Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Lessons learned from applying semantics to blogs


Published on

Lessons learned from using semantics on a blog network; pertains to targeted advertising, engagement levels, and predicting social media amplification.

  • Be the first to comment

  • Be the first to like this

Lessons learned from applying semantics to blogs

  1. 1. Semantics & BlogsLotico San Francisco Semantic Web Meetup Hosted by Federated Media Publishing August 8th 2012 Tim Musgrove Chief Scientist, Federated Media Publishing ©2012 Federated Media Publishing, Inc. All Rights Reserved.
  2. 2. Who is FM?Founded in 2005 to help hundreds of high-quality independentpublishers band together to earn sponsorship and ad revenueNow a top-ten US audience (comScore) and growing fastHistorically: have paid out well over $100 million to our authorsand publishers John Battelle
  3. 3. A couple of lessons I’ve learned about semantics & blogs1. With display ads on blogs, you need semantics even more so than with a general audience2. Higher level semantic features, such as reading level, can predict how much and where a blog post will be shared in social media
  4. 4. Part OneDisplay Ads On Blogs
  5. 5. An open secret • Supply-side advertising players (including content networks like FM) often don’t like targeting to be too focused • Why not? • Because the math isn’t nice to them Let’s see that…
  6. 6. Ouch…….To sum up:Broad targeting $1,400,000increases bothCPMs and soldinventory, but…. $1,200,000Narrow targeting $1,000,000brings a sharpdecline in sold $800,000inventory $600,000So, even thoughit has a higherCPM, narrow $400,000targeting isn’tdone much. $200,000 $0 Run-of-network Broad targeting Narrow targeting
  7. 7. Is there another way to implement this? • Q: What’s the one case when narrow targeting can have the effect of expanding the inventory purchased by the advertiser, instead of shrinking it? • A: When it makes the advertiser comfortable going outside their usual content category Let’s see that….
  8. 8. Typical buy pattern (without conversation targeting) Suppose: Advertiser comes to FM w/a “cloud computing” campaign. This is what their buy usually looks like: Tech Finance Sports $ $ $ $ $ $$ $ $ $ $ $ $ Government & Lifestyle Small Business Law
  9. 9. Modified buy pattern (with conversation targeting) But when the advertiser learns about Conversation Targeting, this is what their buy starts to look like: Tech Finance Sports $ $ $ $ $ $$ $ $ $ $ $ $ $ $ $ $ $ $ $ Government & Lifestyle Small Business Law $ $ $ $ $ $
  10. 10. Another implementation technique : Target conversations, not topics• What’s the difference? – Topic = a particular theme of subject matter – Conversation = a cluster of related topics comprised of various themes• Results: simplicity for advertiser, and way more inventory! • Let’s see an example….
  11. 11. The Small Business Conversation (just a subset of the topics)affiliate marketing facebook marketing series online advertising small business weekbetter business bureau family business payroll tax social entrepreneurshipbusiness advice family businesses product marketing social media marketingbusiness succession planning franchise association quickbooks startup catalystcenter for entrepreneurship franchise business review referral engine startup companychamber of commerce franchise consulting rieva lesonsky startup magazinecustomer service franchisee association ryan hanley startup weekenddell small business franchisee of the year s corporation susan paytondigital marketing global entrepreneurship week sba 504 loan tj mccuedirect-response marketing guy kawasaki search marketing toilet paper entrepreneurduct tape marketing ivan walsh small biz tech tour twitter marketingentrepreneur corner linkedin small business book awards u.s. chamber of commerceentrepreneur of the year marketing management small business expo viral marketingentrepreneurs marketing strategy small business influencer viral video marketingentrepreneurs roundtable multi-level marketing small business jobs bill yahoo small businessfacebook marketing national small business week small business trends young entrepreneur council
  12. 12. Does it work?• Aligned CTR: 49% higher than broad targeted ads• Global CTR (including non-aligned): still 39% higher than broad targeted ads – “Aligned” means the subject matter targeted has an obvious connection to the ad campaign. Occasionally this isn’t the case!
  13. 13. Exceptional lift in CTR This charts the first 50 campaigns we put through Conversation Targeting at FM0.70%0.60% RED: the industry standard CTR YELLOW: the typical CTR for the parent category in FM (Sports, Business, etc.)0.50% BLUE: the CTR attained by Conversation Targeting at FM0.40% Channel CTR DFA-IB CTR0.30% CT CTR0.20%0.10%0.00%
  14. 14. Exceptional lift in CTR CT usually outperforms the channel it is in.0.70%0.60%0.50%0.40% Channel CTR DFA-IB CTR0.30% CT CTR0.20%0.10%0.00%
  15. 15. The exception that proves the rule CT rarely underperforms the channel it is in. The few exceptions are unaligned campaigns0.70% An unaligned0.60% campaign is, for example, a car ad0.50% targeted to art, museums, fashion0.40% Channel CTR DFA-IB CTR0.30% CT CTR0.20%0.10%0.00%
  16. 16. The blog audience dilemma• A network like FM’s – communities that are passionate about a common interest – presents a dilemma to advertisers• One the one hand, it represents a denser concentration of influencers and thus is desirable for advertisers• On the other hand, this type of audience may react differently to advertising – They sometimes can be seen to engage more in endemic, conversational material…. – And less with conventional display ads (if targeted using only conventional methods)
  17. 17. Two ways that FM resolves the blog audience dilemma1. FM encourages advertisers to utilize content-rich, conversation-oriented marketing pieces, e.g. sponsored content series, etc.2. For display advertising, FM encourages advertisers to use “Conversation Targeting” to make sure their ads fit better into the specific conversation happening on a blog
  18. 18. Success of CT in the blog audience CT can lift up audience engagement from below the norm, to above it0.70%0.60%0.50%0.40% Channel CTR DFA-IB CTR0.30% CT CTR0.20%0.10%0.00%
  19. 19. Lesson learned• Those who supposed that avid blog readers are necessarily more adverse to clicking on ads than the average internet user, were not quite right• It turns out, when the ads are made really relevant via semantic targeting, the CTR’s pop right up – even above Internet averages• But it means you need semantics even more for blog-based ads than for general web-based ads
  20. 20. How far down the tree will they go?• Advertisers presently Tech are moving below top- Computing Green- Tech level to “mid-level Bio-tech ontology” Enterprise Consumer for targeting Computing Computing• As this market Cloud Servers & matures, it seems Computing Net-Ops they’ll keep driving down Cloud Storage Cloud SaaS Cloud PaaS the tree Google This is going to break the AWS Storage Dropbox Cloud Storage “classifier” approach to targeting, in favor of more AWS AWS scalable approaches (can you Storage Storage Tool Kits Integrators build a 1-million node classifier tree, which morphs daily?)
  21. 21. Part Two What makes a blog post get shared more (or less)on various social networks?
  22. 22. The case studyMeasured two hundred blog posts from leadingbloggers, by these criteria: – Length of post – Estimated grade level of post (6th grade, 7th grade, etc.) – Number of: • Tweets • Facebook shares • LinkedIn shares • StumbleUpon’s • Google+1’s
  23. 23. How grade level was established: a voting engineImplemented several formulasestablished in the literature Formula #1 #2Updated the vocabulary lists #3where applicable etc….Discarded the high and lowTook a weighted average ofthe remaining scores(weightings tuned manually)
  24. 24. Effects found: post lengthWe all know that, in general, writing a longer text can turn out worse than writing less: Long version: Short version:
  25. 25. Effects found: post length• In posts ranging from 100 to 600 words in length, there is more social media amplification as the posts get longer• But this effect tapers off in the 600-1200 word- length range• For posts longer than 1200 words, the effect reverses, i.e, adding more words seems to hurt amplification• The above pattern held across all the social networks measured
  26. 26. Effects found: grade level• Generally the grade level on FM’s network was about a grade-and-a-half higher than the Internet average, sometimes more• Taken en masse, the amplification level of posts did not correlate strongly with grade level• However, on closer inspection, we saw that different audiences with opposite preferences were cancelling each other out• The biggest contrast was between Facebook-sharers and Linked-In sharers
  27. 27. Effects found: grade level Facebook vs. LinkedIn• Facebook shares, especially on posts longer than 600 words, were inversely correlated with grade level• LinkedIn shares, especially on posts shorter than 1200 words, were positively correlated with grade level
  28. 28. Effects found: grade level Summary• Grade level works together with post-length to affect amplification with some audiences• Facebook sharing is weak on longer posts unless they are at a lower grade level• LinkedIn sharing is weak on shorter posts unless the writing is at a higher grade level• Googlers and Stumblers and Tweeters lie at various points in between these extremes
  29. 29. Finding the sweet spotThe “sweet spot” wouldappear to be:• Posts between 600 and 1200 words in length• Written at a slightly higher-than-average (but not too high) grade level
  30. 30. Finding the sweet spot• Posts in the “sweet spot” get the best overall amplification – avoiding a “penalty” from any of the audiences measured 600 1200 Words words amplification 4th grade 8th grade 12th grade• Caveat: this was a limited sample and a broader study is required to validate these early results
  31. 31. ContactTim MusgroveE-mail: tmusgrove@federatedmedia.netTwitter: @tmusgrove