So that about wraps it up. I hope you found this useful. By all means drop all of your

14:30

questions in Webmaster and Help forums and we'll swing by and take a look. Thanks.