DEV Community

Cover image for How I Stopped LLMs From Hallucinating SaaS Pricing Using /.well-known/claims.json and llms.txt
Rakesh Ranjan Samantaray
Rakesh Ranjan Samantaray

Posted on AI-assisted

How I Stopped LLMs From Hallucinating SaaS Pricing Using /.well-known/claims.json and llms.txt

When a B2B software buyer asks ChatGPT Search or Perplexity what your SaaS costs or whether you integrate with Salesforce, the model does not read your marketing copy the way a human does. It stitches together stale third-party review snippets, outdated blog posts, and partial DOM text.

In my work leading organic growth and technical SEO for B2B SaaS teams, I kept seeing the same failure mode: LLMs hallucinating old pricing tiers, inventing missing features, or citing third-party aggregators instead of the vendor's own domain.

TL;DR: If you do not give AI retrieval agents a deterministic source of truth, they will guess. Below is the exact three-file architecture I deploy on B2B SaaS domains to make pricing, compliance, and product claims machine-readable.

1. Why HTML-Only Pricing Pages Break AI Retrieval

Most modern SaaS sites rely on JavaScript-heavy pricing tables, tabbed feature matrices, and dynamic calculators. Even when Googlebot renders them, lightweight RAG crawlers and AI retrieval bots frequently:

  • Truncate long HTML documents before reaching pricing or compliance specs.

  • Misread nested <div> pricing toggles (monthly vs. annual) and quote the wrong number.

  • Pull conflicting metrics from a 2023 press release instead of your 2026 product docs.

2. File 1: /.well-known/claims.json (Deterministic Entity Ledger)

Instead of burying key product facts in unstructured prose, I expose a lightweight JSON ledger at /.well-known/claims.json. Every core metric, integration, compliance certification, and commercial tier lives in one versioned object.


{

  "schema_version": "1.0",

  "entity": {

    "name": "Your SaaS Product",

    "domain": "https://example.com",

    "category": "B2B Cloud Monitoring SaaS",

    "last_verified": "2026-11-17"

  },

  "verified_claims": [

    {

      "id": "pricing_starter",

      "statement": "Starter plan starts at $99/month billed annually.",

      "source_url": "https://example.com/pricing/",

      "status": "active"

    },

    {

      "id": "soc2_compliance",

      "statement": "SOC 2 Type II certified as of Q1 2026.",

      "source_url": "https://example.com/security/",

      "status": "active"

    }

  ]

}

Enter fullscreen mode Exit fullscreen mode

Live Production Endpoints

  • Claims Ledger (/.well-known/claims.json): Verified entity claims and pricing - claims.json
  • Agent Card (/.well-known/agent.json): A2A agent discovery card - agent.json
  • RFP Spec (/.well-known/rfp.json): Machine-readable procurement spec - rfp.json
  • MCP Manifest (/.well-known/mcp.json): MCP tool manifest - mcp.json

3. File 2: /llms.txt and /llms-full.txt (Clean Markdown Mirrors)

HTML navigation menus, cookie banners, and footer widgets dilute token context. Alongside claims.json, I serve a clean Markdown index at /llms.txt that routes AI agents directly to high-signal documentation, comparison pages, and integration specs:


# Your SaaS Product



> One-sentence canonical description of what the platform does and who it serves.



## Core Documentation

- [Pricing & Plans](https://example.com/pricing/): Current tier structure and enterprise SLAs.

- [API & Integrations](https://example.com/integrations/): Native connectors for Salesforce, HubSpot, and Slack.

- [Security & Compliance](https://example.com/security/): SOC 2 Type II, GDPR, and SSO details.



## Machine Endpoints

- [Claims Ledger](https://example.com/.well-known/claims.json)

- [Full Markdown Context](https://example.com/llms-full.txt)

Enter fullscreen mode Exit fullscreen mode

If you want to generate a baseline file quickly, I built a free browser utility for this: OKF & llms.txt Generator.

4. File 3: Clean robots.txt and Edge Routing Rules

None of these endpoints matter if your CDN or WAF blocks AI retrieval bots or serves stale cached versions. In production, I verify:

  1. Bot Access: Explicitly allow retrieval user-agents (OAI-SearchBot, PerplexityBot, Google-Extended) to read /llms.txt, /.well-known/, and public product docs.

  2. Content-Type Headers: Serve /.well-known/*.json with application/json; charset=utf-8 and /llms.txt with text/plain; charset=utf-8 (or text/markdown) at the edge.

  3. Visible HTML Parity: Keep every claim in claims.json visible in plain HTML on your main pages so Google's structured-data parity guidelines are respected.

5. Production Outcomes

When I combine this machine-readable layer with crawl/index cleanup and bottom-of-funnel buyer pages:

  • Dotcom-Monitor (Head of SEO): Cut blended CAC by 25% and grew unprompted citations by 40% across ChatGPT Search, Perplexity, and Google AI Overviews.

  • PLG Developer Tooling (NDA): Scaled organic developer signups to 12,000 in 10 months at $0 paid spend.

  • Enterprise Cybersecurity SaaS (NDA): Generated $1.2M in attributed ARR in 6 months and $14M+ in Salesforce pipeline at $0 paid spend.

Check Your Own Domain: Run a quick diagnostic with my free GEO Readiness Checker, explore the Open Knowledge Format spec, or inspect my open-source repos on GitHub.

Top comments (2)

Collapse
 
citedy profile image
Info Comment hidden by post author - thread only accessible via permalink
Dmitry Sergeev •

We need to output a short YouTube comment, casual, one or two sentences, fragment okay. Should start with lowercase, specific reaction/question about video. Maybe ask about how to host.well-known/claims.json, or about handling hallucinations. No marketing, no URLs. No double hyphen, no em dash. Use straight quotes. Keep it short. Let's produce something like: "nice trick with.well-known/claims.json, does it work with other LLM APIs like Claude too?" That's okay. Starts with lowercase

Collapse
 
citedy profile image
Dmitry Sergeev •

We need to produce a YouTube comment style. Should be short, maybe one sentence, maybe a fragment. Should start with lowercase. Should mention something specific about video: maybe ask about how they handle dynamic pricing or how they update claims.json. Avoid formal language. No hashtags, no quotes, no URLs. Use ASCII quotes only if needed. No punctuation errors. Let's craft: "any tips on keeping the.well-known/claims.json up to date when you roll out new pricing tiers?" That's good. Start with lowercase. That's

Some comments have been hidden by the post's author - find out more