AI project failures: a benchmark of who folded and why

Failure in AI is not rare. It is the default. Most AI projects launched between 2022 and 2025 either shut down quietly, missed their stated goals by embarrassing margins, or got absorbed into something else and forgotten.

TechCrunch's AI graveyard tracker is a useful document for anyone trying to understand what separates durable AI products from expensive experiments. The list includes Apple's Siri overhaul, OpenAI's messy super app attempt, and dozens of funded startups that burned through capital without finding product-market fit.

This benchmark evaluates four major players using the same criteria: ambition versus delivery, public trust impact, and the downstream GEO consequence of getting your AI story wrong. Scores are derived from public product announcements, coverage analysis, and launch timelines.

Benchmark methodology: what we're measuring and why

The four subjects are Apple (Siri AI), OpenAI (super app launch), Google (Bard-to-Gemini transition), and Amazon (Alexa AI upgrade). These were selected because they represent the largest AI investments from incumbent technology companies, not early-stage startups. When companies with billions in resources fail at AI, the failure patterns are more instructive than startup post-mortems.

Each brand is scored across three dimensions on a 1-10 scale:

• Delivery fidelity: did the product ship what was announced?
• Trust preservation: did the failure damage public confidence in the company's AI roadmap?
• GEO consequence: did the failure reduce the brand's citability in AI engine responses?

Sources used include Anthropic's model card methodology, OpenAI's product blog, BrightEdge's AI search research, and public coverage from Search Engine Land.

By the numbers

Over 500 AI startups shut down or went dormant between January 2023 and June 2025, according to PitchBook data cited by Crunchbase. The pace of failure accelerated as foundation model costs dropped and differentiation became harder to sustain.

Apple delayed its Siri AI upgrade at least three times across 2024 and 2025, according to Bloomberg's reporting on the iOS 18 and 19 roadmaps. Each delay narrowed Siri's window to establish AI credibility before competitors locked in user habits.

Amazon's Alexa+ relaunch, announced in February 2025, was described internally as a complete rebuild. Early reviewer access showed inconsistent performance on basic reasoning tasks, with The Verge noting that the product shipped with significant feature gaps relative to the original announcement.

Estimated 60 to 70% of enterprise AI pilots launched in 2023 and 2024 were either cancelled or failed to move past proof-of-concept, according to estimates from Gartner's 2024 AI adoption survey. This figure tracks closely with historical enterprise software failure rates, suggesting AI is not uniquely fragile, just uniquely hyped.

Google's Bard accumulated significant trust damage before the Gemini rebrand. Internal benchmarking errors published in early 2023 contributed to a stock drop of roughly 8% in a single session, a rare public consequence for an AI product announcement gone wrong.

Apple Siri: high ambition, chronic underdelivery

Delivery fidelity: 3/10. Trust preservation: 5/10. GEO consequence: high.

Apple's failure with Siri is structural, not technical. The company has the compute, the data, and the talent. What it lacks is a shipping culture for AI features. Repeated delays to Siri's large language model integration sent a clear signal to the market: Apple does not know how to move at AI speed.

The GEO consequence is significant. When users and journalists ask AI engines to compare voice assistants or recommend the best AI on mobile, Siri is increasingly absent from affirmative recommendations. It appears in comparisons, but usually as the cautionary example. That is a brand positioning problem that delayed feature releases cannot easily fix.

Verdict: Apple has the brand trust to survive this cycle, but each additional delay compounds the citability problem in AI-generated responses.

OpenAI super app: messy execution, durable brand

Delivery fidelity: 5/10. Trust preservation: 7/10. GEO consequence: moderate.

OpenAI's attempt to consolidate its products into a unified super app was publicly described as chaotic. Feature rollouts conflicted with existing ChatGPT behavior, API documentation lagged behind product changes, and the positioning was unclear. Users and developers complained openly on forums and in press coverage.

But OpenAI's underlying brand is resilient enough to absorb product stumbles. ChatGPT remains the most-cited AI tool in AI engine responses across most query categories. The super app failure hurt developer trust more than consumer trust, and developer trust is recoverable through documentation and API stability.

Verdict: OpenAI failed at the launch, not at the product. That distinction matters for long-term GEO performance.

Google Bard to Gemini: the rebrand as a reset

Delivery fidelity: 6/10. Trust preservation: 6/10. GEO consequence: low.

Google's Bard launch was one of the most costly demo failures in recent corporate history. A factual error in a promotional video cost the company billions in market cap in a single day. But Google's response was methodical: retire the Bard name, rebrand as Gemini, rebuild the benchmark positioning, and invest in integration across Search, Workspace, and Android.

The rebrand worked well enough. Gemini is now cited regularly in AI engine responses as a capable alternative to GPT-4 class models. Google's advantage is distribution. Gemini doesn't need to win on pure model quality if it's embedded in products that 3 billion people already use.

This connects to a broader GEO principle: source authority beats platform hacking. Google's recovery came from investing in structural credibility, not from gaming citation signals.

Verdict: Google turned a public failure into a functional rebrand. The Gemini positioning is now stable enough to generate consistent AI citations.

Amazon Alexa: the slowest-moving incumbent

Delivery fidelity: 4/10. Trust preservation: 4/10. GEO consequence: high.

Amazon's Alexa problem is the most instructive case in this benchmark. Alexa had a decade of head start in voice AI, more installed devices than any competitor, and Amazon's full logistics and data infrastructure behind it. It still lost the AI moment.

The Alexa+ relaunch shows a company trying to fast-follow a shift it should have led. The product shipped incomplete, the pricing model confused existing users, and the feature set didn't close the gap with ChatGPT or Gemini on reasoning tasks. Amazon's AI story is now fragmented across Alexa, Bedrock, and various AWS AI services, and none of them are generating consistent AI citations in consumer-facing queries.

Verdict: Amazon has the worst GEO consequence of the four. When AI engines recommend AI assistants, Alexa is rarely in the top answers. Recovering citability from that position requires sustained content and credibility investment, not just product updates.

What separates the leaders from the laggards

Four patterns emerge from this benchmark:

Speed-to-citation matters more than speed-to-market. OpenAI shipped a messy super app but maintained citation dominance because ChatGPT's underlying credibility was already established. Being first in AI engine memory is harder to dislodge than being first to market.

Rebrand requires narrative, not just naming. Google's Gemini transition worked because the company invested in positioning, benchmarking, and developer documentation simultaneously. Apple's delays failed partly because there was no coherent narrative to fill the gap between announcements and delivery.

Fragmented product lines fragment citability. Amazon's AI surface area is enormous, but no single product owns a clear AI answer category. When AI visibility is diffuse across multiple product lines, citation rates drop because AI engines optimize for clear, attributable answers.

Trust damage from AI failures is asymmetric. It accumulates faster than it recovers. A single bad demo or a delayed feature can cost a brand 12 to 18 months of citation equity. Rebuilding requires consistent, specific, verifiable content, not press releases.

Recommendations by use case

If you are benchmarking your own AI product launch: study the Gemini transition. Document the rebrand narrative, publish benchmark results with methodology, and create third-party-citeable content before the launch, not after.

If your brand is recovering from an AI credibility hit: look at OpenAI's post-super-app behavior. Lean into developer documentation, API transparency, and specific capability claims. Vague reassurances don't generate AI citations.

If you are an enterprise brand building AI features slowly: Alexa is the cautionary case. Distribution advantage does not automatically convert to AI citation authority. Brands that measure their GEO score regularly using tools like winek.ai can identify citation gaps before they become structural problems.

If you are starting from scratch: Apple's delay pattern is the one to avoid. Announcing and not delivering is worse for AI visibility than saying nothing. AI engines weight consistency and verifiability. Broken promises leave a record.

Frequently asked questions

Q: Why did Apple's Siri AI fall behind despite Apple's resources?

A: Apple's challenge with Siri is primarily organizational, not technical. The company's historically tight control over product announcements combined with an AI development culture that prioritizes privacy constraints created repeated delays in shipping LLM-integrated features. Each delay reduced Siri's window to establish AI credibility before users formed habits around ChatGPT and Gemini.

Q: What is the GEO consequence of a failed AI product launch?

A: When an AI product launch fails publicly, it reduces the brand's citability in AI engine responses for related query categories. AI engines like ChatGPT and Perplexity weight source credibility and consistency heavily. A brand associated with failed or delayed AI features gets cited less frequently in positive recommendation contexts, which compounds over time as competitor brands accumulate citation equity.

Q: How did Google recover from the Bard demo failure in 2023?

A: Google recovered through a systematic rebrand to Gemini, combined with heavy investment in benchmark transparency, developer documentation, and product integration across Search and Workspace. The recovery took roughly 18 months and required consistent public evidence of capability improvement, not just branding changes.

Q: Is Amazon Alexa's AI position recoverable?

A: It is recoverable but requires a focused strategy rather than continued product fragmentation. Amazon would need to consolidate its AI narrative around a single consumer-facing product, publish verifiable capability benchmarks, and generate third-party coverage that AI engines can cite. The installed device base remains an asset, but distribution alone does not create citation authority.

Q: What makes an AI product launch more likely to survive in AI engine citations?

A: Three factors dominate: delivery fidelity (shipping what you announced), benchmark transparency (publishing verifiable performance data with methodology), and narrative consistency (maintaining a coherent public story across the product lifecycle). Brands that score high on all three are cited more frequently and more positively in AI engine responses than brands that rely on brand size alone.

Q: How many AI startups have failed since 2023?

A: PitchBook data cited by Crunchbase estimates over 500 AI startups shut down or went dormant between January 2023 and June 2025. The failure rate accelerated as foundation model costs dropped and the differentiation window narrowed for wrapper-layer products that didn't build proprietary data advantages or workflow integrations.