Using Contextual to Define Audiences Demands Choosing the Right Pages

Not all content pages are created equal.  We’d like to focus on how we wind up choosing individual content pages for a segment, why we do it that way and what we find the benefits are.

There are a couple of things at the top that are important to understand about our platform’s approach.  

First is we tightly manage the domains that we take pages from using tools available in the ecosystem and more importantly, tools we’ve built ourselves.  We live in a well lit part of the ecosystem and our partners all work to remove garbage domains.  We then maintain a list of domains that we don’t find acceptable past the partner efforts.  Finally each segment is inspected to see if anything slipped through the first two filters and adjusted as needed.  

Second is we only ever include individual content pages in a segment.  We use individual article pages at scale to define an audience.  We never include  home pages or a channel fronts.  Individual content pages have stable content on them which can be run through our topic and keyword models toward identifying and targeting an audience.

After we’re done removing domains that we know aren’t useful at best and fraud at worst we wind up with 125 thousand sites and 3.5 billion pages in our index.  Our segment creation tool is what lets us find both pages with the right topics and the right keywords on them to deliver an audience for a brand.

We built the creation tool to support two contextual processes; first determine the topic of a page and then determine the keywords on that page.  The right keywords are often found in content that has little to do with the keywords from a commercial perspective.  No travel brand wants to target good travel keywords in a news article about families escaping wars which occur everyday.

The analyst building the segment leverages the platform through these steps;

1. Training set pages - each segment gets a handful of pages that our analyst finds that are about the topic that brand needs to align with.  Our topic model breaks each training set page into a series of topics that can be inspected before using.  So we always know the environment is suitable.  Suitability to us means travel keywords in travel content.

2. Keyword adjustments - once the foundation of a segment is created using training set pages, we can inspect the keywords that are found in the segment.  We can add or delete keywords and phrases and we make it very easy to make those adjustments.  

3. Final adjustments - once we have a candidate segment we make one last check to see if there’s any remaining content the brand doesn’t want to be near.  Here we’ll remove anything that looks irrelevant or looks suspicious.  

We think about choosing pages in some detail as you can tell.  And if you’re thinking our platform must be transparent then you’re correct.  There are benefits from taking all the measures we do.

1. Higher engagement - our approach consistently delivers higher engagement metrics than other contextual or other targeting options.  Higher engagement is the intent metric that can be mined from content and our approach delivers that higher engagement.

2. Less waste - we target keywords better because we understand the environment they appear in using our topic engine.  That understanding means there’s less wasted spend than other contextual engines.

The intersection of better engagement and less waste is the reason most of our brand customers have been with us for three years.

Defining and Targeting Audiences Using Context

We are always thinking about what contextual means and how to best use it to define and target an audience.  Contextual analysis is simply understanding the pages that people are reading when they’re at a particular task.  It sounds simple, and, in concept, it is.  

For the entire history of contextual analysis including today, most contextual analysis occurs by assigning keywords to domains or pre-categorizing channels of domains into one segment that can be targeted.

We think these approaches fall apart in terms of accurately defining an audience. If you assign keywords to a domain then you assume that every page under that publisher’s domain is about the same keywords.  Clearly, that’s not the case.  If you group sections into a contextual segment for health, then you’re assuming that each page is also about the same thing.  Also, demonstrably not true.  We realized that the problem we had to solve was how to not think about the context of domains or channels and only think about pages that are similar.  That led us to our platform innovation.  

We decided our customers would be best served if we could look at content at the individual article level and with more granularity than a traditional keyword approach could provide.  Our initial response to that challenge was to build an engine which could process billions of pages at the rate of 1,000,000 impressions per second.  

Our second solution was to build an AI topic model with extensive keyword capabilities so we could understand both the topic of the page and the keywords on that page. This allowed us to to easily allow and block any keyword that creates noise and waste; it allowed us to better tune a segment for performance.  Finally, we can filter attributes of content such as tone and sentiment.  Our topic model provides dependable guard rails and the keywords create additional performance.

The combination of looking at content consumption at the page level, with the granularity of both topic and keyword analysis was an effective solution for our clients.  Rather than provide an inclusion list of domains or channels with a few keywords attached, we began making segments with extensive lists of individual content pages (urls) from many publishers which had similar topics and keywords.

It’s a very practical solution that scales and performs.  We know it scales and performs since our brand and agency customers continue to spend more and add brands to their business with us.  We have a brand customer retention rate of greater than 70% with customers which have used us for three or more years.  

We also made it a very efficient exercise.  Our tools don’t require anyone to be keyword or search engine experts nor does it require extensive tables of allowed or blocked keywords.  We figured that simple is better; We understand how the system works and can explain it to clients.

When we create a segment, we have to find pages with similar topics. Our first step is to we feed our engine several seed articles. We’ve got topic classifications on about 3.5 billion individual article pages so when we put the seed pages into the engine, it searches the article database and finds all the pages where the topics match.  Now we have the basis of a segment which can have thousands or millions of pages that match the topics.

The second step is to understand and adjust the keywords that appear on the pages with similar topics. Together, with our clients, we share the keywords in the articles we think are relevant. Once we’ve shared the keyword analysis and gotten their feedback, we adjust the number of keywords in the segment, both included and excluded.

Between these two processes our brand and agency customers now have a scaled, well defined audience they can target in a number of ways.  And it works.  With a large, enterprise software company we deliver audiences with higher engagement than any publisher on the plan, which is their main KPI.  In travel we deliver a premium audience for a luxury hotel measured by a consistently low cost per visit with high ROAS.  

Contextual becomes increasingly important as platform changes and privacy laws put pressure on traditional targeting in digital and social media.  With less data available due to new laws and a push by platforms to eliminate user guided targeting, we provide a scale and performant option for brands to define and target an audience across multiple media channels.

What, Exactly are We Curating?

We know a fair bit about curation.  It’s what our platform does and has done since 2018.  In fact, it was built to curate as a core function.  So we’ve had experience in curation and our brand customers benefit from it.  More than 70% of our brand customers have been with us for 2.5+ years.  That retention rate indicates we’re able to deliver what works at scale and make sure our customers avoid the steady torrent of crap that we all read about.

To give you a sense of our scale,  we curate content at the rate of 1mm impressions per second across 3.5bn URLs.  We aggregate those, dynamically, into segments that deliver scale and performance to our brand customers.  For curation, our contextual topic model and keyword analysis are the only two pieces of data we’ve ever used.  

The content we curate are individual pages.  We curate only article pages and never home pages or channel fronts.  We have learned that the strongest signal comes from article pages; that home pages and channel fronts have little signal.  Article pages are what consumers focus on when they’re at task.  They gain information from them and that information forms a mindset with the reader. That’s what we’re after.  

We continue to curate our content segments in real-time based on the scale and performance our brand customers provide us as feedback.  This creates a cycle where optimizations can and do happen by adjusting the words and the weighting of words our topic and keyword models deem relevant based on scale and performance.

With this cycle of continuous curation of page-level content into segments, we’re able to help our brand customers navigate the landscape of legacy data that’s either changing or going away.  Over time, we’ve begun to understand there are topics and KWs that consistently indicate the mindset of the consumer who’s reading about a topic.  

We see similarities across brands in the same categories with words and topics that create different responses.  With the knowledge of a category across many brands we can create other filters to better understand a page.  The cycle goes on.

That’s our version of curation and it has been proven over the years to be effective.  

What we don’t do is use curation as a marketing tool for inventory and domains.  In fact, we think focusing on domains and inventory misses the point of curation.  And might be a bit of a dodge.  

To us, the inventory a brand targets is a tactical optimization and they should be able to separate that from the audience they’re after.  Some segments might respond best with one ad size and another brand in the same category might find success with another size.  So why try to be the arbiter of optimization?  Our function is to provide curated content that delivers an audience.  The inventory is up to the brand and they like that control.

Then we’ve read that curation rids segments of MFA sites.  To that we say anyone that needs curation to remove MFAs isn’t trying hard enough.  MFAs have to be excised at the platform or ecosystem level first, as close to the publisher as possible.  If the only place in a workflow to excise worthless content is the segment, the logical outcome is you won’t remove all of it.  

We should all get behind curation.  But we should raise the bar on what curation means and not lower it to suit the result of our current efforts.  Or problems.  Feel free to reach out to us and we’ll show you how curation works.