# Generative Engine Optimization for Developer Tools > GEO for developer tools and B2B software: what AI cites at the evaluation stage, how coding agents pick dependencies, and how to write docs agents can use. Source: https://metavert.io/industry/generative-engine-optimization-for-developer-tools Published: 2026-10-07 Updated: 2026-10-07 Industry Application Generative Engine Optimization Developer Tools **Generative engine optimization for developer tools** is the practice of getting an API, SDK, platform or B2B software product selected by AI systems, and it differs from every other vertical in one respect: there are two machine audiences. The first is the chat assistant a human buyer consults while evaluating vendors. The second is the [AI coding agent](https://metavert.io/ai-coding-agents) that chooses a dependency and installs it, often without a human comparing anything. The two read different material and reward different work, and most of the general [generative engine optimization](https://metavert.io/generative-engine-optimization) advice addresses only the first. ## What AI Cites at the Evaluation Stage In most industries a brand's own site is a minor source for AI answers. Kumar's five-engine study put brand-owned domains at 2.9% of citations. B2B software evaluation looks different. Ten Speed's analysis of 7,387 citations across 220 evaluation-stage prompts (September 2026; agency research, modest sample) found vendor product pages were the largest category at 24.1%, followed by comparison content at 13.3%, listicles at 13.2%, G2 and Capterra at 7.2% and community sources at 4.2%. The plausible reading is that evaluation questions are specific (does it support this framework, what does the next tier cost, how does it compare with a named rival) and the vendor's own pages are often the only place the answer exists. That puts weight on product, pricing and comparison pages that state facts plainly and stay current. Review sites carry less citation volume than their reputation suggests, but they carry trust: in G2's April 2026 survey of 1,076 B2B software decision-makers, 45% said a review-site citation was the most confidence-inspiring signal in an AI response. ## Coding Agents Are a Second Channel When an agent scaffolds a project it makes vendor decisions. Armature's September 2026 study of 5,292 sessions across [Claude Code](https://metavert.io/claude-code), [Codex](https://metavert.io/codex) and [Cursor](https://metavert.io/cursor) (vendor research) found [Stripe](https://metavert.io/stripe) taking roughly nine in ten payments picks and Neon 66% of database picks. PayPal appeared 139 times in agent conversations and was never selected. Choices tracked language ecosystems, with Resend winning email in TypeScript projects and SendGrid in Python, and [Supabase](https://metavert.io/supabase) lost database picks when it was presented as a bundle rather than a database. Three findings explain why this channel is hard to move. Models favor what they saw most in training: an ACL 2026 paper found Flask in 88% of responses against 9% for FastAPI, a preference for older, widely documented libraries that is discussed under [training data frequency](https://metavert.io/training-data-frequency). Model providers tilt toward their own ecosystems: University of Zurich researchers found six of ten provider-affiliated models significantly favored them, by up to 18.8 percentage points. And the default competitor is no vendor at all. In Amplifying's July 2026 edition of its Claude Code study, custom code accounted for 21.4% of picks, although 32.5% of those named a vendor to swap in later. ## Documentation Written for Agents Documentation traffic has changed character. Mintlify reported that 66% of traffic to the developer docs it hosts came from agents in July 2026, up from 15.2% at the start of the year (vendor data from its own customer base). Ahrefs' log study of 137,210 domains found coding agents and agent infrastructure were the largest AI category of requests at 10.5%, far above AI retrieval bots at 1.1%. This is the context in which [llms.txt](https://metavert.io/llms-txt) should be judged. The same Ahrefs study found 97% of llms.txt files received zero requests, and no AI bot goes looking for one that does not exist, so it is not a visibility lever for chat answers. For documentation sites the picture is narrower and more favorable: Mintlify's benchmark found adding llms.txt cut agent 404 errors by about 90%. The file helps an agent that is already on the site find the right page. Publishing agent-readable copies is increasingly a platform feature rather than a build step; [LightCMS](https://metavert.io/lightcms), the CMS serving this site, generates llms.txt and a clean [Markdown](https://metavert.io/markdown) copy of every page automatically. Beyond reading docs, agents increasingly call tools directly. The Linux Foundation counted more than 10,000 published [MCP](https://metavert.io/mcp) servers and more than 60,000 projects using [AGENTS.md](https://metavert.io/agents-md) when it formed the Agentic AI Foundation in December 2025. An official MCP server or a documented quickstart that an agent can execute end to end is product work that doubles as distribution, the subject of [agent experience](https://metavert.io/agent-experience). ## What Is Not Supported, and How to Measure The content-rewriting tactics from early GEO guides have not survived testing. A September 2026 re-measurement found the original levers move citation on none of ten modern engine families, and Ahrefs found no major citation uplift from adding schema markup. Outcome data is mostly self-reported: Tally said in September 2026 that 43% of its new users arrived via AI assistants, and Webflow attributed about 10% of signups to AI in late 2025. These are single-company statements, useful as existence proofs and not as benchmarks. In Armature's data the three agents agreed on the same tool only 42% of the time, so measurement needs to cover each agent separately and repeat each prompt. ## Applications & Use Cases #### Evaluation-Ready Product Pages State supported languages and frameworks, limits, pricing tiers and compliance facts in plain text on indexable pages. These are the pages AI cites most when a buyer asks a specific evaluation question. #### Honest Comparison Pages Comparison content is the second-largest citation category at the evaluation stage. Pages that concede where a rival fits better are more useful to the buyer and harder for an AI summary to contradict. #### Agent-Readable Documentation Markdown versions of every page, an llms.txt index, stable URLs and copy-paste quickstarts reduce the dead ends agents hit. The measurable result is fewer agent 404s and more completed integrations. #### Official MCP Servers and AGENTS.md Giving agents a supported way to call the product, and repository instructions on how to use it, moves selection from what the model remembers to what it can verify in the session. #### Ecosystem-Specific Positioning Agents pick differently by language. Maintaining first-class SDKs, examples and templates in each target ecosystem matters more than a single generic landing page. #### Review and Community Presence Current reviews on G2 and Capterra, maintained GitHub repositories and answered community questions supply the third-party evidence buyers say they trust most in an AI answer. ## Key Players - **Claude Code, Codex, Cursor and GitHub Copilot** — The coding agents making dependency choices; JetBrains' August 2026 survey put Claude Code at 39% of professional developers, Copilot at 21%, Codex at 16% and Cursor at 12%. - **ChatGPT, Gemini, Claude and Perplexity** — The chat assistants buyers use for vendor research. - **G2 and Capterra** — Review sites cited in evaluation-stage answers and rated by buyers as the most confidence-inspiring source. - **GitHub** — Where repositories, examples and AGENTS.md files live. - **Model Context Protocol** — The open standard, now under the Linux Foundation's Agentic AI Foundation, through which agents call tools. - **Mintlify** — Documentation platform publishing data on agent traffic to developer docs. - **Stripe** — The clearest example of a near-default agent pick in a category, in Armature's data. - **Armature and Amplifying** — Publishers of the main public studies of which tools coding agents select. ## Challenges & Considerations - **Incumbency in the Weights** — Models prefer libraries and vendors that were heavily represented when they were trained. A newer tool can be better documented today and still be passed over until training data catches up or the agent searches the web. - **Build Instead of Buy** — Agents frequently write the functionality themselves. A product whose value is a few hundred lines of glue code is competing with the agent, not with other vendors. - **Provider Bias** — Models from companies with their own cloud or developer ecosystems measurably favor them, and early ecosystem choices in a session persisted up to 90.3% of the time in the Zurich study. Vendors cannot optimize this away. - **Vendor-Sourced Evidence** — Nearly every dataset on agent tool selection and docs traffic comes from a company selling something adjacent. The directions are consistent; the exact percentages should be held loosely. - **Two Programs, One Team** — Winning chat citations (comparison pages, reviews, earned coverage) and winning agent picks (docs, SDKs, MCP) are different projects with different owners. Developer relations and marketing rarely share a plan for both. ## Related Topics - [Generative Engine Optimization](https://metavert.io/generative-engine-optimization) — the general discipline - [AI Search for Developer Tools](https://metavert.io/industry/ai-search-for-developer-tools) — the demand side: how developers and buyers research - [AI Coding Agents](https://metavert.io/ai-coding-agents) — the second audience for developer-tool marketing - [Agent Experience](https://metavert.io/agent-experience) — designing products and docs for agent users - [llms.txt](https://metavert.io/llms-txt) — what it does and does not do - [llms.txt vs MCP](https://metavert.io/compare/llms-txt-vs-mcp) — reading docs versus calling tools - [AGENTS.md](https://metavert.io/agents-md) — repository instructions for coding agents - [Training Data Frequency](https://metavert.io/training-data-frequency) — why agents default to incumbents - [LightCMS](https://metavert.io/lightcms) — agent-native CMS with generated llms.txt and Markdown copies ## Further Reading - [What AI cites at the B2B evaluation stage](https://www.tenspeed.io/blog/what-ai-cites-b2b-evaluation-stage) — Ten Speed, September 2026 - [Half of B2B software buyers now start their research with AI chatbots](https://www.prnewswire.com/news-releases/new-g2-research-half-of-b2b-software-buyers-now-start-their-research-with-ai-chatbots-302742807.html) — G2, April 2026 - [Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines](https://arxiv.org/html/2606.20065) — Kumar, arXiv, June 2026 - [Which tools coding agents install](https://armature.tech/blog/which-tools-coding-agents-install) — Armature, September 2026 - [Claude Code picks: Fable-5 edition](https://amplifying.ai/research/claude-code-picks-fable) — Amplifying, July 2026 - [A Study of LLMs' Preferences for Libraries and Programming Languages](https://arxiv.org/html/2503.17181v3) — ACL 2026 - [Do LLMs Favor Their Providers? Measuring Vertical Integration Bias in Code Generation](https://arxiv.org/pdf/2605.28515) — University of Zurich, arXiv, May 2026 - [State of docs traffic](https://www.mintlify.com/blog/state-of-docs-traffic) — Mintlify, July 2026 - [llms.txt study](https://ahrefs.com/blog/llmstxt-study/) — Ahrefs, June 2026 - [Linux Foundation announces the formation of the Agentic AI Foundation](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation) — Linux Foundation, December 2025 - [AI coding agent adoption 2026](https://blog.jetbrains.com/research/2026/08/ai-coding-agent-adoption-2026/) — JetBrains, August 2026 - [Scoring Without the Engine](https://arxiv.org/abs/2609.07559) — Bajemon & Rochet, arXiv, September 2026 - [Schema and AI citations study](https://ahrefs.com/blog/schema-ai-citations/) — Ahrefs, May 2026 - [6 years in, 6 million far](https://blog.tally.so/6-years-in-6-million-far/) — Tally, September 2026 - [Traffic is no longer reliable](https://www.growthunhinged.com/p/traffic-is-no-longer-reliable) — Growth Unhinged, November 2025