SEO writing with AI: what changes when your reader is an assistant

Writing with AI for AI search changes one thing above all others: your reader is a machine looking for a single passage it can lift and still be correct. Put the answer in the first two sentences, keep one claim per sentence, and write every section so it still makes sense after somebody pulls it out of the page. The drafting tool you use matters far less than whether your page gives an assistant something exact to repeat.

Two different jobs share one name

People searching for AI and SEO writing together usually want one of two things.

Faster drafts. Using a model for outlines, first drafts, meta descriptions and variations. This is a production question. It saves time and it does not, by itself, change whether anybody is named in an answer.

Pages an assistant will quote. Writing so ChatGPT, Gemini, Perplexity, Claude or Google AI Overviews can use your page when a buyer asks a question in your category. This is the one that decides whether your name appears.

Most people arrive wanting the first and need the second, which is what the rest of this page is about. We have written separately about the structure of a page an assistant will quote, and this page goes down to the sentence level.

What actually changes when the reader is an assistant

Four changes, all of them mechanical rather than stylistic.

It reads a passage, not a page. An assistant pulls the part of your page that looks relevant and works with that. Whatever is in the extracted passage is all it has.

It has to be able to be right. A careful assistant drops a claim it cannot stand behind. A sentence with a number, a unit and a date is safe to repeat. A sentence with an adjective is not.

There is no scrolling. A person skimming can be made curious and kept reading. A machine has no reason to read past the part that answers its question.

Hedging costs you. "Solutions tailored to your needs" gives an assistant nothing to say. A specific scope, count or price gives it a sentence.

Answer-first paragraphs

The first paragraph of the page is the answer to the question in the title, written so it could be quoted alone. Not context, not a preamble.

Apply the same rule one level down. Under each heading, the first sentence is the answer to that heading and the sentences after it are the evidence. That gives an assistant a usable passage at every depth of the page rather than only at the top.

A useful test: delete everything except the first sentence under each heading. If what remains reads as a correct, if blunt, answer to the page's question, the structure is right.

One claim per sentence

A sentence carrying three claims cannot be partly quoted. If one claim is unsupported, a careful assistant drops the whole sentence, including the two that were fine. Splitting them limits the damage.

Hard to quote	Easy to quote
Our comprehensive platform supports a wide range of integrations and is trusted by leading companies across multiple industries.	The platform connects to 31 tools. The list is published on our integrations page. It was last checked on 1 October 2026.

The right hand column is less impressive and more usable. That is the trade this kind of writing asks of you.

Sections that survive being quoted out of context

Write each section as though it were the only part of the page anybody will see. Four habits do most of the work.

• Name the subject in the section, not only in the heading. "It covers 25 states" is useless alone. "The service covers 25 states" can be quoted.
• Repeat units and dates locally. A price stated in paragraph two and called "that figure" in paragraph nine loses its meaning when paragraph nine is extracted.
• Avoid back references. "As mentioned above" breaks the moment the passage travels.
• Keep lists complete. "The following six" followed by six is checkable. "A range of factors" is not.

What gets a passage cited

Five properties keep showing up in the passages that get used.

A number with a unit and a date. Exact rather than rounded. A narrower figure with a date beats a generous figure without one.

A named list instead of a total. "The following twelve" can be checked. "Over a dozen" cannot.

A source beside anything about somebody else. Competitor prices, market figures, regulatory facts. Say where it came from and the day you checked. We wrote up a case where correct but unsourced competitor prices were set aside at read time in our note on how to get recommended by Claude.

A stated limit. What the product does not do, and who should buy something else. Pages that say this get quoted in answers about the limits of a category, a question almost nobody else answers.

One question per page. A page answering eight questions loosely tends to lose to eight pages that each answer one tightly.

Where AI drafting helps, and where it hurts

It helps with the structural parts. Turning buyer questions into page titles, outlining the sub questions, rewriting a paragraph that buried its answer, finding sentences that carry more than one claim.

It hurts in one predictable place. A model asked to write about your business will produce confident specifics it has no way of knowing: coverage figures, client counts, years in business, competitor prices. Those are exactly the claims an assistant will later check, and one unsupportable number can cause a careful reader to discount the rest of the page.

So let a model shape the page and never let it supply a fact. Every number should come from somebody in your company who can say where it came from.

What this page does not cover

This page is about writing. It does not cover whether a crawler can read your page at all, it does not cover getting your name onto the third party sites assistants already read, and it does not rank AI writing tools. It also makes no claim about how any specific engine weighs sources. None of these companies publish that, the behaviour changes, and everything here should be read as a pattern rather than a rule.

Asked and answered: SEO writing with AI

Does AI written content get penalised?

Google's published guidance is about quality and usefulness rather than how a page was produced. The practical risk is not the tool. It is that generated drafts tend to be vague and sometimes invent specifics, and vague pages give an assistant nothing to quote.

Should I still do keyword research?

Yes, then write to the question rather than the phrase. Keyword tools still tell you which wording buyers use. What changes is that matching the phrase is no longer the mechanism. Being the page that lets an assistant answer confidently is.

Can I just add an FAQ section to existing pages?

It helps and it is not enough on its own. A question and answer block is easy to extract, so the shape is good. But if the answers are vague the block is still unquotable.

How do I tell whether any of this is working?

Ask your buyers' questions on a fixed list, without naming your own company in the prompt, and record how often you are named. Naming yourself in the question contaminates the result. Track that next to search clicks, because the two can move in opposite directions.

Where does AI Knows Us fit?

Writing for an assistant is a guess until you read what the assistant produces. Take the question your page was written to answer, ask it plainly, and see whether your page is the source or whether something else is. That one habit tells you more about the writing than any score. AI Knows Us does it across a set of pages at once and keeps the citations beside each answer.

What to do first

• Read only the first paragraph of your most important commercial page. If it does not answer the question the page is named after, rewrite it so it does.
• Find every sentence carrying more than one claim and split it.
• Find every claim that cannot be checked. Replace it with a number, a named list or a date, or delete it.
• Add a section saying what you do not cover and who should buy something else.

That edit on one page is usually worth more than ten new pages written the old way.