How does a database company get chosen when AI helps pick the stack?

Who chooses a database, and what is a customer worth?

Developers usually make the first choice; architects and procurement turn it into an enterprise standard and a contract.

A database rarely starts with a purchase order. A developer picks one for a prototype, the prototype goes into production, and the choice hardens as data piles up. Moving data later is expensive and risky, so the first pick tends to stay. That is why database companies spend so much effort on free tiers, documentation and developer communities. DevOps tools grow the same way, as our guide to how DevOps tools get recommended to engineers shows.

The enterprise stage looks different. An architecture or platform team reviews the database for security, compliance, support, cost at scale and fit with the rest of the stack. Procurement then signs a multi-year subscription or a cloud commitment. Security tools sold to developers face a similar double review, covered in how application security vendors win demand.

Public filings show what that second stage is worth. MongoDB reported revenue of $771.8 million for its second quarter of fiscal 2027, up 30%, from more than 70,600 customers. Its remaining performance obligations, contracted revenue not yet recognized, reached $1,519.2 million, up 91% year on year. Its count of customers paying $100,000 or more a year has climbed every quarter in the table it publishes.

Private companies show what the developer stage is worth. Supabase raised $500 million at a $10 billion pre-money valuation in June 2026, with nearly 10 million developers building on it. Large data platforms are buying their way into the same market: Databricks agreed to acquire Neon, and Snowflake announced the acquisition of Crunchy Data (opens in a new tab), whose press release put PostgreSQL use at 49% of all developers and called it “a massive $350 billion market opportunity.” That figure is Snowflake’s own estimate.

How has AI changed the way databases get picked?

AI coding tools now create many new databases, so an agent’s default choice has become a sales channel.

The clearest evidence comes from the vendors themselves:

• Neon. Databricks said internal telemetry showed over 80% of databases provisioned on Neon were created by AI agents rather than humans.
• Supabase. Its CEO wrote in June 2026 that more than 60% of new databases are launched by an AI tool, and that growth accelerated as Claude Code and Codex expanded the number of people who can build.
• Integrations. Supabase became an official ChatGPT app (opens in a new tab) with 29 tools for running queries and managing projects. MongoDB launched a hosted service that connects coding agents such as Claude Code and Codex to live Atlas data.

These are company-reported figures from firms with an interest in the trend, and they describe developer-led products more than enterprise estates. Still, the direction is plain: when a person asks an AI tool to build an app, the tool usually picks the database too. The same goes for other parts of the stack, as our guide to how developer tools win AI-first users explains.

Developers themselves use AI heavily but cautiously. In Stack Overflow’s 2025 survey, 84% used or planned to use AI tools, up from 76% in 2024. Only 31% used AI agents at the time, and 35% visited Stack Overflow after running into problems with AI answers. Human communities remain where developers go to check.

Which questions do developers and architects ask AI about databases?

Workload, comparison, compatibility, cost and migration questions. We wrote the sample prompts below to show how developers and architects tend to phrase database questions; none were taken from real query logs.

Two kinds of answer are at stake. A developer’s question often ends in a setup command, so being named means being installed. An architect’s question ends in a shortlist for a proof of concept. Pricing answers are fragile for usage-priced products: in our pricing study, only 61.9% of plan prices quoted by four assistants for 45 software products were fully faithful to the vendor’s page.

How does an agent’s database pick turn into usage revenue and contracts?

Through a long funnel: an AI tool’s default becomes a free project, then production use, then a contract.

1. A developer or a coding agent needs a database and asks for one, or the agent picks one itself.
2. The tool names or installs a database, usually one it has seen in documentation, templates and examples.
3. The project grows on a free or usage-priced tier.
4. If the app succeeds, usage and the bill grow with it.
5. The company standardizes, and procurement signs a larger commitment with support, security and compliance terms.

MongoDB’s customer table shows the shape of the later steps: more than 70,600 customers, of which 2,999 pay $100,000 or more a year. AI visibility, on this reading, matters most at steps 1 and 2, where volume is decided, and again when architects research alternatives for a large workload. That last point is our inference, not a measured link. The clouds those workloads run on face their own shortlist, covered in how cloud platforms win enterprise buyers.

The revenue rarely arrives with a referral tag. A database chosen by an agent in March may show up as an enterprise deal a year later. We cover the attribution problem in what lost clicks mean for pipeline.

What decides whether an AI assistant recommends a database?

No platform documents how it picks databases; studies and developer behavior point to public, checkable sources.

Documented by the platforms. Google says its AI features may issue several related searches for one question, and OpenAI says ChatGPT search rewrites a question into targeted queries. Neither says how a database gets recommended. Coding agents have no published selection rules either.

Observed in studies. Software is where assistants agree most. In our consistency study, B2B software had the most stable brand lists across repeated runs, with a mean overlap of 0.708. Stable lists are good news for established names and hard for newcomers to break into. Developer communities also matter as sources: in our Reddit study, 35.4% of Google’s AI Overviews for B2B software and technology searches cited at least one Reddit thread.

What developers trust. Developers check what AI tells them. Stack Overflow found 46% distrusted the accuracy of AI output, up from 31% a year earlier. Documentation, community threads and benchmarks they can rerun carry more weight than claims.

Our inference. An assistant or agent can only recommend a database it has seen used. Clear setup guides, framework templates, code examples, honest comparison pages and independent benchmarks are the raw material it works from. A reasonable expectation is that databases with that material in public, current and consistent places get named more often. No study has yet tested this for databases. Training data also lags: an assistant answering from memory may describe last year’s product, as we explain in why ChatGPT misses new products.

What does it cost a database company to be missing?

Mostly lost defaults, which compound, though no study has put a number on the loss.

• Defaults compound. If more than 60% of new databases on a platform are launched by AI tools, a database that agents do not reach for misses a growing share of new projects at the moment of choice, we infer.
• Switching is rare. Once data and code depend on a database, moving is costly. A project lost at the prototype stage is often lost for its lifetime, our reading of how database adoption works.
• Popularity is self-reinforcing. PostgreSQL topped Stack Overflow’s “want to use” and “want to keep using” lists for three years running. When assistants and agents favor what is common, the common choice gets more common.
• Wrong answers cost credibility. A developer who gets a wrong limit, price or feature from an AI answer and then hits it in production blames the product, not the assistant.

How does GEO work for a database company?

It makes your database easy for assistants and agents to understand, set up and compare honestly. No one can promise that an agent will install it.

1. One clear identity per workload. Say plainly what your database is best for, and what it is not for, in the same words across docs, your site, cloud marketplace listings and community answers.
2. Agent-ready documentation. Keep quick-starts, connection strings, limits and pricing on readable pages; publish framework templates and integrations where coding tools look. Docs that only render with scripts are a common blind spot, as when AI agents can’t read your site explains.
3. Benchmarks others can rerun. Publish methods and scripts, and encourage independent tests. Developers trust results they can reproduce, and independent write-ups give assistants a source other than you.
4. Honest comparisons and migration guides. Explain trade-offs against the databases buyers already use. Our research summary on whether comparison pages help B2B citations covers what such pages can and cannot do.
5. Real community presence. Answer questions on developer forums and Reddit as identified staff, and fix the misunderstandings you find. Why those threads matter is covered in how Reddit shapes Google’s AI Overviews.
6. Measurement in assistants and agents. Track which databases ChatGPT, Gemini, Claude, Perplexity, Copilot and Google’s AI features name for your workloads, and what coding agents install when asked to build typical apps. For sample sizes, read how many prompts to track.

Which questions about databases and coding agents has nobody answered yet?

No public study shows how coding agents pick a database, or whether AI visibility moves enterprise database revenue.

• Agent figures come from vendors. The Neon and Supabase numbers are company-reported and cover their own platforms.
• No study of agent selection. We found no published test of which databases coding agents pick, or why.
• We could not use one standard source. DB-Engines publishes a widely watched popularity ranking, but we could not save its page for verification, so we do not quote its scores.
• The link to subscription and consumption revenue is thin. The evidence is reviewed in does AI visibility drive business results and the B2B SaaS view of AI search revenue.

How can a database company find out what agents install and assistants recommend?

Have assistants and coding agents build the apps your customers build, and log which database each one picks.

Pick a dozen typical projects and architecture questions for your strongest workloads. Give them to ChatGPT, Gemini, Claude and Copilot and to two or three coding agents, repeating each run several times. Note which database is named or installed, which sources are cited, and whether your limits, prices and features are described correctly. The gaps usually point to missing templates, unreadable docs or a lack of independent benchmarks.

We can run those tests with you: ask us to audit what agents and assistants say about your database. We will show where AI answers and coding agents place you on the questions that lead to new projects and enterprise evaluations, and which gaps most likely cost you developer adoption and paid usage. Agent-ready documentation, rerunnable benchmarks and honest migration guides make up most of our generative engine optimization service for a database company, and that page explains how each is run.

Frequently asked questions

Do AI coding agents really choose the database?

Often. Neon reported over 80% of its databases were created by AI agents, and Supabase more than 60% of new ones by AI tools.

Should we be Postgres-compatible to be recommended?

Not necessarily, but PostgreSQL is the common default: 66% of developers who used it wanted to keep using it in Stack Overflow’s survey.

Do benchmarks help AI visibility?

Untested, but independent, reproducible benchmarks give assistants a source beyond your own site, and developers distrust unverified claims.

Will enterprise architects trust an AI recommendation?

Not on its own. 46% of developers distrust AI output, so expect architects to verify every claim in a proof of concept.

How do we know what agents install today?

Test it. Ask several coding agents to build typical apps and record the database each one sets up, repeating runs over time.

Sources

• Supabase, Paul Copplestone (2026-06-04), Supabase Series F (opens in a new tab)
• Supabase, Greg Richardson (2026-05-08), Supabase Is Now an Official ChatGPT App (opens in a new tab)
• Databricks (2025-05-14), Databricks Agrees to Acquire Neon to Deliver Serverless Postgres for Developers + AI Agents (opens in a new tab)
• Snowflake (2025-06-02), Snowflake Acquires Crunchy Data to Bring Enterprise Ready Postgres Offering to the AI Data Cloud (opens in a new tab)
• MongoDB (2026-09-01), MongoDB, Inc. Announces Second Quarter Fiscal 2027 Financial Results
• Stack Overflow (2025-07-29), Stack Overflow’s 2025 Developer Survey (opens in a new tab)
• Google Search Central (2025), AI features and your website (opens in a new tab)
• OpenAI Help Center (2025), ChatGPT search (opens in a new tab)
• Underneath (2026), How faithfully do AI assistants quote software prices?
• Underneath (2026), Ask an AI the same question 5 times: do the brands change?
• Underneath (2026), When does a Reddit thread become evidence in Google’s AI?