ML Search Ad Targeting

Overview

Meta was facing a companywide decline in ad revenue, and soon reported the first year-over-year revenue drop in its history. Two shifts were driving it at once. Apple's iOS 14 update cut off access to the cross app data that ad targeting depended on, a change Meta estimated would cost it about $10 billion in revenue in a single year. And as Reels took off, people's attention moved away from Feed, where ads were mature and monetized well, to a newer surface that carried a lighter ad load and lower intent. Meanwhile, the usual playbook of optimizing ad targeting in Feed was hitting diminishing returns.

Feed ads were mostly shown to people passively scrolling, with no strong purchase intent. But a meaningful share of Instagram's users were actively telling us what they wanted, searching for brands, products, and interests, and that surface was completely unmonetized. I saw the opportunity: instead of pushing more ads into Feed, where satisfaction was already strained, introduce ads where intent was explicit and high signal. I self-originated Instagram Search Ads, the first ads inside one of Instagram's utility features, using machine learning on keywords, engagement history, and user interests to show people sponsored results that matched what they were looking for. The bigger goal was to prove that monetization opportunities exist inside utility features, and that Meta should take advantage of every high intent signal, wherever it shows up.

AI Hypothesis

The bet the product was built to test:

If ads in Instagram Search match what someone is actively searching for, they will read as relevant results rather than interruptions, which lets Meta grow ad revenue without hurting search engagement, retention, or user satisfaction, and proves that utility features can be monetized: Meta should act on high intent signals wherever they appear, not only in Feed.

Persona & Customer Journey

Search Ads served two groups with different needs:

• Instagram users who search with intent, looking for a specific brand, product, or interest. They decide whether an ad in Search feels like a useful result or an interruption.
• Advertisers, who wanted to reach people at the moment they were looking, after iOS 14 made it harder to find them anywhere else.

We built the MVP around the user who searches with intent. The whole bet depended on them: if a sponsored result didn't feel relevant, search engagement, retention, and satisfaction would drop, and the experiment would stop no matter how well it monetized. If it did feel relevant, advertisers would follow, because those users convert.

Primary persona · composite of surveys and interviews

Maya, the intentional searcher

Uses Instagram

Daily

Uses Search for

Brands, products, ideas

Shops on Instagram

Sometimes

Ad tolerance

Low, unless it's useful

A day in her life

Maya scrolls Feed and Reels to pass the time, but opens Search with a purpose: a sneaker brand a friend mentioned, a recipe idea, a gift for a birthday next week. She wants to find the right account or product fast, and she skips past anything that doesn't match what she typed.

Goals

• Find the brand, product, or idea she's looking for quickly
• Discover options she didn't know about
• Trust that results match what she asked for
• Act on what she finds: follow, visit, or buy

Frustrations

• Ads that interrupt instead of help
• Feeling like more of her feed is ads every month
• Results that miss what she actually searched for
• Having to leave Instagram to finish what she started

Customer journey

How Maya moves from a search to an action, and where a sponsored result has to earn its place.

AI Models Overview

Search Ads plugged into Meta's existing ads system rather than starting from scratch. Three models worked in sequence: one to find ads that matched the search, one to predict which of those a person would act on, and one to keep ads out of sensitive categories.

AI Input & AI Output

Data Pipeline

Data moves through four stages, from a search typed into Instagram to a sponsored result and the metrics that decided whether it could scale.

1. Collect

   • Search queries are captured as people type them into Instagram Search.
   • Engagement history and inferred interests come from activity on Instagram.
   • Advertiser campaigns supply each ad's creative, objective, targeting, and bid.
   • Policy's rules and keyword lists define the sensitive categories.

2. Prepare

   • Queries are cleaned and normalized so spelling and phrasing variations map to the same intent.
   • The classifier screens queries and ads, and anything in a sensitive category is removed before matching.
   • Ads are turned into embeddings ahead of time, so they can be matched against a search instantly.
   • Past clicks and conversions are assembled into training data for the ranking model.

3. Match & rank

   • The two-tower model retrieves the ads closest to the search, the person's engagement history, and their interests.
   • The ranking model predicts click-through and conversion for each candidate.
   • Meta's ad auction weighs those predictions with the bid and ad quality, within the ad load agreed with the search team.

4. Deliver & measure

   • The winning ad appears as a sponsored result alongside organic results.
   • Clicks and conversions are logged and fed back into ranking model training.
   • Business metrics and user experience guardrails are tracked together, and shared with the search team as the data comes in.

System Architecture

How a search becomes a sponsored result: sources on the left, safety screening and preparation next, the models and auction inside Meta's ads system, and the results and guardrails on the right, with a feedback loop that keeps the ranking model learning.

• Ad serving path
• Policy and safety screening
• Feedback loop

Product Opportunities & Prioritization

Revenue had to come from somewhere, and there were four ways to go after it. The question was which one to bet on first, with a surface that had never carried ads before.

Why Search came first

The core insight was relevance over volume. More ads in Feed had a measured cost to user satisfaction, and Discovery depended on data that was no longer reliable. Between the two intent rich surfaces, Search won the priority over Discovery: it had the cleanest intent signal and the fastest, most defensible path to proof: if an ad matched the query, it worked more like a relevant result than an interruption. A survey confirmed that people responded positively to ads closely matched to their search intent, which gave us the evidence to get approval. Sequencing mattered too: proving Search first built a credible case for monetizing utility features, and I launched Profile Visit Ads about a year later once its measurement was solved.

Tradeoffs

Product Roadmap

I self-originated Instagram Search Ads and owned its product strategy and roadmap. The roadmap followed the prioritization: prove ads could work in Search first, with guardrails the search team trusted, then scale. Every phase had to clear both business metrics and the guardrails the search team cared about before the next one began.

User Stories & Acceptance Criteria

The stories center on the Instagram user who searches with intent, the primary persona, alongside the advertisers who buy the ads and the teams whose trust the launch depended on. Each acceptance criterion is something the team could test, and many come straight from the guardrails agreed with the search team.

Risks & Mitigation

Putting ads into Search for the first time meant risking two things at once: the experience people relied on, and the trust of the team that owned it. Each risk below was rated on how much damage it would do and how likely it was, and most mitigations were guardrails agreed before the test began.

Business Model / Lean Canvas

Search Ads was a new revenue surface inside an existing ads business, so it sold through Meta's existing advertiser channels rather than a new go-to-market. The canvas below sums up the business on one page.

AI Product Scalability

Instagram Search handles an enormous volume of queries, and every one is a chance to show an ad. The product was built so that more searches made it more accurate and more valuable, without slowing search down or loosening its safeguards.

MVP

The MVP did one job: put a single, relevant sponsored result into Instagram Search, and prove it could earn revenue without hurting the search experience. The storyboard follows Maya from a search to a purchase, and what the system does at each step.

• Sponsored results matched to the search query, engagement history, and interests
• Sensitive category filter screening every query and ad before matching
• Ranking and auction through Meta's existing ads system, within the ad load agreed with Search
• Guardrails and a stop rule tracked alongside revenue from the first day of the test

Results

I led the development and launch of Instagram Search Ads, a machine learning ad product that matches ads to search intent using keywords, engagement history, and user interests. I directed a cross functional team of AI/ML engineers and data scientists through an A/B test with 5% of users, and scaled to a full launch once both the business metrics and the user experience guardrails proved healthy:

Guardrail metrics such as search engagement and retention stayed stable, proving that monetization and search quality could coexist. The project also produced a reusable governance framework for monetizing utility features at Meta, which let Discovery Ads launch later with far less friction.