Writing passages AI answers can quote

What does an engine lift from a real page?

Whole sentences, and often old ones. Here is what Perplexity said about us when asked "what is Attensira" on 2 and 3 October 2026. Attensira takes three samples per daily reading of that question, issued together (n=3); these six answers were read over the Attensira MCP on 4 October.

Every one of the six named us. Read the descriptions closely, though, and you can see the pages they came from.

• One answer called Attensira "an agent that ships the fix", not just a dashboard, citing our about page. That line is an older tagline of ours; it was still in the about page's title when Attensira's page reader last read it, on 1 October 2026.
• Several described our audience as technical teams "whose site or docs live in a Git repo", citing our homepage. That is an earlier positioning of ours, and not how we describe the product now: an AI-native SEO agency run by software, which does the work across your site, content and third-party pages.
• One said Attensira "does not automatically merge changes", citing the about page. True, and lifted almost word for word.
• Two of the six opened by saying "Attensira" could also be the name of an ADHD medication, and cited medical sites to say so.

Nothing here is a hallucination. The engine did exactly what it is built to do: find pages about the subject, pick out self-contained sentences that answer the question, and stitch them together with links. Our sentences survived extraction. Some of them were simply the wrong sentences to have left lying around.

That is the whole lesson of this chapter in one example. Engines quote passages, not pages. The passages that get quoted are the ones that name their subject, state something plainly and make sense alone. Write those deliberately, keep them current, and remove the ones you no longer stand behind.

Why do engines work with passages rather than pages?

Because an answer is built from evidence for specific claims, and a claim lives in a sentence.

When an engine answers a question, it retrieves a set of pages, often by running several related searches, and then selects spans of text from them that support what it is about to say. It attaches some of those pages as sources. Attensira records, for every citation, the sentence of the answer that carries the link. Looking at those sentences shows the pattern: each cited page is propping up one particular claim, rarely the page's overall argument.

A page that ranks well as a whole can still contribute nothing if none of its sentences works on its own. The ranking chapter covers that gap from the retrieval side. This chapter is the writing side.

What makes a passage survive?

Five properties. Most pages fail on the first two.

1. It names its subject. "It integrates with Xero" is useless once lifted, because "it" is gone. "Ledgerly integrates with Xero" carries your name into the answer. Our "an agent that ships the fix" line survived because the answer could put "Attensira is" in front of it, and the page made that easy.

2. It answers first. The first sentence of a section should be the answer to the section's question. Context, caveats and history come after. An engine that takes one sentence should take the one that answers.

3. It carries a specific. A number with its unit and date, a named method, a named comparison. "Fast setup" gives an engine nothing to attribute. "Setup takes one Xero connection and an import of your client list" does.

4. It stays on one idea. A paragraph that answers two questions is quoted for neither, or quoted in a way that blurs both.

5. It stands outside its layout. A sentence that depends on a heading above it, a tab you have to click, or an image beside it loses that support when lifted. Google asks site owners to make sure important content is available as text for the same reason.

What does the difference look like on a real paragraph?

Here is one passage written two ways, about a fictional invoicing tool.

Hard to lift:

Built for modern teams, our platform makes getting paid simple. With powerful automations and deep integrations, it's the last tool you'll need. Get started today and see why teams love it.

No name, no specific, no answer to any question a buyer asks. An engine asked "what is the best invoicing tool for small agencies" has nothing to take.

Easy to lift:

Ledgerly is an invoicing tool for agencies of 5 to 50 people. It sends invoices from tracked time, chases late payments by email on a schedule you set, and syncs paid invoices to Xero and QuickBooks. It does not handle payroll.

Each sentence stands alone and carries the name. It says who it is for, what it does, what it connects to, and what it does not do. An engine answering a shortlist question, a comparison question or a "does it do payroll" question can take a sentence from it and cite it correctly.

How should a whole page be structured?

As a stack of passages, each answering one question, in the order a buyer would ask them.

1. A lead answer of 40 to 60 words directly under the title, answering the page's main question completely. Every page in this guide opens this way.
2. Question-shaped headings for each section, such as "How much does it cost?" or "Who is it for?". They mirror the sub-questions an engine searches for and tell the reader what the section answers.
3. The answer as the first sentence of each section, then the explanation.
4. Specifics in text, not only in images, tabs or interactive widgets.
5. A short FAQ for the questions that do not deserve a section but do get asked.

This is not a template for its own sake. Each element exists because it produces a sentence that makes sense when everything around it is gone.

Do statistics, quotations and sources really help?

The best evidence we have says yes, with caveats.

The GEO study, the first systematic test of content changes against a generative engine, tried nine kinds of rewrite. The top three, adding citations from credible sources, adding quotations and adding statistics, raised its main visibility measure by 30 to 40% relative to the original page. Making the tone more authoritative and persuasive produced no significant improvement; keyword stuffing produced little to none. The best method also varied by topic: statistics helped most on law, government and opinion topics, quotations on people, society and history.

Two caveats. The study measured its own test engine, not today's ChatGPT or Perplexity, so treat the direction as useful and the numbers as illustrative. And adding a statistic only helps if it is true, sourced and dated. A made-up number is a passage you will one day have to explain.

What about the passages you did not write?

They count too, and they are where our example went wrong.

The old tagline on our about page was ours, and we can fix it. The ADHD medication result is not ours at all: there are pages using a similar name for a medicine, and an engine asked "what is Attensira" reasonably mentions them. We cannot remove those pages. What we can do is make the answer about us so clear and so consistently stated, on our site and on the third-party pages that describe us, that the engine has an unambiguous passage to lead with.

Three practical steps:

1. Audit what engines say about you, sentence by sentence. Store the answers to your branded questions and highlight every phrase that came from a page. Attensira keeps every answer verbatim and the sentence each link supported, which makes this a reading exercise rather than a guessing one.
2. Fix the source passage first. Change the sentence on your own page, then the copies of it on directories, review sites and profiles you control.
3. Re-read after the pages are re-fetched. An engine keeps quoting the old sentence until it has read the new one. Expect weeks, not days.

Which pages need this treatment first?

Not every page. Start where a lifted sentence matters most.

1. The pages engines already cite. If an engine is reading a page, the passages on it are being quoted now. Our pricing page and about page were both in that group.
2. The pages that answer your branded questions: about, pricing, product overview, and the comparison pages for your main competitors. Answers to "what is" and "X vs Y" lean on them.
3. The pages written for your most valuable category questions, where you want to be named and are not yet.

Leave blog archives and old announcements for last, but read them once: an outdated sentence there can be the one an engine finds.

What should you check on each important page?

A short checklist, one pass per page:

• Does the first paragraph answer the page's main question in 40 to 60 words, using your name?
• Does every section heading ask a question a buyer asks?
• Does every section's first sentence answer its heading?
• Is every pronoun that matters replaced by the name it stands for?
• Does every claim that matters carry a specific: a number with a date, a named integration, a named limit?
• Is any important fact only in an image, a tab or a script?
• Is there a sentence on the page you would not want quoted back to a buyer? Remove it.

The next chapter, schema and technical access, covers making sure engines can read these passages at all.