Meta Just Explained How It Reads Your Ad Creative to Guess Who Wants It. Here's What That Actually Means

Meta’s engineering team published a research post in July 2026 describing a new system called Hierarchical Interest Representation, or HIR.

It’s dense, genuinely technical material, the kind of thing written for machine learning engineers, not media buyers.

But the core idea is simple enough to explain plainly, and it matters for anyone running Meta ads: Meta is teaching its ad system to read the actual content of your creative, the words, the images, the video, and use that understanding to figure out who’s likely to want what you’re selling, even when it has almost no direct evidence that person is interested.

Here’s what HIR actually does, why Meta needed to build it, and what it means for how you think about creative and targeting going forward.

We’ll use a few made-up, illustrative examples along the way to keep the technical parts grounded, none of these are real advertisers or real Meta data.

The Problem Meta Was Actually Trying to Solve

Start with the problem, because it explains why this system exists at all.

Meta runs ads for millions of advertisers, showing them to billions of people, every month. Most of the time, most people don’t convert.

A purchase, a sign-up, a real “deep funnel” action, is a rare event compared to the sheer volume of ads shown. That means the signal Meta actually has to learn from, real people taking real actions on real ads, is thin relative to how much guessing the system has to do.

Picture a fictional example…

Imagine a small skincare brand called Verdant Botanicals running ads to a broad U.S. audience. Out of a million people who see the ad this month, maybe a few thousand click, and a few hundred actually buy.

Meta’s system has to somehow generalize from those few hundred purchases to make good guesses about the next million people it shows the ad to, most of whom look nothing like anyone who’s bought yet.

That’s the scale of the sparsity problem HIR is built to address.

The Core Idea: Compressing a Messy Graph Into Something Learnable

Underneath Meta’s ad system is essentially a giant graph, users connected to ads, ads connected to advertisers, advertisers connected to products, all linked by actions like viewing, clicking, or buying.

At Meta’s scale, this graph is enormous and the connections between any two points are usually sparse. Most users have never interacted with most products.

HIR’s first move is to compress that raw, sparse graph into something denser and more useful.

Instead of trying to reason directly about the relationship between one specific user and one specific product, HIR learns a smaller set of what Meta’s engineers call “interest primitives,” essentially, learned clusters that sit between users and products.

A user doesn’t map directly to a product.
Both the user and the product map into a shared space of these learned interest clusters,
And the system reasons about relationships through that shared space instead.

Think of it like a librarian who’s read every book in an enormous library and organized them not just by title, but by deep, overlapping themes, so that even a reader who’s never picked up a specific book can still be pointed toward it based on themes they’ve shown interest in elsewhere.

The interest primitives are Meta’s version of those themes, except they’re learned automatically from data rather than assigned by a human.

Where Creative Actually Comes In

This is the part that matches what you were describing.

HIR doesn’t just learn from behavior, clicks, views, purchases. It also pulls in what Meta’s post calls “world knowledge”: the actual content of the ad and the product itself, text, images, and video, processed through language and vision models and folded into the same representation.

In plain terms: the system isn’t just watching what people do around a product. It’s also actually looking at what the product and the ad creative say and show, and using that understanding to fill in gaps where behavioral data is thin or missing entirely.

Here’s a fictional scenario to make that concrete.

Imagine a new advertiser launches a product Meta has never seen before, a weighted recovery blanket from a made-up brand called Restwell.

There’s no click history,
no purchase history,
nothing behavioral to learn from yet.

Under the old approach, Meta would have very little to go on.

Under HIR, the system can process the actual creative,
images of the blanket,
copy about deep pressure and better sleep,
and connect that content to the interest primitives it’s already learned are associated with sleep aids, stress relief, and recovery products…

Even though it’s never seen a single person engage with this specific blanket before.

The creative itself is doing real inferential work, not just serving as the thing a person clicks on.

How This Connects to Meta’s Other Systems

HIR isn’t a replacement for the systems you may have already heard of.

Meta describes it as an upstream layer, meaning it feeds into the systems that actually decide which ads get shown to whom: GEM (Meta’s Generative Ads Model), Andromeda (the ad retrieval engine), and the Adaptive Ranking Model.

HIR’s output, compact “tokens” representing a user’s or an ad’s interest profile, are designed to plug into those downstream systems to make their targeting and ranking decisions sharper, especially for the sparse, harder-to-predict conversions further down the funnel.

What This Actually Means for How You Run Ads

None of this changes the basic mechanics of setting up a Meta campaign.

But it does reinforce something worth internalizing if you haven’t already: your creative is doing more targeting work than it used to, and that work increasingly happens at a semantic level, not just a behavioral one.

A few practical implications:

Your ad copy and imagery are literally training data for who Meta thinks should see it. Vague, generic creative gives the system less to work with when it’s trying to match your product’s actual content to the interest primitives that fit it. Specific, clear creative gives it more to reason from, independent of how much behavioral data your account has accumulated.

New products and new advertisers benefit more from this than established ones. The whole point of pulling in creative content and product information is to handle situations where behavioral data is thin or nonexistent, which is exactly the position a new account or new product launch is in.

This is one more reason broad targeting with strong creative tends to outperform narrow manual targeting right now. If the system is increasingly inferring interest from the content of your ad itself, narrowing the audience manually works against the very mechanism designed to find the right people for you. Practitioner data backs this up directionally: one 2026 creative benchmark analyzing over 550,000 ads found only 5 to 8 percent of ads become real winners, with roughly half getting little or no spend at all, underscoring how much weight creative variety now carries in the system.

None of this tells you whether it’s actually working for your account. HIR is a research system aimed at improving Meta’s internal targeting and ranking. It has no bearing on whether the platform is reporting your results accurately. A more sophisticated targeting engine can still sit on top of the same attribution and measurement questions AdBeacon has written about elsewhere, since a better-targeted ad and an accurately-measured ad are two completely different problems.

If you want to see how your actual first-party conversion data compares to what Meta’s increasingly sophisticated targeting systems report, book a live AdBeacon demo.

—-

FAQ

What is Hierarchical Interest Representation (HIR)?

It’s a research system Meta published in July 2026 that learns compressed, shared representations connecting users and ad entities (advertisers, products, campaigns) by combining behavioral engagement data with the actual content of ads and products, processed through language and vision models.

Does HIR replace Andromeda or GEM?

No. Meta describes HIR as an upstream layer that feeds into those systems, providing them with richer interest signals to use in retrieval and ranking, rather than replacing any of them.

Does this mean Meta reads my ad copy to decide who sees it?

Broadly, yes. HIR pulls in text, image, and video content from ads and products and uses it, alongside behavioral signals, to infer which users are likely to have genuine interest, particularly in cases where behavioral data alone is too sparse to make a confident prediction.

Does this change how I should target my campaigns?

It reinforces an existing trend rather than introducing a brand-new one: broad targeting paired with clear, specific creative tends to give Meta’s systems more to work with than narrow manual targeting, since creative content itself is increasingly part of how the system infers audience fit.

Does a better targeting system mean Meta’s reported results are more accurate?

Not necessarily. Targeting and attribution are separate problems. A more sophisticated system for finding likely-interested users doesn’t change how conversions get measured or reported, which is a separate question worth verifying independently.