SIMS 141 - Search Advertising: Dr. Hal Varian

Uploaded by UCBerkeley on 20.08.2007

Doctor Hal Varian.

HAL VARIAN: Thank you. Actually I started working for
Google almost three-and-a-half years ago now on a consulting
basis and helped build a group there called AdStats at one
time we called BizStats I guess, but now we call it
AdStats We do all of the analytic work on the ads engine.
So that includes forecasting, analysis, experimentation,
tweaking it in all sorts of different ways.
So what I want to do is today is just give you a little
outline of how the advertising model works at Google and
more or less, at Yahoo.
They aren't so terribly different.
I'll bring out some differences as I go through the talk, but
I'm primarily familiar with the Google model.
So if you look at online advertising there's basically
three different categories: there's banner ads, and an
example of a company that sells banner ads is DoubleClick and
what that is are those images you know, those so called
annoying images that show up on your web page.
Maybe they flash, maybe they pulsate, whatever.
And they're typically a standardized shape.
The industry got together at the very beginning and picked a
few shapes so that people could build their ads to that shape.
And they tend to be loosely related to the content.
If you go to a page that's about automobiles, hey,
you might see ads for cars. If you go to a page about
Hawaii, you'll see ads for tickets or something like that.
But those are all sort of human targeted to the content that
is available on that particular website.

The other second model is what we call context linked ads and
in Google it goes by the term Google AdSense.
And what that is are typically text ads, although there are
some image ads and the text ads will be linked by textual
analysis, by computer analysis of the contents of that page
and the ads would typically have something to do
with the contents. So if you look in the New York
Times and you happen to be looking at a story about
Italian cooking, they'll be ads for Italian food supplies:
olive oil and parmesan and all that kind of thing will pop up.
So they'd be contextually-based ads and they work pretty much
like the search engine works where they look at the words
and the keywords that the advertiser uses or look at the
contents in the page and they'll do some matching.
I'll give you some more details on that in a few minutes.
And finally, there are search linked ads and that's Google
AdWords and those are ads that are related directly
to the search terms.
So you'll type in a search term like Honolulu and you'll get
ads for flights to Hawaii. OK?
So that's the three differences.
And what I'm going to talk about today are these two here.
Not so much about the banner ads.

OK, so here's a couple examples.
This is actually New York Times table of contents
in this particular case and here's a banner ad.
It's about Vonage, the broadband phone company.
Doesn't really have anything to do with the
content on that page. It just happened
to show up there.
Maybe they bought-- they said, we want to show this ad on
the contents page of the New York Times.
These ads down here are the content ads.
It says up here, ads by Google and it describes some of these
particular textual-based ads or the same kind of ads you'll see
in the search context but they're related to some of the
stories that happen to show in the context.
It picked up and showed those ads because there were
stories about those topics in the table of contents.
That's kind of interesting in this particular case you say,
well this looks like kind of bad placement because these ads
are way down at the bottom of the page, but if you think
about it, if you talk to a UI expert for example, you can
find out that well people scroll through the page and
they get to the bottom they say oh yeah, I meant to
buy some CDs today.
So they'll click on buying CDs So bottom of the page
is a pretty good placement for these kinds of ads.
Then here's an example of the search ads.
So what happens there is I typed in the queries search
engines and I got back a bunch of search results.
These are the search results here, one, two, three.
And I got back some ads.
So these are sponsored links if you're listed along here.
And then up here is a sponsored link that we
call a promoted ad.
And I'll talk a little bit later about how
ads become promoted. There's a whole machinery
that's setup to rank these sponsored links.
So we try hard to get relevant links at the top and if we have
an ad that's performed really quite well in the past we'll
promote it and stick it up over in this position here.
So not all ads get promoted, only some ads.

Let's see, what else is there to say about that?
Well of course, part of the reason is that your eye very
naturally goes to this place here, so this ad has a very
high probability that somebody will click on it and these ads
have lower probabilities pretty much monotonically
as you go down the page.
So high probability you'll click here, somewhat lower
probably you'll click lower, lower, lower, lower.
In Google there are eight ads in the right-hand side and up
to three ads in the top of the page.
OK, different search engines have different configurations.
AOL uses something different, Yahoo uses something different.
So that's the layout and that's how we do it.

OK, so let's talk about the search link to advertising
because that's the first one, the most interesting
one in some ways.
The nice thing about search-linked advertising is
when you type in that query: search engine, hey, you're
interested in search engines.
Right? I mean that's why you typed that in.
You're interested right then.
So those ads that you see over in the right-hand side tend to
be very, very relevant to your interest.
Because the advertiser who bought those ads,
bought the keywords. In that case, search
engine, most likely.
And that results in very good performance in terms of cost
per acquisition That's marketing speak for cost per
sale or cost per customer or something like that.

In our case I'm referring to the advertising
expenditure per sale.
So in the ad industry there are all sorts of terms: CPM, cost
per thousand impressions, CPC, cost per click.
Cost per acquisition is the one that matters most because
ultimately if you have something to sell you better
hope that the profit you're making on that item that you
sell is at least as large as the cost it took you to
acquire that customer. So the cost per acquisition
is one way to measure that and search-linked advertising has
worked very well in that respect.
Has very rapid growth over the last few years.
Has become kind of the best model or most exciting model
for advertising on the internet, but it's still a very
tiny fraction of the entire marketing expenditure.
Total U.S. advertising spent depending
on how you count, I'm going to show you, break this
down on the next slide.
It's about $220 billion that companies are
spending on advertising.
It's been growing at maybe 1-2% a year, over
the last few years. It's highly cyclical as
the economy is doing well, people advertise more.
When things aren't so good the advertising drops off.
But roughly speaking, 1-2% a year.
Online advertising is only about $10 million now.
That's only 5% of the total, but it grew at 26% last year.
So, it's growing very, very rapidly.
Looks like a very strong performer, but it's
got a long ways to go.

Advertising spend so if you look at it you've got-- I just
took two analyst reports of this market and put
their numbers down. They agree remarkably well for
analyst reports, usually these numbers are all over the map,
but according to Morgan Stanley about $65
billion on newspapers. According to Media
Metrics about $57 billion.
I think the difference there is how you count classifieds
because in newspapers there's going to be the regular
ads that you see in the newspaper then there will be
classified ads, there'll be supplements, Sunday papers.
There's lots of different ways to slice that pie up.
Direct mail, they pretty much agree.
Broadcast TV, radio, cable TV, magazines, yellow pages, for
some reason there wasn't a number for yellow pages here.
And then the internet and somewhere between now $220
million is probably a pretty good number for total
advertising expenditure in the U.S.
So as you can see we're still very, very tiny compared
to other things. There's a lot of
room for growth.

And hey, what do you know?
Stock market recognized that and Google went public in
August 2004, about $85 a share.
Went up to $100 on it's opening day and now it's up over $300.
So there's a lot of beliefs, hopes, expectations, whatever
you want to call it out there that the model is going to
continue to grow and generate revenue.

All right. So how's it work in detail?
So what you do is you bid on search terms.
And let me be a little precise on that.
We distinguish the query, which is what the user types
in to the keyword, which is what the advertiser buys.
And there's an engine there that maps the
queries to the keywords. so it says, oh this query, then
these are the keywords that are relevant to that query.
And it does things like stemming and correcting
for plurals and things of that sort.
So there are matches at a conceptual level, not
necessarily at an exact word level.
You can choose if you want to have an exact match,
so I want to match this word and only this word.
Or you can choose to have a broad match, which you'll do
some kinds of synonym expansion in that stuff to give you
a match that's in the same general category.
So that's the choice that's up to the advertiser.
You can also choose a negative match.
You can say socks, but not red.
You can look for socks, but not red socks.

And then what happens is you can use a traffic estimator
and you can get an estimate of how many clicks you'll get.
So you say, OK, if I want this keyword and I bid this much
then it's going to get me about this many clicks.
So you have an idea of how much money you'd have to spend
to get that many clicks.

And then what happens is when somebody types in a query and
a query matches up with some keywords then there's a number
of ads that match typically and those ads are ranked.
Now there are different ranking functions that
are used in the industry.
In Overture the highest bidder gets the first position,
second highest bidder, second position and so on.
And then there's something that we don't quite understand
that they do get some preference to exact match.
So exact match gets first.
And then they're ordered by bids.
And then broad match gets second and they're
ordered by bids. Something like that.
I don't know exactly the algorithm that they use, but
it's more or less like that. That's what they said in public.
Google, we do it a little differently.
What we do is we rank by bid times of predicted
click-through rate.
So we have an engine, which tries to estimate the predicted
click-through rate for every ad and then if it's a really good
ad that's gotten a lot of clicks in the past, that's
likely to do well then it's going to get pushed
up to the top. So this is the idea that
the most relevant ads get displayed at the top.

Another way to think about it is you think about
click-through rate, that's clicks per impression.
How often the ad gets clicked on if it's shown.
And if you multiply that times the bid, which cost per click,
cost per click times clicks per impression is
cost per impression.
So we're ranking it by how much the advertisers willing
to pay for an impression.
And that's for showing this ad to somebody.
And that's relevant because for a lot of other media: TV,
radio, newspapers, magazines, everything's priced on a
cost per impression basis. So advertisers tend
to think that way.

OK, so you think about it, Google sells ads impressions.
That's what we've got to sell.
We've got the fact that somebody typed in this query
and there's a page of search results and there's some ads
that might be relevant to those search results, so we've got
space to show those ads.
That's what we've got to sell.
And the seller, he cares about how many sales he gets, right?
Doesn't care about whether anybody sees the ad or not, he
cares about whether they buy.
And kind of as an intermediate case, he might care about
whether they click because if they click well, that's an
opportunity they might buy.
So it's kind of like a storefront.
You know, you walk by and you see the store.
Well, so what?
You walk inside the store then the vendor has a chance to sell
you something and so that's a lead, something of that sort
and then ultimately you might make the sale or not.
So what we do is we look at this incentive problem
where Google's got impressions to sell.
Seller wants to pay for clicks or really conversions and so
we rank advertisers by how much they're willing to pay for
impressions since that's what we've got to offer them.
And this is just what I said earlier, you look
at the value per click.
So how many sales you get per click times the clicks per
impression is the value per impression and we're tending to
give the best positions to the guys who are willing to
pay the most for them. Have the highest value
for that impression.
Maybe because they are willing to pay the most or maybe
because they've got a high probability of click.
Because they've got good creatives, good texts,
good brands, whatever.
They've matched the users interest well.
But the trouble with this is sometimes advertisers
might want to show too many impressions.
If you think about it, I mean, after all if you're an
advertiser you don't really care so much that you are
relevant Google cares that you're relevant because we want
to provide a good service to the customers, but the
advertiser just wants to show his ad on the chance you might
walk into the store or might click on the ad and
go to his page.
So one of the things we also do is if there's too few clicks
per impression, if people don't seem very interested in this
particular ad then we'll disable it.
We won't show it.
So we have a threshold that says if we get this many
clicks per impression we'll show the ad.
If not, we'll disable the ad.

So how do you estimate that predicted click-through rate?
Because that's fairly important for us.
So the model is this little what you might call in
statistics an independent effects model.
Multiplicative effects model.
You have actual CTR is going to be a position effect.
So you get some clicking just from being in the first
position and then there's an ad effect or advertisers specific
effect or position specific effect.
And what you can do is to figure out the position
specific effect, just the propensity to click on
positions on the top of the page, you can randomize the
ads that are shown there. So you just show random ads
at the top of the page.
See how many times people click on that position just by virtue
of it being first and that gives you what we call a
position normalizer and then you look at the ad effect.
So you try to estimate the characteristics of the ad
that influence the probability that gets clicked on.
And there's a big computational engine that does that.
I can't really tell you how it's done, but I will tell
you it's quite complicated and involved and requires
a fair amount of analysis.
We think of that not only as this predictor for
this ranking, but also as a quality signal.
Because if people click on the ad a lot then we think it's
likely a pretty good ad because it gets all those clicks.
So well, the last thing just says what I mentioned.
That this isn't an easy thing to do.
It requires a lot of computation and a fair
amount of analysis.

So then there's another kind of interesting twist, that the
price you pay for the ad-- so when somebody clicks
Google gets paid by the advertiser for that click.
The price you pay depends on the bid of the
person below you.
Not on what you bid, but on the bid of the person below you.
And in Overture it just directly depends on it.
You pay the bid of the person below you.
How much they bid per click.
And in Google you pay the minimum necessary
to keep your position.
Remember we don't rank just based on bid, we rank
based on bid times PCTR.
So the amount of money you have to pay to keep your position is
enough to keep your bid times your click-through larger
than the guy below you's bid times his click-through.
So we just adjust it down.
Why not pay your actual bid?
Well then what happens is you'd see people
adjusting all the time. You don't want to pay more than
necessary to keep in this position, so if I'm paying $1
and the guy below me is paying $0.80 well I'll notice this and
lower my bid down to $0.81.
So since they would want to do that anyway the ad
words discounter just does it for them.
And it's very much like the proxy bidding on eBay.
Do we have eBay users out here?
Yeah, so eBay is the same thing.
Highest bidder gets it, pays the second highest price.
Here it's pretty much the same idea.
You don't need to constantly inspect the bid landscape, you
just let Google do it for you.

So that's similar in spirit to the proxy bidder on eBay
as I just mentioned and what's called a Vickrey
auction in economics.

How many people have heard of a Vickrey auction?

All right, that's a reasonable number.
OK, so how does it work?
So I'm going to move away from the Google model for a moment
and just think about a regular old auction, where I'm
an auction off this half drunk bottle of water. OK?
You can find these on eBay, I'm sure.

The way we do it is we could have an english auction and
people just yell out a bid.
You know, $0.10, $0.12, $0.23 and then we'd
sell the item that way.
We'd keep going until only one person was willing to bid.
So what happens if you think about it, the person who values
the item most highly gets it.
But what price does he have to pay?
Well he only pays the value of the second highest person.
I only have to beat everybody else.
The price will get bid up.
If I thought this, in fact, I do, I'm kind of thirsty.
If I thought this was worth $10 to me I'd only have to pay the
maximum price that anybody else would be willing
to bid out there. That's a standard
english auction.
What a Vickrey auction is, is it says, let's do the same
thing, but instead of iterating we'll just have everybody write
down their value for this item on a card.
We'll pass all those cards in.
We'll take the highest value on those cards.
That's the person who wins the object and he only has to pay
the second highest price.
OK, the second highest number written on the card.
So that's the VIckrey auction.

So it turns out you can show-- and I think I'm going to show
this in the next slide-- that it always pays to be honest
in the VIckrey auction.
You don't want to game it.

The best thing to do is to write down your true value.
And I'll show you this on the next slide, I think. Yes.

And why is that? Well here's a little
bit of logic.
v is your value, b is the bid of the second highest bidder.
You don't know whether you are second highest or not.
Maybe you're highest, maybe you're not.
And you're going to get v minus b2.
Value minus the bid of the second highest bidder times the
probability that your bid is bigger than the second
highest guy OK?
Maybe it is, maybe it isn't. Depends on what you bid.
So think about it.
If you're getting positives value from the item, if v is
bigger than b2 then you want to make that probability
as large as possible.
You know what? No matter how hard that
probability tries the biggest it can get is one.

And if v is bigger than b2, if you set your bid equal to v,
if you set your bid equal to v then indeed v will be greater
than b2 with probability one.
And if v is less than b2, If you value it less than the
second highest bidder, then you want that probability
to be as small as possible.
But if your bid equal to v then your bid will be less than b2.
So that probability would be zero.
So either way it works out so that's in your interest
to report your true value.
Whether your bid gives you a positive surplus or negative
surplus in economics chart.
So it turns out that ad options, the way I described
it, don't quite have that properties, not quite the same.
Requires a little more analysis.
But the analysis turns out to ne similar, so if we went
through a more involved analysis we'd have
a similar outcome.

So this example of what we call an economics mechanism design.
And the way an economics mechanism design problem
works is people report something to the center.
They report their values or their costs or there interests.
The center kind of figures out what the best outcome is
according to some criterion and then it goes back and it
says, OK, this is the outcome we'll have.
You owe me $10 or you owe me $7, you owe me $9.
OK, so what's happening is they're all reporting
their values to the center.
The centers making a choice and then it's assigning
some payments based on the choice the agents report.
And there's a problem that economists call
incentive compatibility.
How can we do this in a way so that people will be
happy with the outcome?
That they are willing to pay however much they charge to
get the outcome that the center ended up with.
Well it turns out that's a very intensively studied
problem in economics.
The auction that I described a few minutes ago is
an example of that.
And the Google Ads auction is another example of that.
It turns out that you can do quite remarkable things with
that economic mechanism design and I will say that it was
mostly of theoretical interest for the first 20 years or so
that this subject existed.
But then when people started doing all the auctions of
airwaves and eBay and the Google AdWords auction, all
this other stuff, it suddenly became a very practical topic.

And now in computer science even there's a subject call
distributed algorithmic mechanism design, which is all
about how do you build these mechanisms to achieve
some desired outcome.
Make sure everybody is behaving in an appropriate way and also
is computationally effective.
Extra problem that economists hadn't thought about too
much, but computer scientists were quite interested in.
Now I claim that the Google auction is the biggest
auction in the world.
In fact, according to comScore there are about 4.8 billion
web searches in the world in total in July 2005.

Roughly half of those searches display ads.
That's about 2.4 billion ad auctions per month.
And I will say that those are pretty conservative estimates.
So when comScore says this they're basing
it on their panel. That's Google, Yahoo,
AOL, everybody.
So this mechanism for ad auctions is the biggest
auction around. Despite what eBay claims.
It's much bigger.
There are hundreds or thousands of these going on every second.

Now from the seller, if you would like to become a Google
AdWords seller it's very easy.
Costs you $5.
And I think it's about the same at Yahoo.
What you need to do is you choose your creative.
That's the marketing term.
You choose the text of your ad.
It's a standardized size.
You have a certain number of characters for each of the
three lines and there are various tips available.
For example, it turns out that plurals get more
clicks than singulars.
So if you're in the business of selling diamonds you probably
want to choose diamonds as your keyword rather than
diamond, right? Because with diamonds you get
all those guys who are going to propose and with diamond you
get the kids who's doing his high school report
on carbon, right?
So you look at this and there are other little tips about you
know, action items, buy x here is better than you know,
x is available here. And active versus passive.
You know, all this stuff. Then there are search engine
marketers out there who will help you, give you tips.
We give a lot of tips on the Google page and we actually
have an optimization program that will help you tune your
creative and pick your keywords.
So then you actually have to choose your keywords: exact,
broad phrase, negative-- I guess I mentioned
all those before. You could only go for exact
match so if you had for example, a trademark word.
I mean, where you own the trademark or a particular
phrase that had a particular meaning in your context.
You might want that exact match, but I think most people
want many of there campaigns to use broad match because it
gives you a broader set of impressions.
And then you have to follow certain ad guidelines.
You can't lie, and you can't advertise tobacco on Google
and hard liquor and a few other things like that.
There are certain things you can't say.

Then they're reviewed by a human to make sure that you
followed these ad guidelines and then they get displayed.
That's it. You just buy your keywords and
for most people I will say, usually the creative isn't so
hard because the text is fairly limited.
It's like writing a haiku. You don't have all that many
characters to deal with and most people have an idea of
what is they want to sell. There's a little bit of effort
that goes into that, but in many cases the challenging
thing it's choosing your keywords.
Because a minute ago, I said you wanted to get diamonds
rather than diamond, but you know, not so clear
because it depends. Maybe diamonds cost a
lot more than diamond. Because everybody has
diamonds to sell is bidding on that keyword.
So you have to compare your value to the cost of the
keyword and think about how many clicks you're going to
get and how much the keyword costs you.
You might want to have a more obscure keyword.
Maybe you'll pick anniversary or maybe you'll pick whatever
it is-- 50th anniversary or something rather than
picking diamonds directly.
And then you can implement conversion tracking.
Conversion tracking is something that Google provides
that says if you want on your check-out page you can put a
little image and then we'll record those conversions and
track them for you and give you a nice report every month about
how many conversions you got and how much you spent and so
you can look at your cost per acquisition or cost per
conversion by doing that.

So you choose your bid, how much you're willing to pay,
your maximum cost per click and you do that probably by looking
at the probability of a sale, which you'd have to experiment
a little bit with different creatives, different things to
figure out how likely you are to get sales.
What your profit would be per sale and you can use the
traffic estimator you can use the discounter and so on.
And then you set a budget because sometimes you might
get a huge number of clicks.
If you're advertising on Hurricane Katrina or something,
all of a sudden, bang, you get all those clicks.
So a lot of people want to set a daily budget so they don't
spend more than they intend and then you let it rip.
So it will take you about 10 minutes to set one
of these campaigns up.
And you can sell whatever you want to sell.
As long as it fits within the guidelines.

Now the AdSense program is a little bit different.
These are the content ads. We still use the keyword
matching engine to pick up the ads that are related to
the content in the page.
But you think about it, there's lots of content on a page.
You might have two or three different stories or might
be different kinds of text.
So it's not as easy as the search link,
the search problem. You've got to put a more
natural language analysis in there to try to figure out what
the page is really about to try to show ads that are relevant
to the content in the page.
The auction needs a little bit of tweaking because again, in
terms of the layout what you think happens in search is
pretty much what happens. The top ads get the most click.
The right-hand sides get the second most and they pretty
much go monotonally down the page.
But with the content ads it's quite different.
There's layouts across the top of the page, down the
right-hand side, at the bottom of the page, lots
of different layout forms.
So yo need to control for that kind of thing when
you're dealing with the position effects.
And typically you have lower click-through rates, lower
conversion rates and so on on the content ads.
The reason is, as I mentioned earlier, if you're interested--
if you type in hotels in Hawaii, we know pretty much
that you're interested in seeing hotels in Hawaii.
If there's an article that you're reading in the New York
Times about hotels in Hawaii well, you may or may
not be interested. You're interested enough to
look at the article, but you might not be immediately
interested in buying.
So the clicks aren't quite as likely to convert in the
content ads and so there's an adjustment that's made for that.

So what do you need to make this all work?
You need an engine to match the ads to the queries and
that involves a fair amount of linguistic analysis.
You've got a promotion policy-- oh, yeah, I
didn't mention that.
I mentioned at the beginning that those ads were promote
were deemed to be very, very high quality ads and basically
we use several indicators for quality ads, but it's driven by
things like click-through rate and performance and so on And
then you need this auction engine, which just ranks the
ads by the cost per impression, records the clicks and then
determines the payments that people make.
So it's basically all those computers that Google has,
a good chunk of them are devoted to the ad system.

And then the question is, isn't it amazing that Google
is able to make money?
The average cost per click-- I'm not going to give you
Google's numbers, but in the industry it's around,
you know, $0.50. Think about it.
So you're making this money $0.50 at a time.
And what are the factors that affect your revenue.
Well one way to think about it, so the monetization or the RPM,
revenue per thousand queries-- this is the industry lingo, I'm
just telling you what it is. It sounds crazy.
But revenue per thousand queries, so that's revenue
per query times the-- sorry, that's defined as revenue per
queries times a thousand.
And one way to write that is revenue per click times
the click per query.

Revenue per click times the clicks per query and you can
write that as revenue per click times queries with ads as a
fraction of total queries.
And then you've got ads as a fraction of queries with
ads and then you've got clicks per ad.
And if you think about it that's your cost
per click, CPC.
This is your coverage, what fraction of the queries
actually have ads attached.
Because there's some queries that are just not commercial
at all, there would be no possible ad you could show.
And you should think about that as being around half of them.
Half of the queries have ads attached.
And then you look at, of those queries that have ads,
how many ads do you see?
Well that's called depth. So there might be four
ads per page, we query that ads on average.
And then finally there's the click-through rate per ad.
And the click-through rate could in turn be decomposed
even further into the ads specific effect or the
position specific effect.
But basically you think about it.
First term is price, second two terms are quantity.
How many ads you get to show.
And the click-through rate is like the quality measure.
That you want ads with high click-through rates.
So if you want to increase your revenue, well you could
increase your cost per click.
And one way to do that is to create a higher conversion
probability for the advertiser.
So there's an advertiser optimization program
that says hey, bozo use plurals, not singulars.
You know, use action verbs, do this, do that.
You could try to capture more value by
increasing competition.
Remember, this is an auction. You're competing for
the top position. The more advertisers you have
that are competing, higher prices are going to be.
So you capture more value by increasing competition.
You do that by acquiring more advertisers.
And you get increased coverage, so you can get advertisers
to choose more keywords.
That's one of the best pieces of advice you can
give to advertisers. Think the way the consumer does.
Don't just think about the way you see your product, but think
about things that the consumer might be interested in.
If you're advertising, you're selling french fries,
advertise on hamburgers, right?

Get more keywords and then match more broadly.
So you could tell people use broad match, not narrow match.
Now there's a trade-off there because that's going to
effect the click-through rate.
If you have a very broad match well maybe that's
good, but it might lower the probability of a click.
It's kind of like precision recall that I'm sure you've
heard about in this course earlier.
Or you could increase depth.
You get more advertisers.
Then you're showing more ads per page.
You get more clicks and you have more competition,
so that's a good thing.
And you can increase the click-through rate by
showing more relevant ads.
Then that says, we don't want to show so many ads, right?
Show fewer ads, but get them more relevant.
So when you look at that formula I had on the
previous page those are not independent terms.
You have to make a balancing and you're always tuning this
ad system to say well, should we make the promotion threshold
a little higher, little lower or should we disable more ads.
Should we do this, do that and all these things are decisions
that are made in terms of tuning how the system works.

So pretty much what I mentioned earlier, there's a minimum.
You can bid. You know, you have to pay at
least $0.05 to show your ad.
Minimum CTR is around 1%.
It varies a bit, but if you don't get more than 1%
of the impressions have clicks you're disabled.

Another point.
This will be my last point, I think.
Is this is all syndicated.
So I just described how the ad system worked on Google, but in
fact you can sell this service to other information providers.
So if you look at AOL, Google provides both the search
service and monetization or ad service for AOL, so we do both.
Ask Jeeves, they have their own search engines, but Google
provides the monetization.
Yahoo, up until a year ago, Google provided-- or maybe
it was a year-and-half-- all right, time flies.
A year-and-a-half Google provided the search but
Yahoo provided its own monetization using a
company called Overture.
And then Yahoo bought Overture and Overture provided the
monetization and Overture of course, bought Alta Vista and
Inktomi and so now they provide their own search
and monetization.

And MSN was served by Overture and Yahoo up until recently
and now they're also providing their own
search and monetization.

So you do a negotiated revenue share with partners.
So if you go look at search engines in other countries, you
see a lot of bidding and usually Google will bid and
Overture will bid and maybe MSN is going to start
bidding for that. So if you go use-- what's
the-- AUDIENCE: Baidu.
HAL VARIAN: Baidu, yeah. Baidu is an example, but I was
thinking of Wanadoo, the one in France where there's
a big search engine.
I mean, a big internet service provided by the France Telecom
and Google and Overture bid on providing the search and
monetization for that.
Then of course as a contact ads network I mentioned that
we show these content ads, but I should also mention that
there's a revenue share there.
If it's New York Times they get a revenue share of the
clicks on those content ads.
And if you have content, if you had good web content and you
wanted to monetize it you can stick your ad, stick in a
little piece of Java script in your website and Google will
stick ads in there and then they'll send you a
check each month. The check will be a revenue
share based on how many people click on those ads.
And typically it's a fixed revenue share and one nice
thing about it is it provides a business model for
these small websites. So it's a way to make
some money from providing information to people.

I guess that's about it.

Search--a lot of competition.
I think of it as like the magazine business because
you have magazines. That's the content, but
we have search results.
And you show ads next to the content.
We show ads next to the search results.
And there's some people like Newsweek and some
people like Time. Some people like Google,
some people like Yahoo.
And so you're always competing for those eyeballs for the
readers and then your revenue is coming from the ads that
those readers look at.
So it's not so different then magazines or TV or other
kinds of medium business.
In particular it's nonexclusive.
So you look at the advertisers, well everybody who's
advertising and who finds it profitable to advertise using
this method on Yahoo would also find it profitable on Google
and pretty much vice versa.
So it's not really exclusive, just like Ford shows ads
in Newsweek and in Time.

You know, if we look at Amazon they'll show ads
on Google and on Yahoo.
So the challenge is to keep it running.
It's a big, complex system.
We've got to keep the advertising guidelines
with a lot of regulatory stuff are coming up now.
Can you advertise on a keyword that's a trademark
of one of your competitors?
The answer is yes in the U.S., no in France.
There's broad issues: credit card fraud, click fraud,
various kinds of spam, people trying to
manipulate the system.
And of course, there's dealing with the advertisers.
Because I mentioned that was a pretty good deal, but as prices
get bid up and up in up in this competitive auction it would
get less of a good deal, right?
Because it's getting more expensive.
So you've always got this situation of trying to make
sure advertisers get the value out of the system and these
are the kinds of issues that we keep working on.
All right. They said I was supposed
to leave 10 minutes for questions, so I've got
little less than that. Maybe seven minutes.

AUDIENCE: Hi, is this working?
AUDIENCE: So I'm a little confused about the business of
honesty in the bidding because it strikes me that if I'm
out there in a competitive landscape and if I'm going to
pay the second highest bid, well I should just bid a
ridiculously high amount. Because I'm not going
to have to pay that.
And then by the same token, if somebody happens to outbid me
I've driven their costs way up, which is a good thing.
HAL VARIAN: Well, suppose a click is worth $10 to you.
You say, why don't I bid $20?
Well, what if somebody comes along and bids $15?
Then it's only worth $10 to you and you've got to
pay $15 for the click.
So you have to pay more than it's worth to
you if you overbid.
AUDIENCE: Yeah, as a community effect I can see that.
HAL VARIAN: No, but even as an individual.
See even as an individual what happens is people figure this
out that well if what I have to pay doesn't depend on what I
have to bid then I should bid higher.
But then if you get above your value, so then you're paying
more than it's worth to you.
Or at least you have the possibility of that happening.
Now it's a good observation though because a lot of times
people do overbid and they get burned a few times and then
they kind of come back in line.
So it's an effect that is present and we observe not only
in this context, but we observe it in laboratory experiments in
economics and it is something where that does happen.
But it's usually corrected.
By the way, I will say that I do want to emphasize that the
ad auction as I described it isn't quite the same as the
straight Vickrey auction, but it's a generalization of it.

AUDIENCE: Do you think the-- can you hear me?

HAL VARIAN: You're on.
AUDIENCE: Do you think the contextual ads are going to go
the way of the banner ads in that people are going to get
more and more accustomed to seeing them and just start
ignoring them after awhile?
HAL VARIAN: So what this is called, the term of art in
the industry is ad blindness.
People develop ad blindness.
And it's like, you know, they say in the Super Bowl they
monitor the water pressure in New York and they can always
tell when the commercials on because all the toilet flush.
You know, it's that kind of thing.
So what happens on Google is actually we found that
if anything click-through rates are going up.
I mean, because the more relevant you can make the ad
and if you really put this emphasis on ad quality.
Of making sure that the ads are highly relevant.
Then in fact, the more clicks you get in the long run.
And on very, very highly commercial queries it's
not uncommon for the ads to be more relevant than
the search results. We're not necessarily happy
about that, but go type in a few really relevant, highly
commercial queries and judge for yourself.
So if the ads are good quality people like them.
And put an emphasis on quality and in general you're going to
get an improvement in click-throughs.

AUDIENCE: Two questions.
I'm wondering if firms-- [INAUDIBLE].
HAL VARIAN: Oh there you are,
OK AUDIENCE: I'm wondering why don't firms just work on
improving the quality, the characteristics of what Google
uses to rank them from a particular query as opposed to
paying for advertisement? And also, what's a firm's
incentive to pay for an advertisement if they
are already at the top of a general query?
HAL VARIAN: Well I will tell you that in some cases there
are companies that have very good search results that
come up at the top of the search results.
Typically, that would influence their spend on ad.
So that if you're looking up categories where they come
up in the top of the search results they won't pay that
much to be in first place on the ads.
You know, they figure that out.
So they're exactly the way you described.
And then the question of trying to manipulate the search
results in order to get higher placement, yeah of course,
people try that, too.
But that's not so easy to do because Google is watching and
Google has an interest just like in the magazine wants
to have honest and accurate content.
We want to have honest and accurate search results.
So we're trying to rank them by relevance and a lot of times
ads are things that are not directly the kind of thing
that's in the search results, but rather something that's
complimentary to it.
So again, travel ads are easy.
You type in you know, backpacking in Hawaii and
they'll show you a bunch of cheap flights and trips
and things like that. So the search results are
here, the ads are here.
The ads are relevant to, but not identical with
the search results.
And you think about it that's the way most advertising works.
AUDIENCE: We've just begun advertising on Google and
everyone in my office has been poking around to see what
number we come up and because it's our own ad we
don't click on it. So now I'm concerned.
You talked about best ads, should we all be
clicking on our ads?
HAL VARIAN: No because if you click on your ads then
we'll find out about it.
Obviously that's monitored.
AUDIENCE: Are we damaging our reputation by not
clicking on the ads?
HAL VARIAN: No, in your particular case-- obviously
there's a bit of an incentive issue here.
You don't want to click on your own ads because
there's a lot of detection.
We have a lot of systems in place to detect click fraud and
part of your agreement said you wouldn't click on your own ads.
Yeah. There's some things where obviously people want to
click on them and make sure everything's working right
and that kind of stuff.
But if you look at it there's some issues of that sort.
And indeed, there obviously are people who do click on their
own ads and they're dealt with.
We tell Guido and Guido takes it from there.

AUDIENCE: I see you talked about ads placement strategy
between Yahoo and Google.
So far, Yahoo case may be user is not happy.
Sometimes the number one position are not
something they want. For the Google case, maybe
the advertisers are not happy.
They can bid more and want to put in the number one position.
So based on your judgment, which one would you see
would be the future standard for the ads placement?
HAL VARIAN: Well I will tell you that once you put one of
these systems in place it's not so easy to change it.
So it's like, I don't think either of them will be
changing to the others model any time soon.

You know, it's like any other differentiator.
Google likes its method and Yahoo likes its method.
I think that both of them recognize that you do want
a relevance ranking for ads.

That is similar to, but not identical with the relevance
ranking for search.
So you both should be concerned about the ads quality to make
sure they're really delivery to the users what they want.
And there are different ways to go about that.

You know, even there you think there's a lot
of fine distinctions. We can say there's
the creative quality.
There's the brand. There's a landing page quality.
You may have a great ad, but it's a lousy web page so then
you're not going to get too many conversions or clicks.
So from the viewpoint of looking at this issue of ads
quality there's a lot of dimensions of that when you
get down and study it deeply.
MARTI HEARST: Last question.

AUDIENCE: Can you speak a little bit about
ads in [INAUDIBLE].

HAL VARIAN: Exactly the same thing as AdSense.
We filter out some things because they were deemed to be
inappropriate in e-mails and somewhat different mediums.
So there's a little extra filtering there, but other than
that, it's using the same ads the same, or not exactly the
same, but similar ad engine-- the nice thing about it is
from the viewpoint of the advertiser, you come up with
one creative and it's going to show up on the search side, on
the contents side, on the Gmail side, on local, on
all these different things.
So it's like we're using information.
You get the same ad can appear in many, many different medium.
So pretty soon we'll have t-shirts that flash on
and off with Google ads. No, don't tell
anybody I said that.
HAL VARIAN: Thank you very much.