AdWords, Geotargeting Myths

Note: A short updated article about Google AdWords Geotargeting in 2008 is now available.

Changing the geotarget can change your paid search performance quite markedly. However, commentators, and even Google, publish some wildly inaccurate information. I’ve read a few strange myths about geotargeting, in various forums. I’ve got some clients for whom geotargeting at country, and smaller areas, is very important, so I’ve done a bunch of experiments to find out what really happens. Here’s a brief account of some of the results.

What Is Geotargeting?

Geotargeting is one of the ways to address adverts to people who are more likely to buy. Other mechanisms within AdWords are the advertising system (keyword targeting, content match, site targeting), publishing system (web, mobile/cellular, print), languages, message type (text, graphic, video) and network (Google only, Search Partners, AdSense sites, etc).

Geotargeting relies on clues from users, most notably the IP address of the users’ web browser, to identify where the user is. Potentially useful additional clues are keywords or content that implies a geographically relevant targeting for the site. For example, “elephants denver” may imply that there is a geographical interest in Denver – though “John Denver” doesn’t imply geographical interest to the same extent. Intent is all important, and notoriously difficult to manage in search engines and in paid search.

Setting Geotargets

Geotargets are set at the campaign level of AdWords. Other advertising systems have different ways to select the geotarget. For example, Yahoo!Search Marketing selects the country geotarget at the account level, so a multi-country target audience needs multiple accounts (which can be tied together under one login). Some paid search systems do not have any way to select the geotarget at all, even at country level, or if they do offer a geotarget, may not enforce it effectively. We’ve measured some lower tier search engines that offer country specific targeting, with 55% of clicks out of the target area. Using third party (and cheap) geolocation services, suggests that Google is around 98% accurate at country level targeting – that’s within the measurement error I expect from the geolocation DB we’re using, so Google could be as much as 100% accurate or as low as 95% (1 in 20 adverts served to the wrong country – enough to run several large click fraud businesses; back of the envelope calculations for accuracy, with some hand waving and unmeasured assumptions, which would take an article longer than this to explore… another time, perhaps).

In AdWords, use the Campaign level “Edit Settings” link, and look at the lower part of the right hand column. You should see a language setting (surprisingly useful for some clients) and the geotargeting settings. This usually defaults to the country in which the business is registered. So a UK business address defaults to a UK (England, Wales, Scotland, Northern Ireland) target.

Measuring Geotarget Performance

There are two main reasons to look at geotargeting performance. Clicks from outside your geotarget should result in wasted spend – if clicks outside your geotarget result in sales, you should consider a larger geotarget. You may also want to specifically send different adverts to sub-regions in your main service area. For example, a national advertiser with local stores may want different adverts when the web user is within the local service area of a store – in this case, national adverts may focus on “buy online” and local adverts may carry specific local offers and “buy online or come to the store” messages.

Some web analytics packages report the country or smaller regions, if they have a relationship with a geolocation service. Google’s web analytics package, Google Analytics (Urchin, as was), now uses the same geolocation service as AdWords. Because the geolocation service is the same, using Google Analytics to help tune regional or country specific targeting is useful; when the web analytics package reports the same location data as is used for targeting, you can retarget and tune. You need to tune with respect to the geolocation service used by the paid search or advertising vehicle… because if their service area definitions differ from that of a third party analysis system, you’ll mistune the delivery.

However, there is a weakness to using and believing a geolocation service shared between the analytics and delivery systems. If the same location service is used for both targeting and measurement, then misdelivery can not be identified. How so?

Assume that you target only the USA. You set the campaign geotarget to the USA. You measure using Google Analytics to see if any adverts are clicked on from outside the USA. When the geolocation services used by delivery and measurement are the same, then errors in location will not be shown – because the assumed location in both cases will be the USA. So you’d be unable to see errors in geotargeting delivery, when the analytics system shares location data.

When checking the delivery effectiveness, you really need to use a different geolocation service. That is, if the AdWords geolocation service shows an address in Mexico as being in the USA, this could not be identified by Google Analytics… but a third party geolocation service could show the misdelivery.

Geotargeting Myths

There’s some strange stuff that people believe about geotargeting. Some of it is promoted by Google – probably not intentionally, but caused by some unstated assumptions that may not always be true.

IP Address Gives Away Too Much Personal Information

My domestic IP address is really sufficient to localise me to the UK only. Most geolocation services show that I’m 150 miles north, or about 60 miles south, of my real position. This is because internet connections don’t map to the physical world. Only in special cases can you use an IP address to accurately give location. There’s some newer techniques, none of which I’ve yet seen in wide usage, that can do things like measuring the response time, to narrow down the location of an IP address. But even those aren’t normally more effective than a few tens of miles (in the UK – could be closer in the USA, I think).

The way that most services seem to derive location is to use ISP’s optional information about the position of their routers, to give the end point location. In places like the UK, with a national telecoms structure, the router may be hundreds of miles from the end user. In places like the US, there is usually a good reason to believe that the end point is probably only tens of miles from the router, or not in Virginia. Why Virginia? Because that’s where AOL and Uunet and some other national ISPs located central services… and so many end users will be shown as coming from their web proxy/cache services, as the ISP tries to keep down their upstream bandwidth usage (that’s not a lot to do with Internet Marketing, but knowing this stuff helps to interpret why European advertising shows US clicks – much of it is US based web/cache proxies for European users).

Complicating all of this are “problematic” services. Some ISP’s use “dynamic IP Addresses”. This is very common with dialup services and also affects some ISPs offering ADSL. A dynamic IP address means that it could be used by people at opposite ends of the country in a relatively short period. Mobile data networks may give different IP addresses in a single session, as a result of many causes, from switching between bands, or changing base stations. There are also deliberate services such as anonymising proxy servers, that disguise the end point IP address, and may shift the address during the session, possibly around the world. Many ISPs use web cache/proxies, resulting in requests coming from the proxy server, not the user. Similarly, many large businesses will have a national gateway or even an international gateway between internal users and the internet – for example, I know of a business in the UK, with dozens of offices around the UK, with a gateway in Peterborough and I believe that some of the larger computer businesses with international networks have gateways in the San Jose/Mountain View area for all international offices.

While it is possible to get geolocation data, potentially very accurately (down to 10 metre/30 foot range, with the right technologies), for mobile/cellular phones, the cost has been high. The last project I was on that looked at using mobile phone position had a cost of around $0.20 per query (2002 price – cost is likely to have changed, probably downwards; see the comment by Josh, below). If your average advert cost is in that range, it would make geotargeted mobile advertising rather too expensive, though the location cost could be shared by all advertisers targeting that user/location.

The use of obfuscating technologies, such as dynamic IP addresses or anonymising proxies, whether intentional or not, decreases confidence that all IP addresses can be mapped to any location much more specific than the solar system… IIRC there have been some satellites and other space vehicles with IP addresses. 🙂

Static IP addresses, stable for long periods, offer the opportunity for the location to be more correctly determined, over time. But is there any sign that services from Google use this?

Google collects a lot of information from me, when I do local searches, or list my business, or register to do other physically addressed activities that would let them infer the location of this IP address. Despite having had the same IP address for years, and having supplied plenty of location information, Google Analytics clearly shows my IP address as being in the UK, and nothing more precise. This suggests that Google is not yet using all the sources of geolocation information that it could. Sample of one, though – not a lot of confidence in this answer!

Local Adverts Are Prioritised

Google staff often assert that local targeted adverts are shown in preference to adverts for a larger area.

This is an effect, not a policy.

If you run an advert for a local service, in a local area, you should expect that the CTR will be higher than a national advert, with a non-local message (or even with a local message). The higher CTR for a local advert will result in a higher QS for the advert, so the locally targeted advert will run in preference to the national advert.

However, there is nothing within Google that specifically prioritises local keyword search adverts. If you bid higher for national adverts than local adverts, you can get to a point where national adverts will be shown.

When testing messages, or if you have different offers that should be made locally and nationally, then you need to be aware not only of the CTR, but the bid, to ensure that national adverts are not sent locally.

You might think that you could use campaign level targeting to offer adverts targeted to a region, and another campaign that targets adverts elsewhere. This is an illusion. Because some IP addresses can not be given a location within the country, you need three campaigns – one for the target, one for other identified areas excluding the target and a lower QS (managed by CTR and bid) national advert that captures everyone without a specific location.

With an account cap on campaigns, you’ll rapidly run out of campaign settable targets with any large business of multiple locations. You’ll need multiple linked accounts to service more than a few tens of campaigns and make them sensibly sort into the types of target that they address.

85% Accurate, or 94%, or… What?

Google offers a variety of numbers for the accuracy of geotargeting. This appears to be because there are a lot of numbers needed to specify geotargeting – more than a single number. Representing the performance as a single number means that there are a lot of ways to choose that number and interpretations as to the effect. Worse, I believe that the numbers vary according to the country, because of the way that IP addresses and ISP’s work in different countries.

Given a geographically targeted advert, the following cases can happen – this is basically just a Venn diagram of the cases where a location is known:

* The advert is shown to an IP address identified as being inside the target; the user is physically located in the target; this is good
* The advert is not shown to an IP address that is physically located in the target area, because the IP address is not located there; this is bad
* The advert is shown to a user outside the geotarget, because the IP address is incorrectly assigned to the target; this is bad
* The advert is not shown to a user outside the geotarget; this is good
* The advert is shown, although the IP is outside the geotarget, because the keyword includes location specific names; arguably good.
* The advert is not shown, despite using location targeted names; this is arguably bad.

And there’s odd cases… like what should happen when you only know the location to within, say, a country:

* Advert is not shown, because the location isn’t known to better than a country; arguable
* Advert is shown, despite the location being imprecise; arguably wrong

To really understand geotargeting, Google should be publishing figures that represent the likelihood that an advert will be seen by a target audience, the likelihood that the advert will be seen by users outside the physical target, and the likelihood of missing audience that is in the target area. Why all those numbers? Because when those numbers dip below a certain value, you are better off not using geolocation services. I’ll probably write another article about this… it took me a long time to work out what the critical values are, and why, and it will bloat this article and distract from the key points, if included here!

Self Tuning

I’ve seen several commentators explicitly state that Google will tailor advert delivery to geographical regions. That is, if you use a nationally targeted campaign, but offer adverts for Denver and Chicago, that the system will select Chicago adverts for Chicago addresses and Denver adverts for Denver addresses.

I can’t find any evidence that this works.

One experiment that demonstrates this is to send adverts in multiple languages to areas with two significant populations speaking different languages. This should result in advert serving in the ratio of language use, and with analytics showing that the CTR is at the language-appropriate level, and the clicks come from the major population centres for the language. What actually happens is that the dominant language (highest CTR/highest revenue) advert is selected and served to the whole region. This appears to be true whether the region is within a country or crosses national borders.

You’d expect that if this effect worked, it would be most likely to work, say, when targeting adverts to Switzerland, with mixed populations of German, Italian and French speaking natives and quite a few migrants with other languages. Or when using French and Dutch, and targeting Belgium. Or using German and French and targeting Germany and France… Any way you slice and dice it, there is no way that Google is selecting a language specific advert for a geographically well defined area. I infer that no other location specific offers will be differentiating the delivery, either.

Use of locality keywords

In tests that I’ve done, when I use locality specific keywords, I don’t see locally targeted adverts – or rather I do, but they are targeted for where I am, not for where I’m interested in. This has an effect on, say, realtors/estate agents. If you are moving, you want a local realtor where you are moving to… that means you need to use national advertising with locality keywords.

The implications of Google’s documentation are that geotargeting solves a multiplicity of audience problems. I think it works with one model of geotargeting. Extending the use to infer that it solves other problems is overstating the capability.

Summary

Google should be providing better information, at country level and smaller regions, for the accuracy of geotargeting. This needs to express the likelihood of reaching a user in a target area, and the likelihood of reaching users outside the target area, as a minimum. Ideally, there’d also be explicit information about the likelihood of reaching a user in an area only by using national targeting (that is, the proportion of users in an area who have only national location or a different region for their geolocation).

Google should be considering geotargeting of mobile advertising. This is likely to be a growth area and mobile users will need local information – both local to where they are now, and local to where they will be, soon. Important for iPhone users, eh?

Advertisers should now be considering at least six campaigns for a local businesses that may be used by non-local searchers:

* the locality – keyword
* national – keyword
* the locality – keyword with mobile adverts only (though this is probably pretty pointless at present)
* national – keyword with mobile adverts only
* the locality – content
* national – content

Google should be offering some kind of measure of anonymising and proxy services that are being used. These disguise geolocation and can cause mis-delivery (non-US users being shown US advertising, when US proxies are used, for example). When obfuscated locations rise to a critical value, geolocation services cause more problems than they solve.

Well, I’m pressed for time… so I’ll leave this article for the moment… but there’s a lot more to look at with geolocation, click fraud, ROAS and conversion improvement techniques. For another article, eventually…

Updates

2007-10-26 Edits for clarity; corrected typos and some poor word choices. Also a pointer to a really pertinent comment by Josh, below.

2007-10-29 Added a little more detail on the complications of proxy servers and company gateways, triggered by CPCcurmudgeon’s comment, below. I believe that he’s a network systems programmer with experience working for a search engine – a credible commentator.

2008-02-26 Newer article published on AdWords geotargeting differences in 2008 to reflect more recent experiments and changes in Google’s behaviour.