Advanced PPC Analytics – SMX Advanced
- Have you had enough analytics yet? Of course you haven’t. There’s never enough analytics! Here’s some just for the PPC folks.
Moderator: Matt Van Wagner, President, Find Me Faster
Q&A Moderator: Joseph Kerschbaum, Client Services Director, Clix Marketing
Alex Cohen, Senior Marketing Manager, ClickEquations @digitalalex
Adam Goldberg, Chief Innovation Officer, ClearSaleing @AdamSGoldberg
Frank Kochenash, VP Client Services, Mercent
Wister Walcott, co-founder and EVP of products, Marin Software @t_wister
I’ll be honest and say that I’m probably going to understand this one even less than I usually would. I like feeling challenged by sessions but PPC is another language entirely and on top of that I’m still sick. So if this reads like gibberish, I’m very sorry.
Matt says that we’re not going to dive into the nitty gritty today. Which I appreciate.
Wister is up first. He’s very very tall.
Paid search forecasting – how they do it
- predict clicks, conversions & profit for CPA/margin targets
- bid to an overall spendting target (beug based bidding)
- find optimal CPA to balance conversion volume cost
- maximize profit
- target specific positions
- maximize revenue for a CPA target
How do you approach this?
I’m going to just leave that slide here because I don’t know what it all means.
How many keywords per group
Test for CTR and QS differences by number of keywords per group. They discovered that high volume keywords should probably get their own group.
For a CTR of 5%, no group had a QS of less than 4. Conversely for .5% CTR, no keyword had a QS above 3.
1. Above-average keyword count
2. Below average CTR/QS
3. Start with high click-volume groups first.
other useful analyses
Best practices: Any broad match terms getting more traffic than their exact and phrase siblings? Any groups with just one active ad? Are you identifying loser ads on CTR or conversions? Any budget-constrained campaigns? Adjusting bids for Day of Week and Time of Day? (Requires the time of click) Are you deleting keywords with no impressions for 13 months?
Next up is Alex Cohen.
His friend Adrienne lost a bunch of weight. How? Because the system took all the information and turned it into points that made the excess of data much clearer.
An excess of data leads to a shortage of optimization.
Slim down your CPCs with compound metrics
Look at impression, clicks, CTW, Avg CPC, conversions, conversion rate, CPA, Revenue and Net Profit.
Normalize your date by looking at Profit Per Impression (Profit/Impressions) = PPI
clck.it/ppi-links leads to great articles on this topic.
clck.it/halvarianbidding – great video on incremental cost per click.
There are 2 Fundamental drivers of prioritization – fear of loss or hope of gain.
Example: Search Query Mining – If a query is a question, the ad is the answer. PPCers job is to align those questions and answers. Not every alignment has the same value.
QueryMiner Ad Relevance Tool – useful for figuring out where you should be focusing your time.
Impression Share Run Rate: Throw money at your IS problem.
Impression share is a campaign only metric
clck.it/43tools – PPC tools galore
Next up is a battle! Adam and Frank are going to fight about attribution! Adam goes first.
There are three types of actors – introducers, influencers and closers. There can be many influencers but only one introducer and closer per conversion cycle.
All you need to do is be more accurate than you were yesterday.
The least accurate is “last click wins”. Next is the “even attribution” – everyone gets the same credit. Most vendors start with the even + exclusion attribution. (There are three more “path position”, “attribution pattern” and “algorithmic modeling” but they’re not out of the box and he’s not covering them in this presentation. Algorithmic involves Ph.D.s)
Even with Exclusions
- Ability to set maximum days in path
- Ability to aggregate ad network impressions
- Ability to set hours between views
- Ability to apply “exclusions”
- by ad type
- by ad position
- by time factors
They tell people that if a conversion comes from a branded search, exclude that and apply to previous steps because that’s just a re-finding search.
You may want to time out impressions — does someone really remember an ad from 30 days ago?
PPC often gets slighted by the last click model.
Use exclusions to zero in on the terms that are actually bringing you traffic so you can ignore the navigational searches and bring traffic to the terms that actually drive leads.
Frank is up last.
He’s going with a nuanced metaphor about snake oil. Okay then. Snake oil originally was Chinese medicine and it’s still used. The problem is not the oil, it’s the misinformation that surrounds it.
Today’s takeaway: Measure the algorithm.
Advanced ROI attribution is about predicting probability of conversion. How is it used? Providing weights used in calculation of ROI.
Viewing the customer journey as an evolution of conversion
In a perfect world it looks like this:
- assume the influencers are static
- assume every exposure or user interaction deserves credit
- assume every cookie/consumer is the same
- don’t adequately handle external influences
- assume creative and placement don’t matter
- assume…something else. And more as well.
You need to measure the difference between predicted and measured to find the model error (and the model error of time. You want your model error to get less over time and you want it to be adaptive.)
Buying ROI attritibution algos
- Request model error rates prior to launch and over time
- Ask how does the algo adapt to disturbances or changes
- Ask how it adapts.
How do you use product-keyword algorithms for a particular targeted case?
Take ads out of circulation when a product is out of stock, put them back in when it comes back into stock. Use it to drive ads by SKU. Look at how correlated the SKU is to the keyword. (There may be variations like sizes and colors.)
[There’s a chunk missing here while I swapped my battery. My apologies.]
This is not an 80/20 situation: most of the value requires most of the work.
When buying product-keyword attributions algorithms:
- Request correlation statistics (i.e., how do you know it’s working?)
- Ask how do you establish correlations between things I can control (like keywords and LP) and things I can measure (like product inventory)?
- Ask what data is being collected to enable these calculations?
Above all, measure the algorithm.
The Q&A is most very specific so I’m skipping it.