Hack #008 · AI search

AI answers cite sentences, not websites. Write the sentence.

We publish an llms.txt and we sort AI crawlers into the ones that may cite us and the ones that may not. Neither mattered as much as rewriting the first line under every heading.

2026-09-02 · 4 min · we hit this one ourselves

By Chen Deng · OxOne, Calgary AB

Do this

Under every H2, make the first sentence answer the heading on its own, with the business, the place and the number spelled out. That sentence is the unit an assistant can lift. A well-argued page with no liftable sentence gets read and not cited.

Search is turning into an answer with three or four sources named underneath it. That layout has an uncomfortable property: being on the page is no longer the win. Being the sentence is the win.

Assistants retrieve passages, not documents. Something pulls a few paragraphs out of your page, and those paragraphs get read with everything above and below them removed. So the question to ask about your own writing is not whether the page is good. It is whether any single paragraph survives being cut out of it.

The passage test

Take a paragraph from the middle of one of your pages and read it with the rest covered. Then check:

✗ Reads well, cannot be lifted

“That is why we built it this way, and it is why our clients stay with us.”

✓ Stands alone

“OxOne builds business systems for Calgary trades, at a fixed price, in two to six weeks.”

Most marketing copy fails this badly, because good prose flows and flowing prose is full of pronouns. You do not have to write like a robot. You have to put one self-contained sentence directly under each heading, then flow as much as you like after it.

Then check you are not blocking the ones you want

Open yoursite.com/robots.txt. A lot of sites blocked AI crawlers in one panicked afternoon in 2023, using a single wildcard, and never revisited it. That is a real decision with two different halves inside it:

KindAgentsWhat blocking costs you
Retrieval
answers a live question
OAI-SearchBot
PerplexityBot
ChatGPT-User
Claude-SearchBot
Your name under the answer. Block these and you are not in it.
Training
builds a dataset
GPTBot
ClaudeBot
CCBot
Bytespider
Google-Extended
Nothing. They cite no one.

Our own robots.txt allows the first group and blocks the second, with a comment above each block saying why. You are welcome to copy it. The naming is not intuitive and the agent lists move, so a comment explaining the policy is what keeps the file correct a year from now.

llms.txt, honestly

An llms.txt is a plain-Markdown file at your root that states what you do, who you are, and which pages say what. No assistant is obliged to read it and there is no evidence it moves rankings.

Write one anyway, for two reasons. It is roughly an hour, and it forces you to state your offer, your pricing and your service area in flat declarative sentences with no marketing around them — which is exactly the material that gets quoted, whether it comes from that file or from the pages it describes. Ours made us notice three claims we could not actually substantiate.

Why this is the urgent one

Being named as the source is the only version of position one that will still exist in two years.

The click is being absorbed into the answer, and that is not reversible. Being named as the source in the answer is the only version of position one that will still exist in two years, and it is decided by whether your page contains a sentence worth lifting. Right now, most of your competitors' pages do not either. That gap closes.

Every one of these came out of work we actually did. If you would rather not do it yourself, book a free 30-minute diagnostic and we will tell you which of these your business is losing money to.

More on this Local SEO for Calgary businesses Everything we have written