The Meta Andromeda update just got a turbo boost.
Meta has refined the way its advertising system reads people’s interests and intentions. It can now better distinguish between straightforward “I’m ready to buy” buying intent and more complex or aspirational interest – even more effectively than before. This allows the system to better serve ads that match the user intent closer than ever.
How much more effective is it? How does it achieve it? And what does this mean for advertisers and brands? Read below to find out.
Have you ever heard the saying “we can do it well, we can do it cheap, and we can do it fast, but you can only pick two?” Well, that’s basically the problem Meta faced when it came to using AI to decide which ad to show a user. Only worse…
AI, with it still being a rapidly developing technology, can often only do one of these things well, rather than two.
Meta can build (and has built) highly advanced AI to find the perfect ad to show you, based on all kinds of data it has about your behaviour and intent. However, it certainly won’t be quick, as it takes time to explore all the pathways it has and process unfathomably large datasets. In this use case, anything over 1 second is already considered snail-paced. And as we know too well, AI is not cheap, especially when it’s working so hard. We only need to look at RAM prices, tech stocks, and the billions of pounds and dollars being thrown around AI companies to know that.
This three-pronged issue is what Meta has given the fancy name of the “inference trilemma”, which sounds oh so very Silicon Valley!
Andromeda + Adaptive Ranking Model
With the Andromeda update, Meta has already changed the game in terms of how ads are delivered to users. We now live in an age where Meta knows our target audiences better than we do. Meta’s AI systems know who you are, what you like and dislike, what your goals, ambitions, and fears are, and most importantly now, whether you’re genuinely in the market for that new motorcycle or if you’ve just been tyre kicking for the past 4 months.
Now that may sound like a giant leap for advertiser kind – and that’s because it is/was – but what I’m here to tell you about today is Meta’s new improvement upon this. Although the Andromeda update sounds amazing, Meta isn’t one for resting on its laurels, and recently, Meta has been speaking about implementing LLM-scale models into its ‘Ads Recommender’. They call this newly implemented improvement the Adaptive Ranking Model.
Or in simpler terms, using even more powerful AI to recommend ads to you even more effectively.
Let us save the technical details till last. First, let’s talk about what the result of Meta’s Adaptive Ranking Model actually means for us.
What does this mean for advertisers and brands?
Firstly, Meta has quoted a performance increase in conversions of +3% and an increase in click-through rate of +5%. Bear in mind that although this gain is very incremental, advertisers don’t really need to do anything different to experience this gain. If you think of this as an improvement upon the Andromeda update, as long as you are building your strategy around the changes we’ve seen recently, your campaigns should benefit from this. And although 3% extra conversions sounds like nothing, this can equate to much larger effects when applied to large-scale, large-budget accounts.
The second interesting result of Meta’s new Adaptive Ranking Model is that broad audiences should become more performant. We can add this update to the list of events that make Meta’s broad targeting more powerful (and less clumsy). The AI is now working even more effectively when trying to match your ads to ideal audiences.
Which brings us to the third and final takeaway of the new Adaptive Ranking Model: creative just got even more important (again). If you’re already tired of reading/listening to/watching content like this that borderline nags you to focus more on your creative or gives you the old “creative is the new targeting” cliche, then you won’t like what I’m about to say. This new LLM-scale model improvement means that Meta is even better at understanding which of your creatives suck and which rock.
There are some other things to consider, such as how auctions should now have slightly improved consistency as a result of the new Adaptive Ranking Model, meaning we advertisers will see fewer random performance swings, which is great! Also, this continues Meta’s direction away from manual optimisations and further leans into the agenda of Meta just wanting us to input ads and budget into the platform and let it handle the rest. However, the top three aforementioned impacts are the ones to pay the most attention to.
How does it actually work?
So, those are the impacts, but once you understand a bit more about how it works, it may change the way you plan your next strategy.
To start off, super simplified, before going more in depth, Meta’s new Adaptive Ranking Model essentially works by now having a more advanced AI available and ready to help with serving relevant ads. In the very recent past, what would happen every time you open Facebook or Instagram, or scroll through your feeds, Meta would simply use a fixed-complexity ranking system, which was good at tailoring ad recommendations quickly and cost-efficiently, but it was lacking the intelligence that the new ARM brings.
What happens now, as users scroll through their feeds, is that Andromeda will use ARM to decide if it’s easy to determine which ad should be shown. If it’s easy, Meta will use simpler AI models that are faster and cheaper. If it’s a more complex scenario, Meta will use their new LLM-scale AI models, which are slower and more costly.
This means that both simple and complex user intent scenarios are now well serviced in an efficient manner. Whether you have been buying motorcycle gear non-stop for the past 3 months, or just fantasising about buying that sportbike even though you haven’t got your licence yet, Meta will be able to better serve you the right ad to match your intent. It uses the right tool depending on the job.
If you want to get more technical, one thing that’s really cool about the new Adaptive Ranking Model is that it achieves LLM-level performance at a fraction of the compute time. When using tools like Claude or ChatGPT, the time it takes to answer queries is measured in seconds. With Meta, response times needed to be a fraction of a second; otherwise, users would be facing a significant negative change in their user experience (e.g., waiting for their feed to load or the app taking longer to open).
The tech wizards at Meta have been able to achieve “model complexity equivalent to the O(10 GFLOPs) per token used by top-tier LLMs” while operating at an “order of magnitude faster than standard LLM inference, maintaining O(100 ms) bounded latency.”
This is an incredible feat of technological prowess and one that benefits brands and advertisers, as long as the correct strategies are implemented to coincide with updates like these. For those of you interested in the technical nuances, like request-oriented sequence scaling or hardware-aware graph and kernel specialisation, I would encourage you to read Meta’s write-up.
We pride ourselves on geeking out about AI updates like these and even have marketing technology and innovation specialists to help us not just keep up with the rapidly evolving landscape, but stay ahead of it. At Found, we always tout the importance of taking advantage of these platform updates and are real champions of the importance of great and varied content in Paid Social. If your brand is looking for help with running fine-tuned campaigns for Paid Social, why not get in touch with us?
