Meta's AI automation draws skepticism from advertisers despite performance claims

Meta unveiled Andromeda on December 2, 2024, describing the system as a machine learning retrieval engine that delivers what the company calls "a step-function improvement in value" for advertisers. Yet the announcement arrived amid mounting industry skepticism about the platform's aggressive push toward AI-driven automation that reduces advertiser control over targeting, creative, and placement decisions.

According to the Andromeda announcement, the system processes tens of millions of ad candidates and narrows them to thousands of relevant options within strict latency constraints. Built specifically for Nvidia Grace Hopper Superchips, the retrieval engine achieved a 6% recall improvement and delivered an 8% ads quality improvement on selected segments across Instagram and Facebook applications.

The technical achievement addresses real scalability challenges. Meta's Advantage+ suite automates audience creation, budget allocation, dynamic placement across surfaces, and creative generation. According to the announcement, advertisers who activated Advantage+ creative features experienced a 22% increase in return on ad spend. Businesses using image generation saw a 7% increase in conversions. More than one million advertisers created over 15 million ads in a single month using generative AI tools.

Industry professionals voice control concerns

Digital marketing specialist Bram Van der Hallen challenged what he characterized as "absolute nonsense" surrounding Andromeda-style campaign consolidation in a November 27, 2025 LinkedIn post. According to Van der Hallen, marketers promoting single-campaign approaches claim these setups can reach entire audiences across every funnel phase, but the reality proves more complex.

"Reach expands, but impact doesn't follow," Van der Hallen wrote, explaining that increased reach driven by combining broader audiences with more creative combinations generates noise rather than relevance. When click-through rates rise, the improvement may result from ad formats generating cheaper clicks rather than increased intent. Cost per result drops can mask underlying problems when campaigns mixing top-of-funnel, middle-of-funnel, and bottom-of-funnel creatives show improved metrics driven primarily by retargeting audiences.

The critique highlights a fundamental tension. Meta's AI advertising approach emphasizes automation over manual control, with the platform reporting a $20 billion annual run rate and 22% return on ad spend improvements. However, advertising experts warn of "brand damage at scale," citing demographic mismatches, budget catastrophes, and creative disasters.

Attribution and incrementality concerns raise questions about reported performance. One incrementality test found Advantage+ generated only 17% of conversions reported by Meta's attribution—a dramatic discrepancy suggesting the system captures existing demand rather than creating new demand. Backbone Media's study showed cost-per-thousand impressions inflated 10x in February 2024 with volatile return on ad spend.

Technical architecture versus transparency demands

Andromeda employs what Meta describes as "a highly customized deep neural network with sublinear inference cost," enabling a 10,000x increase in model capacity for enhanced personalization. The system uses GPU preprocessing for feature extraction, storing all precomputed ad embeddings and features in the Grace Hopper Superchip's local memory. This approach addresses traditional scaling constraints of limited CPU-to-GPU interconnect bandwidth and low GPU utilization.

Hierarchical indexing organizes ads into multiple layers, reducing inference steps by focusing only on most relevant nodes. According to the announcement, the hierarchical index and retrieval models train jointly, aligning index representations with neural networks to improve both precision and recall compared to two-tower neural networks or approximate nearest neighbor search.

Model elasticity enables segment-aware design that leverages higher complexity models for high-value ad segments. The system automatically adjusts model complexity and inference steps in real-time based on available resources. Together with hierarchical structured neural networks, model elasticity boosts model inference efficiency by 10x.

Yet this technical sophistication operates within what critics characterize as a black box. The lack of transparency makes it impossible to diagnose whether performance drops stem from creative fatigue, audience saturation, competitive pressure, or algorithmic changes. Advertisers cannot determine true incrementality versus harvesting low-hanging fruit.

Harvard Business School research identified five pitfalls specific to AI marketing automation: people blame AI first when things go wrong; when one AI fails, people lose faith in others; people place more blame on companies that overstate AI capabilities; people judge humanized AI more harshly; and people feel outraged by deceptive AI practices.

Platform evolution accelerates automation defaults

Meta removed detailed targeting exclusions in January 2025, citing 22% better performance for campaigns without the feature. The company enabled Dynamic Media by default for Advantage+ Catalog ads starting September 2025, with 100% enforcement by October 20, 2025.

In October 2025, Meta introduced a default option allowing the platform to spend up to 5% of budget on each excluded placement in sales and leads campaigns. The feature activates by default unless advertisers manually opt out. According to marketing professionals commenting on the implementation, the approach represents Meta taking control "under the pretext of improving performance."

The placement control update raised questions about brand safety concerns, particularly for advertisers who exclude Meta's Audience Network for specific reasons. Audience Network places ads on external apps and websites beyond Meta's owned properties, creating potential brand suitability challenges for advertisers with strict content adjacency requirements.

Meta deprecated legacy campaign APIs for Advantage+ structure consolidation in October 2025, with full migration required by Q1 2026. This technical consolidation reflects the company's strategic direction toward AI-powered campaign management as the default advertising experience.

Measurement integrity concerns intensify

A former Meta employee alleged in an August 20, 2025 whistleblower complaint that the company artificially inflated return on ad spend metrics for Shops ads by counting shipping fees and taxes as revenue. According to the complaint, Meta data scientists discovered during internal reviews in early 2024 that Shops ads return on ad spend had been inflated between 17% and 19%.

The discrepancy resulted from Meta including shipping fees and taxes in sales calculations, even though merchants never received that money. Meta's standard advertising products excluded these figures, following the same approach as competitors. Without shipping fees and taxes, Shops ads performed no better than Meta's traditional advertising campaigns.

Meta announced restrictions on October 13, 2025, that will eliminate two view-through attribution windows starting January 12, 2026. The 7-day view-through window and 28-day view-through window will stop returning data. Recent discoveries revealed that Meta counts likes, shares, and saves as "clicks" within attribution windows, meaning conversions from users who never left Meta's platforms get credited to advertising campaigns based on engagement actions alone.

Internal company documents viewed by Reuters revealed on November 6, 2025, that Meta internally projected earning approximately 10% of its 2024 annual revenue—roughly $16 billion—from advertisements promoting scams and banned goods. According to the documents, Meta's platforms expose users to an estimated 15 billion "higher risk" scam advertisements daily.

Brand safety deteriorates alongside automation acceleration

Meta's January 2025 decision to discontinue internal fact-checking in favor of "Community Notes" increases likelihood of ads appearing near controversial content. The simultaneous push for AI chat integration—where every conversation through Facebook, Instagram, Messenger, and Ray-Ban glasses feeds advertising algorithms starting December 16, 2025, with no opt-out—compounds concerns.

According to University of Washington linguist Emily Bender, this creates "financial incentives for Meta to keep people chatting with chatbots—to optimize on engagement, which is one of the vectors for harm."

CEO Mark Zuckerberg articulated the automation endgame in a May 2025 Stratechery interview: complete automation where businesses "connect to your bank account, you don't need any creative, you don't need any targeting demographic, you don't need any measurement, except to be able to read the results that we spit out."

Anonymous agency executives told The Verge that "no clients will trust what they spit out as they are basically checking their own homework," characterizing Meta's attitude as ranging "from moderate condescension to active antagonism to 'we'll fucking kill you.'" CEO Matthias Schrader of OH-SO Digital called the vision "brutal," while Friedrich Dromm predicted "as early as 2028, there will no longer be classic advertising agencies as we know them."

Platform performance data contradicts advertiser concerns

Meta reported second quarter 2025 advertising revenue of $46.6 billion on July 30, representing a 22% increase compared to the same period last year. According to CFO Susan Li, the company's new AI-powered recommendation model for ads expanded to new surfaces and improved performance significantly during the quarter.

"This quarter, we expanded our new AI-powered recommendation model for ads to new surfaces and improved its performance by using more signals and a longer context. It's driven roughly 5% more ad conversions on Instagram and 3% on Facebook," Zuckerberg explained during the earnings call.

The company's Generative Ads Recommendation System, which powers the ranking stage of Meta's advertising system, received substantial improvements. Li detailed the technical advances: "In Q2, we improved the performance of GEM by further scaling our training capacity and adding organic and ads engagement data on Instagram. These improvements translated to meaningful performance gains for advertisers."

Meta's third quarter results showed advertising revenue reaching $50.1 billion, up 26% year-over-year. Ad impressions increased 14% while the average price per ad rose 10%. Andromeda delivered a 14% increase in ads quality on Facebook surfaces during the quarter.

Advertiser response strategies emerge

Van der Hallen's LinkedIn post recommended that marketing professionals avoid converting setups overnight despite automation promises. "Your testing roadmap can absolutely include 'Andromeda' setups, just don't let yourself be influenced too heavily by the current hype," he wrote.

Meta's official response maintains advertisers can set guardrails for age, location, and budgets while benefiting from optimization. The hybrid approach recommendation—60-70% Advantage+, 30-40% manual—acknowledges value in maintaining human oversight. Performance 5 framework emphasizes measurement including A/B testing, conversion lift studies, and marketing mix modeling to validate effectiveness.

Meta introduced restricted word controls on August 20, 2025, addressing advertiser concerns about AI-generated text options containing inappropriate content for specific brands or industries. Advertisers can now input specific words or phrases they want excluded from automated text generation, providing brand safety controls for campaigns using Meta's artificial intelligence creative tools.

Industry experts converge on several essential practices for advertisers using AI automation while protecting brand interests. Adopt gradually rather than wholesale migration. Test one Advantage+ campaign against manual setup to understand system behavior before scaling. Maintain detailed performance tracking across both automated and manual approaches to identify when automation delivers genuine improvements versus capturing existing demand.

The Andromeda announcement outlined future development plans. According to Meta, the model architecture will transition to support an autoregressive loss function, leading to more efficient and faster inferencing that delivers more diverse ad candidates. Increased ad diversity can improve user experience with ads and drive better advertiser outcomes.

Integration with Meta Training and Inference Accelerator and future commercially-available GPUs will continue pushing boundaries of scaling retrieval. Meta estimates another 1,000x increase in model complexity through this hardware evolution.

Yet these technical promises arrive as DoubleVerify's 2025 Global Insights Report revealed 65% of marketers expressing brand suitability concerns about advertising adjacent to AI-generated content. The research surveyed 1,970 marketing and advertising decision-makers worldwide, examining how brand safety measurement capabilities must adapt to real-time content evaluation before ad serving occurs.

Timeline

• December 2, 2024: Meta announces Andromeda retrieval system built with Nvidia Grace Hopper Superchips
• February 14, 2024: Valentine's Day budget catastrophe with 10x CPM inflation
• April 23, 2024: Major incident depletes daily budgets in hours across thousands of accounts
• January 21, 2025: Meta removes detailed targeting exclusions
• May 3, 2025: Zuckerberg announces vision to replace creative agencies with AI
• June 16, 2025: WhatsApp introduces ads using Instagram and Facebook data
• July 30, 2025: Meta reports 22% advertising revenue growth to $46.6 billion
• August 20, 2025: Former employee alleges artificial ROAS inflation for Shops ads
• August 20, 2025: Meta introduces restricted word controls for AI-generated text
• September-October 2025: Meta deprecates legacy campaign APIs for Advantage+ structure
• October 13, 2025: Meta announces attribution window restrictions effective January 2026
• October 16, 2025: Meta introduces default option to spend up to 5% on excluded placements
• October 26, 2025: Industry analysis examines brand control versus algorithmic efficiency tension
• October 30, 2025: Meta reports 26% revenue growth with Andromeda delivering 14% ads quality increase
• November 6, 2025: Reuters reports Meta earned billions from scam ads
• November 27, 2025: Digital marketing specialist Bram Van der Hallen challenges Andromeda consolidation approach on LinkedIn
• Q1 2026: Full migration to unified Advantage+ campaign structure required

Summary

Who: Meta's advertising platform engineers announced Andromeda while digital marketing professionals including Bram Van der Hallen expressed skepticism about consolidated campaign approaches that reduce advertiser control over targeting, creative, and placement decisions.

What: Andromeda is a machine learning retrieval engine built with Nvidia Grace Hopper Superchips that achieved 6% recall improvement and 8% ads quality improvement on selected segments. The system employs hierarchical indexing, deep neural networks with 10,000x increased model capacity, and model elasticity that adjusts complexity in real-time. However, advertising professionals question whether reported performance improvements represent genuine incremental value or simply capture existing demand through retargeting audiences, particularly following allegations of ROAS inflation and attribution methodology concerns.

When: Meta announced Andromeda on December 2, 2024, with the system already deployed across Instagram and Facebook applications. Van der Hallen published his critique on November 27, 2025. The announcement arrived as Meta accelerated automation throughout 2025, removing detailed targeting exclusions in January, deprecating legacy APIs in October, and planning full Advantage+ migration by Q1 2026.

Where: The Andromeda system operates within Meta's advertising infrastructure across Facebook and Instagram applications. The broader automation concerns affect advertisers globally managing campaigns through Meta Ads Manager, with placement implications extending to Messenger and Audience Network properties.

Why: Meta developed Andromeda to address scalability challenges presented by exponential growth of ad creatives through Advantage+ automation and generative AI tools. More than one million advertisers created over 15 million ads in a month using Meta's AI tools. The technical architecture handles processing three orders of magnitude more ads than subsequent recommendation stages while maintaining strict latency constraints. Industry skepticism stems from multiple sources: attribution concerns showing Advantage+ may generate only 17% of claimed conversions, lack of transparency in algorithmic decision-making, brand safety deterioration following fact-checking discontinuation, whistleblower allegations of artificially inflated return on ad spend metrics, internal documents revealing billions earned from scam advertisements, and fundamental tensions between automation efficiency and advertiser control over strategic campaign parameters that determine brand identity and safety.