Automated Bidding on AdWords

Published on September 22nd, 2008 by Jeremy Chatfield

We have done a lot of our work on AdWords using the API and programs that we’ve written. We’ve done some work with fast paced advertising – advertisers with a large inventory of rapidly changing stock with different prices (think “travel”). I was reminded of this when I saw a comment on this blog by Dr Gerda Arts of Sayu.co.uk and followed it to the website. I found a brief white paper over there, about automated bidding. It’s a pretty good read. But it’s wrong.

Well, when I say “wrong”, I suppose I really mean “less than complete”. I think the reason is quite subtle – and not explicitly recognised in the white paper. That doesn’t mean that they don’t know, and it doesn’t mean that they haven’t worked it out… it just got me thinking about some work we’ve done. That, in turn, means a brief look at recent history.

Google’s dominance of paid search, especially in the UK, means that many users won’t be aware of the original innovator in the arena. “First mover advantage” (yeah, that’s irony for you) went to GoTo.com, who eventually became “Overture” and then “Yahoo!Search Marketing” (Y!SM).

Bidding Models

GoTo’s original model for an advert auction (pre-Panama) was that bids would be compared and the longest standing, highest bid, would appear first. A simple auction, easily understood. Also, an auction in which automated bidding would be hugely important. Why so? Because the majority of clicks go the highest ranked listing, but the value of the listing would depend on conversion (or affiliate reward). Getting your advert to the right position would be a matter of fast paced juggling.

Automation can help reduce costs and mistakes in an auction like that, for all advertisers. Tactics, such as choosing when to bid, and how frequently, become important. And little issues like bidding gaps become hugely significant – because you pay 0.01 more than the advert below you. So if you can find a spot in which the two adverts above you are paying just 0.01 more, but you have a 0.05 spread to the advert below you, you pay much less than your higher competitors, which may justify the lower click volume on a lower position. Lots of calculations to be performed quickly. Automation, absolutely useful.

GoTo’s model for search revenue was also quite different from that for Google. Initially, GoTo had to partner with higher volume third party sites. This lead to the acquisition by Yahoo. Google’s first advertising was on their own site, and they could then sell the higher value adverts to search partners who were already prepared, by GoTo, to handle externally supplied advert inventory. This results in an apparent difference in results between Google and Yahoo. Where Google impressions may soar markedly with position, GoTo’s already broad spread across third party properties meant that the impression volume was not so accelerated by position. That was also subtly influenced by Google’s auction mechanism – covered elsewhere on this blog – which offers absolutely the highest returned value for every advert space used.

As an early advertising system, GoTo offered a single advert. It costs the publisher cold hard cash to vet adverts. So running a single advert for each advertiser for each keyword, reduces costs. Important for a startup. However, ask anyone in direct marketing what their most powerful technique is; it is the parallel A/B test. The ability to test copy for effectiveness can transform a failing campaign into a successful campaign, sometimes with the most subtle of changes. Google’s AdWords allows alternate adverts – dozens of them for massively parallel A/B/C/…. testing. This changes the way that the bidding system works – because you need to separate the behaviour for each creative that is used.

Budgets and Bids

The way in which budgets are set and satisfied also results in some intriguing differences between the bidding systems. In GoTo’s model, it appeared that adverts participated in all auctions – even if they did not have the budget to do so. Google’s model appears to run the auction only between participants with enough budget to be able to offer an advert (there’s evidence for this that I may explain in another article – it’s moderately complex reasoning and somewhat off-topic for this thread).

Something really important to carry away is that CPM and CPC (cost per thousand impressions and cost per click) models are interchangeable – if you can make some assumptions, or some measurements. For example, if you can assume a 2.5% CTR. If you do, then for every thousand impressions you’ll see 25 clicks. If you know the CPM, you can then infer the CPC. Google makes assumptions like this, all over the place – you can even infer what the assumed values are with some experiments – I may write those up. It was pretty cute working out what Google’s hidden numbers are :)

Why CPC?

Psychologically, advertisers are happier with CPC presentations of cost. For example, small advertisers may not want to buy 1,000 impressions. When presenting to board level decision makers, you can tell them that you are buying clicks, not impressions which may or may not turn into business. Clicks might turn into business and definitely turn up on the web site – and if they don’t, you can challenge the publishing system for fraud. But impressions – a lot softer… harder to explain why you don’t get clicks or conversions. So Google’s system has gone for CPC presentation – but I’m pretty certain that the auction itself runs as a CPM model.

Broad Match – Broader Than You Might Expect

Finally, Google emphasises broad match (it is the default, it is touted in training, and gives Google control over advertisement appearances for arbitrary searches – which Google can use to its’ advantage, smoothes the landscape for bidding, etc). The broad match auction allows Google to spread the value over whatever it deems relevant. Bid on your company name in broad match, and if you’ve bid enough, your advert will appear on searches for competitor names. No one who has done any courses in Marketing Communications should be happy at this point – MarCom tells you that you need to focus an advert on a competitors specific weakness, relative to your business, if you want to capture someone searching for a competitor.

These differences mean that the nature of the auction on Google is less about fast paced changes – for most advertisers, who do not have stock that ages, and in turn leads to significant reasons for most advertisers to avoid automation or to reduce emphasis on manipulating bids as a technique. The focus for success on Google shifts from bidding strategy and into ad copy, keywords, match types and landing pages. Arguably, this is the right emphasis anyway. The nature of GoTo’s auction system drove attention to bidding, and their inability to offer parallel A/B testing of adverts, demoted interest in improving CTR.

Aggregated reports also make a huge difference, too. GoTo’s model was pretty simple. You could infer, for a given time period, what the position was, and the paid price, and the impressions and the clicks. For Google, we have a variable delay time (between around 15 minutes and 3 hours, with a minimum of 45 minutes for reliable reporting), and aggregation – you can’t get to results for the last time period, for periods of less than a day. This means that to infer what has happened late in the day, you need to know what happened earlier in the day – because you’ll subtract all of those from the aggregated result. However, this leads to a blurring of values late in the day. By 10pm, you’ve had 95% or more of the clicks that you’ll get, and so changes made to bidding have a relatively tiny effect on the numbers that are reported – things like the averaged position over the day are more heavily influenced by the earlier days’ higher activity. This means that the effective measurement times to get insightful data are reduced – you can’t make sensible decisions about changes when 90% of your results have been collected.

Seasonality

What really cripples automation for many small advertisers, and can cause some interesting problems for large advertisers, is seasonality. To effectively select the right price every Christmas, you need to have a model of the change in impressions for the previous Christmas. Building up that model is expensive. Retailers carry this in their heads, but it needs to be expressed to the model before it reaches the date… otherwise you find that the model is no longer predictive, but reactive, and always behind the curve.

Take, for example, the classic New Year resolution to stop smoking. Keywords relating to stopping smoking have a low volume of search over the year. Starting just before the New Year, it ramps, and declines by the end of January, reaching a peak that is vastly higher than the normal monthly run – just for this period. If you use a predictive model, it’ll keep projecting that tomorrow will be like last month… it’s had 11 months of being right. It will typically project that tomorrow, the volume will be down. And of course, it rises again and again through January. So the algorithm spends all its’ time chasing the rising curve. How it reacts to the declining interest is also problematic – with no history and a short time planning horizon. Most algorithms will be dominated by “tomorrow will be like today, with minor changes” – so a declining curve has them retaining bids for too long as the market declines.

Because the auction is really for CPM, and broad match means that bids are compared between wildly different searches, the whole effect is to smooth the bidding landscape and slow the variation of bidding changes. Bidding on Google is pretty stable, after the first frantic period of working out th messages and building a CTR history.

Noise

Complicating all of this, is that the Internet is a “noisy” signalling environment. Bursts of searches may reflect real changes of user interest, or just statistical variation that means nothing. Designing algorithms that take care of high noise environments. with unknown seasonality, is pretty tricky. Getting retailers to describe their expected annual search impressions is also sometimes and exercise in futility – because over- and under-statements cause problems that can be worse than having a near-Bayesian model.

None of that is the defect

The missing point is that each keyword has several built-in characteristics. Different searches have a different implication, that predisposes the search user to specific types of advert. For example, a search for a brand name, especially a business’ own name, usually indicates a high intent to hear from that business, not a competitor. The CTR for a competitor will be tiny, and the CTR for the business will be high. This leads to an inverted paid price. That is, the bidding landscape will tend to have the lowest average Cost Per Click at Position 1. At other positions, the AvCPC tends to be higher, as the advert is under competitive pressure.

Even worse, if the product being sold is a niche, the Average CPC becomes somewhat less significant – because it is important that the higher adverts weed out the segments that you can’t afford to handle. The optimum position is then dictated by competition messages – inferred by increasing conversion rates in lower positions.

Getting the bid right can’t rely on a simple Newton-Raphson optimisation, but has to look for non-local minima. Programming systems to find non-local minima is tougher than most businesses can handle. It is possible that Sayu.co.uk has handled this – but a number of the largest automated bidding systems, especially those popular with large agencies, seem to lack any serious bidding model.

When the non-local optimimum happens to coincide with the traffic maximum, and automated bidding systems fail to identify that, it is a pretty poor failing.

Anyway – thanks to Gerda Arts and Sayu.co.uk – it’s good to have someone explaining what they do in more detail than the handwaving at bigger competitors.