AI models now pick database winners. Neon is recommended first 70% of the time.

A new analysis shows AI models recommend serverless newcomers like Neon and Upstash over incumbents. The playbook is winning the discourse, not just the feature set.

In a test across four major AI models, serverless Postgres provider Neon was recommended first in 14 of 20 prompts for database solutions. This result, from a systematic query of ChatGPT, Claude, Gemini, and Perplexity, reveals a new discovery battleground for developer tools. The winners are not the established incumbents, but the brands most present in the models' training data.

The serverless challengers are the new default

An analysis posted by Bersyn, a market intelligence service, tested four LLMs with five common buyer questions for database and storage solutions. The results showed a clear preference for modern, serverless-focused providers over legacy names.

Neon, a serverless Postgres company, was the top recommendation in 70% of conversations (14 of 20). Upstash, for serverless Redis, was named first in 55% of tests (11 of 20). Turso, an edge-native SQLite database, was the first choice in 45% of queries (9 of 20).

These findings demonstrate that a company founded years after its primary incumbent can become the default answer for AI-driven developer discovery.

Incumbents still own specialized queries

The dominance of newcomers was not absolute. When queries shifted from general-purpose databases to specialized tasks, the models reverted to more established or category-specific leaders.

For vector search, models preferred Pinecone, Milvus, and Weaviate. Challenger Qdrant was recommended first only six times. For object storage, the models consistently named Amazon S3, Cloudflare R2, and Backblaze. A newer alternative like Tigris was never mentioned.

The pattern held for other sub-categories. The models recommended ClickHouse for real-time analytics over Tinybird, and Prisma for ORMs over Drizzle. Specificity appears to trigger a more conservative, incumbent-focused response from the models.

The analysis: discourse beats features

The author of the analysis posits that these outcomes are not based on feature-for-feature product evaluations by the AI. Instead, the models are reflecting the discourse present in their training data.

Neon and Upstash won because a critical mass of independent blog posts, tutorials, and forum discussions had already named them as the answer before the models were trained. Product quality is secondary to presence in the corpus. The buyer who types 'best vector database' and takes the first answer never sees Qdrant, no matter how good it is, until the inputs change.

What We'd Change

This playbook is a lagging indicator. The content and community discussions that placed Neon in the training data were created months or years ago. To win the recommendation of a 2027 model, a company must be building its public knowledge footprint now. This is a long-term content and community strategy, not a short-term hack. The work must precede the model's training cut-off.

The channel is vulnerable to manipulation. As AI recommendations become a primary discovery channel, it will attract SEO-like optimization. Expect a rise in low-quality content designed to flood the training corpus with mentions of a specific tool. This could degrade the channel's utility over time, making the current advantage for authentic community winners a temporary one.

Focus on the specialized query moat. The data shows that general queries are won by broad mindshare, but specialized queries are won by deep, technical authority. For a bootstrapped or early-stage tool, competing with Neon on general "best database" discourse is inefficient. A more durable strategy is to own the answer to a specific, high-intent problem, becoming the default recommendation for a niche where incumbents are too broad to compete effectively.

Landing

The rise of LLMs as a developer's first stop for tool selection redefines GTM strategy. Marketing is no longer just about reaching a human developer through ads or content. It is now also about educating the AI assistants that developers consult. The most effective strategy is to become an indispensable part of the public technical record. This means creating the documentation, tutorials, and case studies that are so clear and authoritative that they become the foundational text for a model's understanding of a problem space.

The investor read

This analysis signals a shift in go-to-market strategy for developer tools, establishing the 'AI recommendation engine' as a new, zero-cost distribution channel. Companies like Neon and Upstash have captured it, likely through strong, long-term content and community engagement that pre-dated the current LLM cycle. For investors, a key diligence question for new devtools becomes: 'What is your strategy for being the default AI answer?' A product without a significant footprint in public technical discourse may face a permanent discovery disadvantage. This represents a new kind of moat based on corpus penetration, not just technology or sales.

Pull quote: “The buyer who types 'best vector database' and takes the first answer never sees Qdrant, no matter how good it is, until the inputs change.”