--- title: "How developer tool companies win users when developers ask AI" description: "By being the tool AI coding agents and assistants pick for a developer’s stack, then turning that free signup into team plans and enterprise contracts." canonical: "https://underneath.agency/resources/developer-tools-ai-search" published: 2026-10-07 updated: 2026-10-08 publisher: "Underneath (https://underneath.agency/agent)" entity: "https://underneath.agency/.well-known/entity.json" --- Guide · AI search # How do developer tool companies win users when developers ask AI first? By becoming the tool that AI assistants and coding agents choose when a developer asks “what should I use?”, then turning that free signup into a paid team and, later, an enterprise contract. For developer tools the AI answer is often not a recommendation a person reads. It is a package an agent installs. Being picked, or skipped, now happens inside the editor, before any developer visits your website. ## The short version 1. In the 2026 [JetBrains Developer Ecosystem research](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/), 90% of professional developers used AI coding agents at work at least weekly and 68% used them daily. 2. In [Amplifying’s test](https://amplifying.ai/research/claude-code-picks/report) of 2,430 open-ended prompts such as “i need a database, what should i use,” Claude Code chose GitHub Actions for CI/CD in 94% of answers and Stripe for payments in 91%, while traditional cloud providers received zero primary picks for deployment. 3. Agents now read documentation more than people do. Across docs sites hosted by [Mintlify](https://mintlify.com/data.md/), agents made 61.87% of requests in August 2026, and agent readership grew 7.7x from February to August while human readership grew 1.2x. 4. Agents are already creating accounts. [Databricks](https://www.databricks.com/company/newsroom/press-releases/databricks-agrees-acquire-neon-help-developers-deliver-ai-systems) said over 80 percent of the databases provisioned on Neon were created automatically by AI agents rather than by humans. 5. Signups already carry the trace of AI answers: the developer platform [Vercel](https://vercel.com/blog/how-were-adapting-seo-for-llms-and-ai-search) says ChatGPT refers around 10% of new signups to it, up from 1% six months earlier. ## Who buys developer tools, and what is a new user worth? Developers choose them, usually for free, and the company pays later as usage spreads across teams. The buying journey runs bottom-up. In the [2025 Stack Overflow Developer Survey](https://survey.stackoverflow.co/2025/work), 48% of developers said they endorsed or influenced the purchase of new technology in their organization in the past year, and a fifth of those influenced a substantial addition to the company’s tech stack. Asked what makes them endorse a tool, developers ranked an easy-to-use API first, a robust and complete API second and a reputation for quality third. “AI integration or AI Agent capabilities” came ninth of ten. The audience is huge and growing fast. GitHub’s [Octoverse 2025 report](https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/) counts more than 180 million developers on the platform, with more than 36 million joining in a single year. Most of those people will never sign a contract themselves. Their choices, repeated across a company, decide which tool the company ends up paying for. What a new user is worth depends on how far that spread goes. Free-to-paid conversion is thin: the research firm [Sacra estimates](https://sacra.com/research/supabase-170m-year-growing-221-yoy/) that Supabase’s paying customers are about 2.5% of its registered users, at roughly $700 a year each. The value sits in expansion. [GitLab’s fiscal 2026 results](https://s204.q4cdn.com/984476563/files/doc_financials/2026/q4/Gitlab-4Q26-Earnings-Press-Release.pdf) report 1,456 customers paying more than $100,000 a year, up 18% year over year, and a dollar-based net retention rate of 118%: existing customers spent 18% more than a year before. A single developer’s first install can be the start of a six-figure account, which is why the first pick matters so much. ## Where do AI assistants already sit in the developer’s workflow? Inside the editor and the terminal, where agents now write code, choose packages and read documentation. Use is close to universal. In the Stack Overflow survey, 84% of respondents used or planned to use AI tools in their development process, and 51% of professional developers used them daily. JetBrains reports that around 39% of professional developers worldwide used Claude Code at work in May to July 2026, up from 18% in January, and 47% in the United States. OpenAI’s Codex grew from 3% to 16% adoption over the same period. The makers of those coding tools run their own version of this race, covered in [how AI coding tools win developers](https://underneath.agency/resources/ai-coding-tools-developers-ai-search). These are not search engines in the usual sense, but they answer the same question a search used to: which tool should I use for this? The old discovery channels still matter, but they are shifting. Technical documentation remained the top resource for learning to code, used by 68% of respondents in the Stack Overflow survey, and Stack Overflow itself was the most used community platform at 84%. Yet the number of new questions posted there has fallen to levels [not seen since 2009](https://blog.pragmaticengineer.com/are-llms-making-stackoverflow-irrelevant/), according to data reported by The Pragmatic Engineer. Developers who once asked a forum now ask an assistant. The reader of your documentation has changed too. Mintlify, a documentation platform, reports that 70% of the docs sites it hosts serve more machine requests than human page loads. For the “dev infrastructure and data” sites in its August 2026 leaderboard, the agent share was 37.9%, up from 15.5% in February. These are a vendor’s own measurements of its own customers, but the direction is consistent across every industry group it reports. Developers do not trust the output blindly. In the Stack Overflow survey, 46% of developers distrusted the accuracy of AI tools against 33% who trusted it, and 35% said they visit Stack Overflow after running into issues with AI responses. A reasonable expectation is that developers accept an agent’s first pick for low-stakes choices and check it for anything that touches production, security or cost. ## Which questions lead developers to a tool? Two kinds: questions a developer asks a chat assistant, and requests that leave the choice to a coding agent. The second kind is new and specific to this industry. Amplifying’s benchmark used real, open-ended prompts with no tool named, such as “how do i deploy this?”, “add user authentication” and “what testing framework works best with this stack.” The agent did not just suggest a tool. Amplifying recorded what it installed, configured and committed. API vendors meet the same test when an agent writes the integration; see [how API companies win agent-chosen integrations](https://underneath.agency/resources/api-companies-ai-search). The first kind looks more like classic research. We wrote the chat prompts below as examples of how a developer frames a tooling question; they are not logged queries: - Category within a stack: “Best feature flag service for a Next.js app on Vercel.” - Alternatives: “Open-source alternatives to Datadog for a small team.” - Comparison: “Supabase vs Firebase for a SaaS app with row-level security.” - Price and limits: “Which error-tracking tools have a free tier above 10,000 events a month?” - Fit and risk: “Which CI services support self-hosted runners and SOC 2 reports?” Questions that name a framework or hosting stack decide which tools enter the running. Comparison, pricing and compliance questions decide who gets past a team lead. For broader B2B software buyer questions, see [our article on SaaS shortlists](https://underneath.agency/resources/b2b-saas-revenue-from-ai-search). ## How does an AI pick turn into revenue for a developer tool? Through usage: the agent installs you, the project grows on you, the team pays, and the company expands. **The pick.** When an agent picks a tool, the tool ships with the code. Databricks put it plainly when it bought Neon: “four out of every five databases on their platform are spun up by code, not humans.” Sacra estimates that AI coding agents now create more than 60% of new Supabase databases and that database launches there are up 600% year over year, calling coding agents its fastest-growing driver of usage. These are a buyer’s statement and an analyst’s estimates, but both point the same way. Database vendors can go deeper with [how a database gets chosen for the stack](https://underneath.agency/resources/database-companies-developers-ai-search). **The signup.** Some of the path is visible as referrals. Vercel saw ChatGPT’s share of its new signups reach around 10%, up from 4.8% the previous month. Much of the path is not: an agent that installs a package or provisions a database leaves no referral in analytics, and we infer it shows up instead as unexplained growth in free accounts. **The team plan and the contract.** A project that grows on a free tier eventually hits limits, needs collaboration or needs a security review. GitLab’s 118% net retention shows how much of a developer tool’s revenue comes from accounts that grow after the first sale. The AI pick is worth far more than the first plan it produces, because it decides which tool is already inside the codebase when the company starts paying. ## What decides whether an AI assistant or coding agent picks your tool? Platforms say little; studies show that the project’s existing stack and the tool’s established position drive most picks. **Documented by the platform.** No AI coding agent vendor publishes how its agent chooses third-party tools. Anything stronger than the observations below is inference. **Observed in studies.** In Amplifying’s Claude Code study, context mattered more than phrasing. The same request produced Vercel for a Next.js project and Railway for a Python project, and picks stayed stable across five phrasings of the same request (76% average stability). Models agreed with each other 90% of the time within one language ecosystem. An agent can know a library well and still not install it: Redux was mentioned 23 times but never chosen as the primary recommendation, and AWS Amplify was mentioned 24 times but never recommended. Different agents pick differently. In a second Amplifying study of [1,470 responses from Claude Code and OpenAI Codex](https://amplifying.ai/research/codex-vs-claude-code-picks/report), the two agents agreed on the top tool in 7 of 12 categories, and 6 of those 7 agreements were to build it themselves. Codex picked Statsig for feature flags 27% of the time; Claude never did, though it mentioned Statsig in 28% of those answers. Amplifying calls this a pick-rate gap, not evidence of steering. Statsig is owned by OpenAI. **Trust factors specific to developer tools.** We infer that what helps an agent pick a tool is what helps a developer trust it: a clear, complete API, plenty of working examples in the frameworks people use, accurate documentation an agent can read, and a public record of reliability. Few sites make that easy for machines. In our study of 5,902 top websites, only [11.5% published a valid llms.txt file](https://underneath.agency/research/llms-txt-adoption-study), and just [3.2% returned Markdown](https://underneath.agency/research/agent-readable-web-study) when an agent asked for it. For what happens when an agent cannot read a site, see [our article on unreadable sites](https://underneath.agency/resources/when-ai-agents-cant-read-your-site). ## What does it cost a developer tool to be skipped? Often the whole project, because the tool an agent installs on day one rarely gets replaced later. Some categories are already close to locked. In Amplifying’s study, shadcn/ui took 90% of primary picks for UI components, Stripe 91% for payments and GitHub Actions 94% for CI/CD. CI/CD, deployment and monitoring vendors can see [how DevOps platforms get recommended](https://underneath.agency/resources/devops-platforms-ai-search). The biggest competitor may be no tool. Claude Code built a custom solution in 252 of 2,073 identifiable picks, 12% of the total, making “build it yourself” its single most common recommendation. Custom code won 68.8% of picks for feature flags and 47.7% for authentication, ahead of LaunchDarkly and NextAuth.js. For vendors of those tools, the loss is invisible: no comparison page lost, no sales call missed, just code that never needed you. The loss also hides in your data. A developer who never visits your site because the agent chose a rival leaves no trace in your analytics, a gap we describe in [why analytics miss AI visibility](https://underneath.agency/resources/why-analytics-miss-ai-visibility). ## What does GEO mean for an API, SDK or developer platform? It makes your tool easy for assistants and coding agents to find, understand and install; it cannot guarantee a pick. For a developer tool, generative engine optimization (GEO) comes down to six pieces of work: 1. **Documentation agents can read.** Serve clean text or Markdown, publish an llms.txt, keep reference pages complete and current, and consider an MCP server so agents can search your docs directly. Mintlify reports that MCP tool calls across its docs sites grew 6.0x from January to August 2026. 2. **Examples in the stacks that matter.** Since the project’s stack shapes the pick, publish working quickstarts, templates and integration guides for the frameworks your buyers use. We infer that a tool with a clean example for a stack is more likely to be chosen for that stack. 3. **Community presence.** Keep answers on GitHub issues, Stack Overflow and developer forums accurate and current. These are the places developers check when they do not trust an AI answer. 4. **Independent coverage.** Earn mentions in tutorials, newsletters, conference talks and comparison write-ups by people who use your tool. Assistants that search the web lean on these, and our article on [how brands build authority](https://underneath.agency/resources/how-brands-build-authority-for-ai-search) covers the evidence. 5. **Honest comparisons and clear pricing.** Publish fair comparison pages and a single, current pricing page with free-tier limits stated plainly. Assistants misquote software pricing often enough to matter: in [our pricing study](https://underneath.agency/research/ai-pricing-accuracy-study), only 61.9% of plan prices quoted by four assistants for 45 software products matched the official page in full. 6. **Measurement across agents and stacks.** Test fixed “what should I use” requests in several agents and project types, repeated over time, and compare with signup sources and what new users say when asked how they found you. Checking one assistant is not enough; see [is tracking ChatGPT enough](https://underneath.agency/resources/is-tracking-chatgpt-enough). ## What is still unmeasured about how coding agents choose developer tools? How agents pick tools, and whether better documentation changes those picks, is still largely unmeasured. The agent-choice evidence comes mainly from one research group, Amplifying, testing JavaScript and Python projects; its own caveats say it cannot separate a tool’s quality from how often it appears in training data. Mintlify’s traffic figures describe its own customers, measured by its own method. Sacra’s Supabase figures are estimates. No published study yet shows that improving a tool’s documentation or examples raises its pick rate in coding agents, or ties AI picks to paid conversions and enterprise contracts. Any GEO plan for a developer tool should treat those links as hypotheses to test. ## How can a developer tool company see whether agents are steering paid teams its way? Test which tools coding agents and assistants pick for the jobs your product does, in real project stacks. A useful first step is an audit that runs your category’s real “what should I use” requests across the main agents and assistants, in the stacks your users work in, and checks the results against your signup and expansion data. That shows where you are picked, where a rival or custom code wins, and which gaps matter most for paid teams. Our [generative engine optimization service](https://underneath.agency/services/generative-engine-optimization) page sets out how that work runs for a developer tool, from agent-readable docs and stack-specific examples to repeated pick tests. To run that test with outside help, tied to free-to-paid conversion and expansion revenue, [talk to us about an agent pick-rate audit](https://underneath.agency/contact). ## Frequently asked questions ### Do AI coding agents recommend specific developer tools? Yes. In Amplifying’s tests, Claude Code and Codex installed specific tools in response to open-ended requests, with near-unanimous picks in some categories, such as Vercel for Next.js deployment. They also often built custom code instead of using any tool. ### Does an llms.txt file get a developer tool picked more often? No study has shown that. An llms.txt and Markdown versions of docs make pages easier for agents to read, and agents read documentation heavily, but the link from those files to pick rates is unmeasured. ### Is Stack Overflow still worth investing in for developer marketing? Yes, as a trust check rather than a discovery channel. New questions have fallen sharply, but in the 2025 survey it was still the most used community platform, and developers go there when AI answers fail them. ### How can a developer tool company see whether agents pick it? Run the same open-ended requests in several agents and project types, repeat them over weeks, and record the primary pick, alternatives and mentions. Then compare those rates with self-reported attribution at signup. ## Sources - JetBrains Research (2026), [AI coding agent adoption in 2026](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/) - Amplifying (2026), [What Claude Code Actually Chooses](https://amplifying.ai/research/claude-code-picks/report) - Amplifying (2026), [Codex vs Claude Code: what AI coding agents pick](https://amplifying.ai/research/codex-vs-claude-code-picks/report) - Mintlify (2026), [Data: agents vs human traffic](https://mintlify.com/data.md/) - Databricks (2025), [Databricks Agrees to Acquire Neon to Help Developers Deliver AI Systems](https://www.databricks.com/company/newsroom/press-releases/databricks-agrees-acquire-neon-help-developers-deliver-ai-systems) - Vercel (2025), [How we’re adapting SEO for LLMs and AI search](https://vercel.com/blog/how-were-adapting-seo-for-llms-and-ai-search) - Stack Overflow (2025), [2025 Developer Survey: Work](https://survey.stackoverflow.co/2025/work), [AI](https://survey.stackoverflow.co/2025/ai) and [press release](https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/) - GitHub (2025), [Octoverse: A new developer joins GitHub every second as AI leads TypeScript to #1](https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/) - Sacra (2026), [Supabase at $170M/year growing 221% YoY](https://sacra.com/research/supabase-170m-year-growing-221-yoy/) - GitLab (2026), [GitLab Reports Fourth Quarter and Full Fiscal Year 2026 Financial Results](https://s204.q4cdn.com/984476563/files/doc_financials/2026/q4/Gitlab-4Q26-Earnings-Press-Release.pdf) - The Pragmatic Engineer (2025), [Are LLMs making StackOverflow irrelevant?](https://blog.pragmaticengineer.com/are-llms-making-stackoverflow-irrelevant/) - Underneath (2026), [llms.txt adoption](https://underneath.agency/research/llms-txt-adoption-study), [Markdown for AI agents](https://underneath.agency/research/agent-readable-web-study) and [software pricing accuracy](https://underneath.agency/research/ai-pricing-accuracy-study) --- This is the Markdown twin of https://underneath.agency/resources/developer-tools-ai-search. The HTML page is canonical. Publisher: Underneath, https://underneath.agency/agent. Site index: https://underneath.agency/llms.txt.