Meta Ads Deep Funnel Optimization: What Hierarchical Interest Signals Mean for Your Creative Strategy

Meta's engineering team recently published a paper that most performance marketers will never read but all of them will feel. It's titled "Exploring Hierarchical Interest Representation for Meta Ads Deep Funnel Optimization," and beneath the academic language sits a consequential shift. Meta is no longer modeling your customer mainly through what they've clicked. It's building a map of how users, products, advertisers, and content relate to each other across the platform, then using that map to find people with "genuine, latent interest" in what you sell, including people who have never touched your brand.

This is not a minor ranking tweak. It changes how Meta decides who is worth showing your ad to, and that has direct consequences for how much creative variety you supply, what signals your assets carry, and how you test from here.

What deep funnel optimization actually means in Meta Ads

"Deep funnel" means the bottom of the conversion path: purchases, subscriptions, high-value leads, repeat customers. These are the events advertisers care about most and the ones Meta has always had the least data on, because they are rare relative to impressions and clicks. An account can rack up huge impression counts and still log only a modest number of purchases each month. That sparsity is the core problem. The algorithm has to learn from a small, noisy signal and generalize it to people who haven't seen the ad yet.

Meta's older answer leaned on pixel data, lookalikes built from purchaser lists, and predefined interest categories. Hierarchical Interest Representation changes the architecture underneath. Instead of treating users, products, and advertisers as loosely connected entities, the research describes a transformer-based graph model that learns a unified embedding space linking all of them. A pair of running shoes isn't just "shoes." It sits in a web of relationships with trail running, endurance sports, outdoor gear, and technical apparel, and the model learns those connections from engagement patterns across the platform, not only from your account.

The detail worth sitting with: the research describes blending engagement signals with multimodal content understanding, using LLMs to process the actual content of ads and product pages to enrich sparse interaction data. Your creative is becoming an input into audience modeling, not just a variable tested against a fixed audience.

How hierarchical interest models change ad delivery

The practical payoff is what the paper calls generalization to "rare and unseen entities." In plain terms, a new advertiser or a new product line with little conversion history doesn't start from zero. The model can infer where your product sits in the broader interest graph from what the creative communicates and how similar products have performed, then route it toward users with latent affinity for that cluster, even if they've never searched for anything like it.

That breaks the old mental model of "feed the pixel volume and wait." Volume still matters. But the system now draws on a richer, cross-advertiser graph to fill gaps, and because that graph is informed partly by ad content, your creative has to carry more information than it used to.

Why creative diversity matters more as targeting gets smarter

Here's the counterintuitive part. As targeting gets more sophisticated, the advertiser's burden doesn't shrink. It moves. Advantage+ has largely absorbed manual audience slicing, but you are now more responsible for supplying the signal diversity the model needs. If the system relies partly on reading ad content to enrich sparse engagement data, an account running three near-identical variations gives it almost nothing. The model can't tell whether your product belongs in "performance running" or "casual athleisure" when every ad looks and sounds the same.

The shift isn't about targeting less precisely. It's about letting creative variety do work manual targeting used to do.

This matches what we've seen with Advantage+ creative versus manual creative setups. Advertisers who held onto granular targeting control were, functionally, working against the direction of Meta's whole stack. Hierarchical interest representation is more evidence that the targeting and creative layers are converging, and creative is fast becoming the main lever advertisers still hold.

Your creative set is no longer just what people see. It's part of how the algorithm decides who gets to see it.

Practical steps: feeding Advantage+ better creative signals

So the question becomes operational: what do you change in how you brief, produce, and upload? Several things matter more now than they did two years ago.

1. Diversify angles, not just execution. An angle is the core argument for buying: price, status, convenience, fear of missing out, social proof. Five ads with different visuals and the same angle teach the model less than two angles with two executions each.
2. Show product attributes in the creative, not only the copy. If your product fits several interest clusters (a shoe that's both performance gear and a lifestyle item), show both contexts across the set instead of picking one and hoping targeting covers the rest.
3. Use format variety on purpose. Static, carousel, and video carry different information density and drive different engagement patterns. Format diversity gives the model more distinct data points, not just more impressions.
4. Make sure Advantage+ settings aren't flattening your variety. Music, text overlays, and image expansion can help, but if they push every ad into the same visual template, you've undone the diversity you built. Review your Advantage+ creative settings so automated enhancements support your variety strategy instead of eroding it.
5. Keep event data clean. None of this works if your Conversions API sends delayed, duplicated, or mismatched events. Meta's own guidance on web and app event optimization is clear that accurate, deduplicated, timely signals are the foundation. Richer modeling enriches sparse data. It doesn't fix bad data.

How this shifts your testing strategy going forward

If creative variety now does some of targeting's job, your framework has to move from "which single ad wins" to "which mix of angles and formats gives the algorithm the richest signal while still hitting efficiency targets." That's a different optimization problem, and it needs a different structure.

Three distinct angles, two formats each, two hooks per format: twelve genuinely different pieces of signal instead of twelve cosmetic variations of one idea. It's the same discipline as our creative testing framework for 2026, now with a sharper reason behind it. You aren't only finding your best ad. You're teaching the ranking system what you sell and to whom.

Plan for one more change: shorter, broader test cycles. Rather than running a single ad for weeks while pixel data trickles in, launch wider creative sets earlier. If the model is less dependent on your account's isolated conversion history, as the "rare and unseen entities" framing suggests, new advertisers and new product lines should get usable signal sooner. That only holds if the creative is varied enough for the model to read.

Where this leaves your 2026 creative strategy

Targeting isn't dead, and account structure still matters. But the lever that moves performance has shifted further toward creative, specifically creative varied in angle and format rather than in surface details. It's the same direction we've tracked across Meta's platform changes, so read this alongside our broader take on Meta ad creative strategy for 2026 if you're rebuilding your approach.

The advertisers who adapt fastest won't be the ones with a clever workaround. They'll be the ones who accept that creative volume and diversity aren't a layer on top of smart targeting. They're the raw material the targeting system now runs on.