What are these files?
llms.txt is a proposed short guide to a website's purpose and useful information. Markdown is a simple way of writing headings, paragraphs, lists and links. A site can offer a Markdown version of a page for software designed to read it. Neither is a replacement for the customer-facing page. The llms.txt proposal
An accessibility tree is different again. Your browser builds a representation of page controls and content for assistive technology. Clear button names and labelled fields can also help a browser agent navigate. You do not commission an “accessibility-tree file”. Agent-friendly website guidance
Why do some audits mention llms.txt?
Read what the audit measures. Google's Search guidance says Search ignores llms.txt and special Markdown for visibility. Chrome's Lighthouse also has an agentic-browsing check, but its published rule treats a missing optional llms.txt file as Not Applicable; it flags a server error when trying to retrieve one. Those are different products and findings. Google Search guidance, Lighthouse's rule
If a supplier says “Google's audit mentions it, so you need it to rank”, ask them to identify the exact rule. A valid optional-file check does not support that sales claim. There may still be a sensible documentation use, but it needs its own reason and cost.
What should a small business do first?
Make the essential buying information easy to find: what you offer, whom it is for, location or service area, price conditions, availability, delivery, contact details and relevant evidence. Use real descriptions and examples instead of publishing many pages that repeat the same promise.
Check your most useful page in Search Console's URL Inspection tool. Confirm the intended public address is reachable and has not accidentally been excluded from indexing. Ask your website provider to investigate an unexpected redirect or a different canonical address; a canonical is the page address you nominate as the main version. A successful live test still does not guarantee inclusion in search. If Google AI eligibility is the concern, ask the provider to check Settings, Search generative AI in Search Console, including any inherited choice. Include is the default; an exclusion can apply separately from ordinary search and is not a training setting. Google’s control explanation
Then try the customer journey on your phone. Can someone understand the offer, choose the right option and reach an enquiry or booking? Fix unclear field labels, hidden terms and broken steps. Those changes address a real customer problem without needing a prediction about how an assistant ranks businesses.
For local discovery, keep the website and relevant business profiles consistent about service area, contact details and hours. If ChatGPT search access is in doubt, ask your provider about OAI-SearchBot. OpenAI documents it separately from GPTBot, its training crawler, so search access does not require agreeing to training access. OpenAI's crawler guidance
How do you judge a proposal with several technical extras?
Use this fictional example. A local repair business has one clear service page and a booking form. A supplier proposes a £300 llms.txt-and-Markdown package plus £25 monthly maintenance. The figures are invented for this buying exercise, not market prices.
The initial audit says “llms.txt missing”. Under the current Lighthouse rule, a normal missing-file response is not the server-error failure described above. Ask for evidence of the actual problem before paying to “clear” it.
On inspection, the public page correctly says the call-out starts at £60 and parts are additional. The proposed Markdown copy says only “repair £60”. That creates a material mistake: a customer or assistant could read it as the full job price. The package should not be accepted merely because the new file exists.
The booking form also contains three controls called “Choose” with no understandable association to the service options. That is a different repair. A clearer interface can help visitors select the right service; rewriting the page in Markdown will not make those controls meaningful.
Decide the scope from the observed need. With no identified agent consuming Markdown, the owner can keep the accurate public price wording and first commission usable form controls. Reject the inaccurate conversion if it is offered. A later documented client may need Markdown; generate it from the same maintained source and test the price conditions before accepting that additional deliverable.
Is automatic Markdown conversion a simpler option?
Some platforms can create it from the existing page. Cloudflare's current feature is an example: it converts enabled HTML responses when a client requests Markdown. If your site supplies no content-use signal, the converted response adds a default saying AI training, search and AI-input use are allowed. Ask your provider whether that matches the owner’s intended policy before enabling it, and check the converted facts. The signal cannot guarantee how others use the content. A format feature is not a search-placement guarantee. Cloudflare's feature documentation
An accurate, automatically maintained file that a client uses may be worth keeping. Check who ensures every offered version changes when your price or refund terms change. Another unreliable price list creates work for the business, even when its initial setup was inexpensive.
How will you know whether anything helped?
Write down the problem and repeat the same check after the repair. Can the tool retrieve the information? Does it preserve the relevant conditions? Can a person or authorised agent use the form? Those are testable results.
Track AI mentions, website visits and customer outcomes separately. A few correct answers in an assistant are encouraging observations, not proof of more sales. Begin with the smallest useful repair and use the weekly measurement approach to decide whether further spending is justified.