GEO & AI Search
Why llms.txt and Direct-Answer Blocks Are the New Technical SEO Standard
, 3 minute read
Direct-answer blocks and llms.txt both exist to make a site easier for language models to use. A direct-answer block is a 40 to 60 word answer under a question heading, and it is the more important of the two. llms.txt is a proposed Markdown index at the site root; it is inexpensive but not widely honored.
Two ideas, unequal weight
The title of this article pairs two practices, and honesty requires separating them straight away. Direct-answer blocks are a writing and markup discipline with a clear mechanism behind them. llms.txt is a proposal that costs an hour and may or may not be read. We do both. We would give up the second long before the first.
Direct-answer blocks
Search engines have extracted passages for years: featured snippets, People Also Ask, voice answers. Generative systems do the same thing at larger scale, retrieving passages and composing a response from them. In every case the unit that gets selected is a short span of text that answers a question on its own.
A direct-answer block is that span, written on purpose:
- It sits immediately under a heading phrased as the question
- It runs 40 to 60 words, long enough to be complete and short enough to quote whole
- Its first sentence is the answer
- It names its subject in full, with no pronouns reaching back to earlier text
- It includes the specific fact that makes it true
- It contains no sales language
Compare two answers to the same question about water heaters.
Weak: we have been helping homeowners with this for years, and it really depends on a lot of factors.
Strong: a tankless water heater typically lasts 20 years or more, compared with 10 to 15 years for a storage tank model, provided it is descaled regularly in hard-water areas.
The second can be lifted into an answer and attributed. The first cannot be used at all.
Mark the block up plainly: a paragraph inside a section with an id, under an h2. No script is needed. We add a data attribute so our own tooling can find and count them, which is how the X-ray panel on this site reports the word count of each block. The full method is under direct-answer engineering, and the markup it relies on is semantic HTML5.
llms.txt
The proposal, published at llmstxt.org in 2024, is a Markdown file served at /llms.txt. Its format is simple:
# Site name
> One-paragraph summary of what the site is and who it serves.
## Services
- [Service one](https://example.com/service-one): what it is, in one line
- [Service two](https://example.com/service-two): what it is, in one line
## Contact
- Phone: 747-555-0100
The intent is to give a language model a curated index, so it does not have to infer the site's structure from navigation and scripts.
Now the caveats. It is not a ratified standard. Google has said it does not use it. Support among AI providers is partial and changes. No one can show you a ranking or a citation that exists because of this file alone.
We publish one anyway, for practical reasons. It is cheap. Writing it forces a business to state what it does in two hundred words, which improves the home page too. And agents that browse on a user's behalf sometimes do look for it. You can read ours at /llms.txt.
Where robots.txt fits
Access rules matter more than either. A site that blocks AI crawlers at the firewall or in robots.txt has opted out, whatever else it publishes. Decide deliberately which bots may fetch your pages, distinguishing training crawlers from those that retrieve pages to answer a live question. That decision belongs in the same review as your crawl and index hygiene.
An order of work
- Confirm crawlers can reach the site and that content is in the raw HTML.
- Rewrite primary headings as questions customers ask.
- Add a direct-answer block under each.
- Align business facts across the site, schema and profiles, the work we call entity co-occurrence.
- Publish llms.txt.
If you stop after step four you have done nearly all of what matters. For the wider picture, see our guide to generative engine optimization and the service page on the llms.txt file.
Questions this article raises
Does Google use llms.txt?
Google has said it does not. Publish the file for other agents and tools if you wish, but do not expect it to influence Google Search or AI Overviews.
Why 40 to 60 words?
It is long enough to give a complete answer with one qualifying detail, and short enough to be quoted in full. Featured snippets have long favored passages of about this length.
Is llms.txt the same as robots.txt?
No. robots.txt tells crawlers what they may fetch. llms.txt is a summary and index intended to help a model understand the site. One controls access, the other offers guidance.
Talk to the person who will actually do the work
No account coordinator, no discovery deck. Call and describe what the phone should be ringing for, or send the site and we'll trace where the signal is leaking.