We seem to be in that sweet moment where artificial intelligence (actually advanced machine learning, but who are we to split hairs?) is more than just buzz and potential, but before the machines have enslaved us and forced us to farm their massive energon cube plantations.
That day is coming—I suggest you start tending a power strip garden now so you are prepared for the energon harvest.
Yes, AI is actually proving useful in chatbots and virtual assistants, and believe it or not: digital advertising. Publishers are excited at having AI comb through their vast supplies of data to optimize inventory pricing in real time on programmatic markets. But hold on, says Roxot Marketing Director Alex Kharitoshin—this practice can have some serious drawbacks.
GAVIN DUNAWAY: What are the chief downsides to AI-based dynamic inventory pricing?
ALEX KHARITOSHIN: Three things.
1. AI-based dynamic inventory pricing doesn’t work effectively with a small amount of data/impressions.
AI needs data to learn: it adapts to how buy-side algorithms react to different floors. The more data, the faster the system understands what floors are the most effective. The less data, the slower the system. A slow system is less effective.
Thus, AI-based dynamic pricing doesn’t fit small publishers. We estimate 20 million monthly AdX impressions to be the bare minimum for AI to be effective.
2. AI-based dynamic inventory pricing is a black-box technology. Nobody will disclose the details about how exactly AI-based dynamic inventory pricing works. We use machine learning and thousands of mathematical models to identify optimal floors for every ad request. There are too many technical details to explain.
Also, the technology and “know-how” are the main competitive advantages. It makes sense to keep them protected. That’s why there’s a certain level of magic/secret sauce when people describe dynamic inventory pricing.
3. You can’t precisely forecast the revenue uplift before testing.
The effectiveness of AI-based dynamic inventory pricing varies from site to site. Tech suppliers can’t promise a certain percentage of revenue lift without testing. We usually provide publishers with the average uplift across our clients and insist on a test. A short test will predict the possible uplift with high precision.
GD: What “advantages” of AI-based dynamic inventory do you think are over-hyped?
AK: It’s hard to say what advantages are over-hyped because there’s only one main advantage—an increase in programmatic revenue. I’d say that certain approaches to dynamic pricing are over-hyped. For example, using a bid from one SSP to floor the other is not effective as people think.
Some companies and publishers pass header-bidding bids multiplied by a coefficient to AdX. We’ve been working on dynamic pricing for 3 years and can confirm this strategy doesn’t push Google to pay more for your inventory.
A dynamic pricing strategy should depend on demand and algorithms inside an SSP you price. It guarantees you consistently price 100% of ad requests and adapt to the performance of this SSP.
Are there other AI benefits outside dynamic pricing that you think publishers are not taking advantage of?
The main benefit of AI/machine learning is the ability to learn from data and find meaningful connections humans can’t find on our own. Publishers have access to ad-auction data even if they do not intentionally collect it—not many realize DFP is a free database. AI is a tool to learn from and act on this data in real time.
What do you think will be the biggest developments in AI for publishers over the next year?
I believe that automated ad stack optimization is an opportunity for publishers to use AI for revenue growth. Using ad auction data from both Google AdX and header bidding, AI technologies can adapt publisher’s stack to every ad request: what SSPs to call, what layout to use, what timeout to set, etc. That’s what publisher will never accomplish manually optimizing ad stack.