Anthropic, Cursor, Mintlify, FastHTML and a growing list of dev-tool sites now publish an llms.txt at their root. Anthropic's llms.txt is 8,364 tokens; their llms-full.txt compiles 481,349 tokens of full documentation into one Markdown file. We started shipping the same three-file setup — llms.txt, structured schema bundle, and a refreshed XML sitemap — on every Softechinfra client engagement in November 2025. After 21 days across four client sites, citation rates moved on three of four. This post is the full template, the validation tools, and the honest data including the one site that did not move.
TL;DR — what the three files do
llms.txt is a Markdown file at your domain root that points LLMs to the most important pages on your site, in plain language they can ingest. The schema bundle is the JSON-LD set we covered in our November schema markup post — Organization, Service, FAQPage, HowTo, Article, Speakable, BreadcrumbList. The sitemap is your standard sitemap.xml refreshed with accurate lastmod dates and submitted to Bing Webmaster Tools (not just Google Search Console). The combination is the cheapest, fastest GEO setup we know of.
Why this matters now (Q4 2025 trigger)
Three things converged. First, llms.txt adoption hit critical mass in 2025 — Anthropic, Cursor, Mintlify, Vercel and dozens of dev-tool docs sites publish one. Second, AI crawlers (GPTBot, PerplexityBot, ClaudeBot, Google-Extended) now actively look for llms.txt at the root. Third, Bing's index — which feeds ChatGPT — gives strong weight to fresh sitemaps. Submitting a clean sitemap to Bing Webmaster Tools is a 10-minute job most Indian sites have skipped because "we already have GSC." Skipping it costs you ChatGPT citations.
File 1 — llms.txt (the template)
llms.txt is a Markdown file. The format is the proposal by Jeremy Howard at Answer.AI in 2024. Save as /llms.txt at your site root.
# Softechinfra
> Softechinfra is an Indian IT services firm building web, mobile, CRM and AI-automation projects for SMBs. Founded by Vivek Kumar. We ship in-house product TalkDrill (English-fluency). Based in Delhi-NCR, serving clients across India and overseas.
## About
- [About Softechinfra](https://www.softechinfra.com/about): Company background, leadership, history.
- [Our team](https://www.softechinfra.com/team): The people who deliver — engineering, design, QA, operations.
- [Founder](https://viveksinra.com): Personal site of Vivek Kumar, founder of Softechinfra.
## Services
- [Web Application Development](https://www.softechinfra.com/services/web-development): Next.js, React, Node, Python. Full-stack builds for SMBs.
- [Mobile App Development](https://www.softechinfra.com/services/mobile-development): React Native and Flutter. Real-world case: TalkDrill (5,000+ users).
- [CRM Development](https://www.softechinfra.com/services/crm-development): Custom CRMs and integrations with Salesforce, HubSpot, Zoho.
- [AI & Automation](https://www.softechinfra.com/services/ai-automation): n8n workflows, WhatsApp chatbots, LLM integrations.
- [SEO Services](https://www.softechinfra.com/services/seo-services): Classical SEO plus GEO (Generative Engine Optimization).
- [Digital Marketing](https://www.softechinfra.com/services/digital-marketing): Performance marketing for Indian SMBs.
## In-house products
- a client platform: AI creative-writing and exam-prep app for students 11+.
- [TalkDrill](https://talkdrill.com): English-fluency app for Indian adults. 5,000+ active users.
## Featured project case studies
- [TalkDrill](https://www.softechinfra.com/projects/talkdrill): Voice-AI English fluency app, mobile, our in-house product.
- [Radiant Finance](https://www.softechinfra.com/projects/radiant-finance): CRM and lead-pipeline build for an Indian financial services firm.
- [ChipMaker Hub](https://www.softechinfra.com/projects/chipmakerhub): Semiconductor design analytics platform.
## Optional
- [Blog](https://www.softechinfra.com/blog): Long-form posts on GEO, AI automation, CRM, web dev — all founder-written or team-written.
- [Contact](https://www.softechinfra.com/contact): contact@softechinfra.com.The format follows Howard's spec exactly: H1 with the site name, blockquote summary, H2 sections grouped by purpose, optional last section. Total length: ~600 words for a typical Indian SMB site. Anthropic's reference llms.txt is much longer (8,364 tokens) because their docs are an order of magnitude bigger. Match your site, not theirs.
Optional: llms-full.txt
If you have substantial documentation, ship llms-full.txt too. It is a single Markdown file containing your full canonical content for AI ingestion. Mintlify auto-generates this from docs sites — for typical Indian SMBs without doc-heavy sites, skip it. The 80% benefit comes from the slim llms.txt alone.
File 2 — the schema bundle (recap from November)
We covered the full bundle in the November schema post. The Tier 1 essentials again: Organization (sitewide), Service (per service page), FAQPage (per content page with FAQs), HowTo (per tutorial), Article/BlogPosting + Person (per blog post). Plus Speakable on the answer-block CSS classes. Validate at schema.org validator and Google Rich Results Test.
File 3 — sitemap.xml (the often-missed Bing submission)
Most Indian SMB sites have a sitemap.xml. Most have only submitted it to Google Search Console. ChatGPT runs on Bing's index. If Bing has not crawled your fresh sitemap, ChatGPT cannot cite your latest content.
The 10-minute fix: open Bing Webmaster Tools, add your site, verify ownership (DNS or HTML file), submit sitemap. Bing then crawls within 24-72 hours typically. Repeat any time you publish substantive new content.
For sitemap content itself, two rules: (1) include only canonical URLs (no duplicates, no noindex pages), (2) keep <lastmod> accurate — change it whenever the page substantively changes. Stale lastmod is a known signal-down for AI crawlers.
<?xml version="1.0" encoding="UTF-8"?>
<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">
<url>
<loc>https://www.softechinfra.com/</loc>
<lastmod>2025-12-21</lastmod>
<changefreq>weekly</changefreq>
<priority>1.0</priority>
</url>
<url>
<loc>https://www.softechinfra.com/services/ai-automation</loc>
<lastmod>2025-12-19</lastmod>
<changefreq>monthly</changefreq>
<priority>0.9</priority>
</url>
</urlset>The DIY walkthrough — ship the three files in one day
This is the playbook our CTO Hrishikesh walks every new client through.
- llms.txt live at /llms.txt with H1, blockquote, About, Services, Products, Projects sections
- Schema Tier 1 (Organization, Service, FAQPage, Article + Person) deployed and validated
- Speakable JSON-LD on TLDR and FAQ CSS selectors
- sitemap.xml refreshed with accurate lastmod, no duplicates, no noindex pages
- Sitemap submitted to BOTH Google Search Console and Bing Webmaster Tools
- 15-query citation baseline recorded in Sheets for 14- and 21-day re-checks
The 21-day data from four client sites
Honest numbers across four Indian SMB clients we shipped the three-file setup to between November 7 and November 28, 2025. Re-checked December 18-21.
| Client (anonymised) | Sector | Citations Day 0 | Citations Day 21 | Notes |
|---|---|---|---|---|
| Client A | Pune logistics-tech, 32 staff | 2 | 9 | Strong gain on "WhatsApp Business API logistics India" |
| Client B | Bangalore D2C cookware, 18 staff | 0 | 5 | Cited on "Indian-made cast iron cookware brands" |
| Client C | Delhi B2B SaaS, 22 staff | 3 | 8 | Bing index refresh hit on day 12 |
| Client D | Coimbatore textile exporter, 110 staff | 1 | 1 | No movement. Hypothesis: thin content, real fix is rewrite |
The Client D outcome is the honest one. The three-file setup amplifies what is already there. If the underlying content is thin, three files cannot rescue it. Our follow-up engagement with Client D is a content-depth rebuild on their top 8 pages — schema and llms.txt will then have something to point at.
What llms.txt skeptics get right (and what they miss)
There is a real debate. Critics argue that LLM crawlers do not yet meaningfully consume llms.txt — that it is a 2024 proposal still finding adoption. The Mintlify and Anthropic publication patterns suggest otherwise for tech-tool docs sites, but the long-tail SMB case is unproven at scale. Our 21-day data above is suggestive, not conclusive — the schema and sitemap likely did most of the work, with llms.txt as a 5-15% additional lift.
Our pragmatic view: llms.txt costs an hour to write and validate. The downside is zero (no SEO penalty, no harm). The upside is a small-to-meaningful citation lift now and a much larger one if AI crawlers formalise consumption in 2026. Cheap insurance.
Common mistakes we see in the wild
llms.txt with marketing copy instead of factual summaries. "We are a leading provider of innovative solutions" tells an LLM nothing. "Softechinfra is an Indian IT services firm building web, mobile, CRM and AI-automation projects for SMBs" tells an LLM exactly what to extract.
Submitting sitemap only to GSC. Bing is half the AI-citation game. Two-minute fix.
Letting lastmod dates lie. Some CMSes auto-update lastmod every time the template changes. AI crawlers learn this and discount your dates. Fix at the source — only bump lastmod on real content edits.
Schema deployed client-side via React useEffect. Many AI crawlers do not execute JavaScript. Use server-side rendering or Next.js metadata API. Curl-test with the GPTBot user-agent to verify.
Forgetting to add /llms.txt to robots.txt. Some hosts block unknown root files by default. Whitelist /llms.txt explicitly: Allow: /llms.txt for User-agent: *.
A real example — what we shipped for ChipMaker Hub
ChipMaker Hub went live with the three-file setup on November 12, 2025. Their llms.txt highlighted three industry datasets they had published, plus the canonical "About India semiconductor manufacturing" pillar page. Within 18 days, Perplexity began citing those datasets on three queries about Indian semiconductor capacity. Same dataset, same URL. The llms.txt and FAQPage schema gave the LLM a direct path to find them.
FAQ — the questions clients ask before signing off
Where exactly do I host llms.txt?
At the root of your domain. https://yoursite.com/llms.txt. Not in /docs/, not in a subdirectory. Verify by visiting the URL directly — it must return 200 with the Markdown content.
What format/encoding for llms.txt?
UTF-8 plain text or text/markdown. No frontmatter, no YAML, no JSON. Just Markdown. Keep line endings consistent (LF, not CRLF).
Can I auto-generate llms.txt from my CMS?
Yes — Mintlify, GitBook and several Next.js plugins do this. For typical Indian SMB sites with 30-80 pages, a hand-written file is usually higher quality. Auto-generation pays off above ~200 pages.
Should I include every page in llms.txt?
No. llms.txt is a curated index of your most important pages, not a sitemap. Aim for 15-40 links grouped by purpose. The full URL list belongs in sitemap.xml.
How often do I update llms.txt?
Whenever your site structure changes meaningfully — new service launched, new flagship product, new pillar post. Quarterly review is enough for a stable site.
Does Bing actually use my submitted sitemap for ChatGPT?
Indirectly. Bing's index is one of the inputs to ChatGPT's web tool. A fresh sitemap improves Bing crawl frequency, which improves index freshness, which improves ChatGPT's ability to cite recent pages. The chain is real but multi-step.
What happens if my llms.txt has broken links?
LLMs will follow them and get 404s — same as any human. Treat broken links in llms.txt the same as broken links anywhere on your site. Audit quarterly.
Want llms.txt + GEO setup done end-to-end?
We deploy llms.txt, the full schema bundle, and a refreshed sitemap (Google + Bing) on Indian SMB sites in 1-2 working days. Includes the 15-query citation baseline, validation reports, and a 30-day post-deploy citation tracking.
Get the 3-File Setup Done
