Or: Collateral Damage by the Cult of Celebrity.
Or: The Hidden Costs of Downloading Porn.
(No, I've not seen the film filked in the
post title.)
I got a
work
call from the advertising manager of
one of my customers.
They are just starting out with
pay-per-click adverts.
I understand that it's been working OK
for the last few months,
increasing their number of sales and
general interest by enough to cover the cost.
In general, each time they refined their
campaigns or spent a little more,
they saw a positive result.
I got a call because
since the start of February,
they were suddenly spending a lot more money,
without any corresponding results.
Was there foul play?
Had a competitor noticed their adverts and
kept clicking them from different browsers?
In short, was someone running a click-fraud
attack on them?
This is the sort of question I enjoy.
(I did a statistics degree, but I don't
use it directly in most of my work.)
I massaged the logs and computed some
general-purpose statistics.
First, I looked for obvious attacks:
were any particular machines or domains
responsible for most of the clicks?
Most non-IT people use the same ISP at home
and at work, in my experience, so if it was
a particular competitor, I'd expect to see
lots of clicks from one domain.
Nothing really stood out.
The next thing I looked at was the geographic
locations, very roughly.
The pay-per-click adverts are targetted at
European users on European sites.
If it was some sort of distributed
denial-of-service attack, I'd expect to see
some increase in American and Asian traffic.
Where were visitors coming from?
I looked up locations
by the first part of the IP address,
which is crude, but good enough for this.
You could even do something similar with the
the xkcd ipv4 map
- it looked like most of the addresses were
in Europe, which is what I'd expect.
There was one interesting US grouping,
but that turned out to be one of the
ad networks checking their ad links were
still valid.
So, I compared the time distributions of
clicks now with clicks a month ago (when the
ads were working fine).
There were increased numbers of clicks just
before 9, just after 17 and midnight-4
local time.
I could dismiss
the clicks just outside 9-17 common work hours
as people looking to buy for their hobbies,
but midnight-4 is a bit strange.
Looking in more detail, the midnight-4 clicks
were coming mostly from ad placements on
one site.
Looking at that site, it seems they specialise
in 'sexy videos' of celebrities.
Needless to say, that was not my customer's
intended target market!
I used that information to look at the whole
dataset and see
how much of the increase was due to that one
site.
Almost all of it was from there.
Scores per night.
Did any of those visitors stay long? Nope.
It even looked like some of them were just
private web spiders, downloading a page and
all links from it, which I guess people might
use to bulk-download videos overnight.
So those results were sent to the advertising
manager and they've updated their campaigns
to exclude that site and any others related
to sexy videos.
It reminds me of the John Wanamaker quote
"Half of my advertising is wasted, I just don't know which half."
Maybe by working out statistics from the logs,
we can reduce it from half for web ads -
or at least limit it to half?