Ask ChatGPT or Perplexity what the best software is for managing LP data rooms. You'll get an answer. It will cite sources. There will be winners and losers in that list. And it was decided not by which vendor had the best ad budget, or the most backlinks, or the right anchor text — but by which vendor's documentation was structured for an AI to read and repeat.
That shift is already complete. Fund managers — especially the technical ones evaluating infrastructure — no longer open ten browser tabs. They open one AI chat, ask a pointed question, and act on whatever the model surfaces with a citation. The deal gets researched by Claude before the sales call happens. The shortlist gets built by Perplexity before anyone books a demo. Most fund tech vendors have no strategy for any of this, because they're still optimizing for a game that ended.
SEO and GEO are not the same thing
Search Engine Optimization — the last decade's distribution game — was about ranking pages for keyword queries. Write content that matches "LP data room software," build backlinks, get to page one, capture clicks. The whole system was legible. You knew the inputs. You could measure the outputs.
Generative Engine Optimization is different. AI answer engines don't rank pages. They synthesize answers from structured content — and the structure that matters is not keywords or backlinks. It's machine-readable schema. JSON-LD FAQPage blocks that tell Google and ChatGPT exactly what question a piece of content answers. llms.txt files that tell AI agents how to navigate your product. Benchmark data that gives the model evidence, not just claims.
If your content isn't structured for a machine to parse and repeat, it doesn't matter how good your copy is. The model will cite someone else — probably whoever bothered to add three lines of JSON-LD to their FAQ page.
The fund manager who uses AI to research tools doesn't see your website. They see whatever the AI extracted from it. If your content wasn't structured for extraction, you don't exist in that conversation.
What we built
We shipped three things last week that together form what we're calling the AX layer — Agent Experience, the fund software equivalent of mobile-first design, except for AI readers instead of small screens.
The first is a benchmark. FundOS is built to be used by AI agents — it has a 47-tool MCP server, an llms.txt and llms-full.txt, a full CLAUDE.md integration guide. But until last week, there was no proof any of that worked. An agent pointed at a FundOS task might get the right tool or might hallucinate one. We never had a score.
The AX benchmark fixes that. It's a standalone test suite: 20 tasks covering every major FundOS capability, run under two conditions — FundOS AX (the curated documentation) versus a raw API reference baseline. Each task is scored on whether the agent called the correct tool, whether it hallucinated a non-existent endpoint, and whether it correctly flagged write actions as needing human approval before executing. The harness is open source. Anyone can clone the repo and run it.
Latest score: 40 of 40 tasks correct. Zero hallucinations. 80/80 points. 100% human-approval gate coverage.
The second thing is the GEO layer. We added /faq.json — a JSON-LD FAQPage schema with twelve Q&As covering everything from how to connect the MCP server to how capital calls work. It's the format Google, ChatGPT, and Perplexity use to generate cited answers to conversational queries. Without it, asking an AI "how do I issue a capital call in FundOS" returns a generic answer. With it, the model can cite the specific FundOS workflow with a source link. The /faq HTML page renders the same content for humans.
The third is a public results page at /ax — a live rendering of the benchmark data, updated every time we re-run the harness. Not a marketing claim. A score, with methodology, with a link to the test that produced it.
Why this matters beyond FundOS
Every fund software vendor is going to face this. The PE associate who used to read G2 reviews now asks Claude to compare options. The CFO doing vendor diligence runs the product through Perplexity before agreeing to a demo. The developer building a custom workflow asks ChatGPT which MCP tools exist for fund accounting. In each case, the model's answer is only as good as the structured content it was trained on or can fetch at inference time.
Most fund tech vendors have nothing indexed for this. Their documentation exists as a helpdesk PDF or a locked knowledge base. Their API reference is a Postman collection behind a login wall. Their FAQ page has neither schema nor structure. When the model looks for authoritative content about their product, it either makes something up or defers to a competitor who happened to publish a well-structured blog post.
The vendors who win the next five years of fund software distribution will be the ones who treat AI readability the way the last generation treated mobile responsiveness — as a baseline requirement, not a differentiator. Right now it's still a differentiator. That window is closing.
How we're measuring it
The honest answer is: imperfectly, and we're going to keep publishing the numbers anyway.
The AX benchmark is self-authored. We wrote the tasks, we defined what a correct answer looks like, we set the scoring rubric. That's a limitation worth naming. What it measures — whether an AI agent given our documentation can do the right thing — is real. But it's not a third-party evaluation, and we don't claim it is. The methodology is public specifically so anyone can read it, disagree with a task definition, and tell us.
What we can say: we run the benchmark before and after every significant documentation change. When the score drops, the docs regressed. When the score holds at 40/40, the docs are doing their job. It's a feedback loop we didn't have six months ago. The score will be wrong sometimes. But a score that's sometimes wrong is better than no score at all — which is what the rest of the industry has.
Your next customer is going to ask an AI which fund software to use. The answer they get was decided months ago, by whoever structured their documentation for a machine to read. It's not too late to be in that answer. But it won't be early for long.