Skip to content
Learn

Writing for AI answers

In this lesson

Section titled “In this lesson”
  • Structure a section so that a system can lift a clean answer from it.
  • Write direct answers, TL;DRs and data tables that help readers and machines.
  • Compare what Google, Microsoft and published research say about writing for AI answers.
  • Explain the two meanings of “citation sources” and “preferred sources”.
  • Use Google’s Preferred Sources feature for your own site.

On day 07 you saw the main rule: you cannot be cited if you are not retrieved. This lesson covers the next step. After retrieval, the AI system reads the passages and picks the ones that it uses in the answer. Clear writing makes your passage easier to pick.

Two views on writing for AI

Section titled “Two views on writing for AI”

Two primary sources give advice on this topic, and their tone is different.

Google says that you do not need to rewrite content “just for AI systems”. Its AI optimization guide (July 2026) lists “chunking” and special AI files as things that you “can ignore”. It asks for one thing in structure: organize content “by paragraphs and sections, along with headings that provide a clear structure to navigate content”. It asks for this because it helps human readers.

Microsoft gives more detailed advice. In October 2025, a Bing team member wrote on the Microsoft Advertising blog that assistants such as Copilot “break content down, a process called parsing, into smaller, structured pieces”. Those pieces “are then assembled into answers, often drawing from multiple sources”. The post recommends clear headings, question-and-answer pairs, lists, tables and short, self-contained sentences.

These views do not conflict much. Both say that clear sections with good headings help. Google says that you do not need special tricks for its systems. Microsoft says what its systems find easy to use. Where the two disagree, follow the advice of the product that you target. For most pages, one clear structure works for both, and for human readers too.

The roadmap.sh item calls this structuring for AI extraction: formatting content so that “an AI system can easily pull out a clean, direct answer”. Extraction means that a system lifts a short passage from your page and uses it in an answer. The main tools are direct answers, TL;DRs and data tables.

Structuring for AI extraction

Section titled “Structuring for AI extraction”

Microsoft’s post gives a useful list. Each item has the same idea: make the start and the end of each idea clear.

Do Avoid
Headings that say what the section answers, for example “What makes this dishwasher quieter than most models?” Vague headings, such as “Learn more”
One idea in each section “Long walls of text” that “blur ideas together”
Key facts in the HTML text Key facts “only in images” or only in PDFs
Important answers visible on the page Answers in tabs or expandable menus, because “AI systems may not render hidden content”
Specific claims with numbers, such as “42 dB” Unanchored words, such as “next-gen” or “cutting-edge”
Simple punctuation Decorative arrows, stars and strings of “!!!”

Some of these points repeat earlier days. Clear headings came on day 04b. Rendering problems with hidden or script-loaded content came on day 03b. The rule about tabs is Microsoft’s advice. Google’s guide does not mention tabs.

What matters most is a self-contained passage: a passage that makes sense when someone reads it without the rest of the page. Microsoft calls this “self-contained phrasing: sentences that make sense even when pulled out of context”. Day 07 gave the same test from Pinecone for chunks. A sentence such as “It is the quietest” fails the test, because the reader does not know what “it” is.

Semantic HTML helps here too. Google says that you do not need “perfectly semantic HTML”, but that semantic HTML is “generally a good idea”, because it helps screen readers. Real headings, lists and tables mark where each piece starts and ends.

Direct answers

Section titled “Direct answers”

A direct answer is a plain reply to one question, in one or two sentences, placed close to the question. The roadmap.sh item says that this content “is more likely to be pulled out” by featured snippets, AI Overviews and other answer systems.

This is an old idea. On day 04 you met featured snippets, which show a passage from a page above the normal result. Google says that you cannot mark a page as a featured snippet: its systems decide. But a passage can only be shown if it exists on the page. Microsoft calls these short passages “snippable” content, and it lists what makes content eligible:

  • “Concise answers: one- to two-sentence responses that directly address a question.”
  • “Structured formatting: lists, tables, and Q&A blocks that can be lifted cleanly.”
  • “Strong headings” that show where an idea starts and ends.
  • “Self-contained phrasing.”

How do you write a direct answer? Use this pattern:

  1. Put the question, or the topic, in the heading.
  2. Write the answer in the first sentence under the heading.
  3. Name the subject in that sentence. Do not start with “it” or “this”.
  4. Give the specific fact: the number, the unit, the date or the name.
  5. Add the details, the reasons and the exceptions after the answer.

Journalists call this order the inverted pyramid: the most important fact comes first. Microsoft’s example is a pair: “Q: How loud is the dishwasher? A: It operates at 42 dB, which is quieter than most dishwashers on the market.”

Edit the section

Toy model with example text. The system lifts the heading and the first two sentences of the section. Tick the edits one at a time. Notice that the passage answers the question alone only when the answer comes first, names the subject and gives the number.

A direct answer is not a short page. Google says that there is “no ideal page length”. The direct answer opens a section. The rest of the section gives the depth that makes the page worth a visit.

TL;DRs

Section titled “TL;DRs”

A TL;DR (“too long; didn’t read”) is a short summary at the top of a piece of content. The roadmap.sh item says that it states the main point “before the reader, or an AI system, reads the full explanation”.

A good TL;DR has two to four sentences or bullets. It gives the conclusions, not a list of the topics. Compare:

  • Weak: “This article covers dishwasher noise, ratings and models.”
  • Strong: “Dishwashers below 45 dB are quiet enough for an open kitchen. The Model Q runs at 42 dB. Noise ratings are not standard across brands, so compare them with care.”

The strong version is a set of direct answers. Each sentence can stand alone.

Is a TL;DR confirmed to help AI citations? No primary source says so. Google does not mention TL;DRs, and it says that you do not need to write in a special way for AI. Microsoft recommends concise answers, but not a summary box. So use a TL;DR when it helps your readers, for example on long guides and reports. Do not add one only for AI systems.

If you do not want a passage to be extracted, use Google’s snippet controls. Google says that nosnippet, data-nosnippet and max-snippet limit what it shows from your page in Search, and these controls include its AI features.

Data tables

Section titled “Data tables”

A data table shows information in rows and columns, not in paragraphs. The roadmap.sh item says that tables help “AI systems extracting specific facts”, because “the structure clearly separates individual data points”. Microsoft recommends “comparison tables” for “how-to queries and feature comparisons”.

Use a table for tabular data: data with the same attributes for several items, such as prices, specifications or results. MDN says that tables are for tabular data, not for page layout. Build tables the correct way:

  • Use real HTML <table> markup, not an image of a table.
  • Use <th> for header cells, so that each value has a label.
  • Add a <caption> that says what the table shows and where the data comes from.
  • Put the unit in the header, for example “Noise (dB)”, and the date of the data.
  • Keep one fact in each cell.

MDN says that screen readers “handle tables well” when the markup is correct. A table that a screen reader can read is also a table that a machine can parse. Google’s guide does not mention tables, so their effect on Google’s AI features is not confirmed. Tables also work well as original data, which you met on day 05.

Citation resources

Section titled “Citation resources”

The roadmap.sh topic file for this item is titled “Citation Sources”. The phrase has two meanings in AI search, and both matter.

Meaning 1: the sources that your page cites

Section titled “Meaning 1: the sources that your page cites”

The best research on content edits is the GEO paper by Aggarwal and others (KDD 2024). The authors built a test generative engine: it took the top five Google results and wrote answers with GPT-3.5. Then they edited one source page with different methods and measured how visible that page was in the answer.

The results:

  • “Cite Sources”, “Quotation Addition” and “Statistics Addition” worked best. They gave a relative improvement of 30% to 40% on the main visibility metric.
  • Keyword stuffing gave “little to no improvement”.
  • An authoritative, persuasive tone gave no significant improvement overall.
  • The gains were largest for pages at lower search ranks. Pages at rank 1 often lost visibility.

The authors also tested Perplexity and found “visibility improvements up to 37%”. Remember the limits. This was a lab test from 2023 and 2024, with an older model. AI products change fast, and no company has confirmed these effects.

Edit method

Data from Table 2 of the GEO paper (KDD 2024). Choose an edit method, then move the slider to your page's search rank. Notice that citing sources, quotations and statistics help pages at rank 5 most, and can hurt the page at rank 1.

Why can a top page lose? The paper does not test the cause. One possible reason: an answer has limited space, so when lower pages become more useful, the top page gets a smaller share.

The practical lesson agrees with Google’s advice on non-commodity content. Facts with a source, real quotes and specific numbers make a page more useful. They also make the claims easy to check.

Meaning 2: the sources that AI systems cite

Section titled “Meaning 2: the sources that AI systems cite”

The roadmap.sh file uses the second meaning: the websites that AI systems “most frequently pull from and cite”. Knowing these sites shows you where your brand needs a real presence.

In March 2026, Peec AI, an AI visibility tool, analyzed 30 million cited sources in the US. These were the top five cited domains on each platform:

Platform Top five cited domains (Peec AI, US, March 2026)
Google AI Mode YouTube, Reddit, Facebook, LinkedIn, Yelp
AI Overviews YouTube, Reddit, Facebook, LinkedIn, Medium
Gemini Reddit, YouTube, Wikipedia, Medium, PCMag or Forbes
Perplexity Reddit, YouTube, LinkedIn, Wikipedia, G2
ChatGPT Wikipedia, Reddit, Forbes, Techradar, LinkedIn

Reddit and YouTube are in the top five of all platforms. The lists also change by industry: Peec says that in hotels, Booking.com and Airbnb “often outrank Reddit”. This is a vendor study, and other studies rank the sites in a different order. The pattern is stable, though: community sites, video and reference sites get many citations. Day 06b showed the Semrush study with a similar result, and it showed how to take part in communities honestly.

Preferred sources

Section titled “Preferred sources”

This term also has two meanings.

The roadmap.sh meaning is the sites that a given AI system “consistently favors” for a topic. You find these with the method from meaning 2 above: ask the system the questions of your audience, and record the cited domains. Then focus content and PR work on those sites. Day 08 covers tools that measure this at scale.

Google’s meaning is a product feature that the user controls. With Preferred Sources, a user picks the sites that they want to see more of. The history:

  • In August 2025, it started in Top Stories, in English, in the US and India. Google’s own launch post was not in the search results, so this date comes from news reports.
  • On 16 December 2025, it became available in English to users worldwide.
  • In 2026, it reached all languages where Google Search is available.
  • In 2026, Google also brought it to AI Overviews and AI Mode, where a chosen site can get a “preferred” badge.

Google says that users had chosen “nearly 90,000 unique sources” by December 2025. When someone picks a preferred source, “they click to that site twice as much on average”.

Can you use this for your site? Yes, with three rules from Google’s docs:

  1. Only domains and subdomains are eligible, such as example.com or code.example.com. A folder such as example.com/blog is not.
  2. For the AI features, your site must be included in generative AI features in Search Console.
  3. Check if your site appears in the source preferences tool.

Then ask your readers to choose you. Google offers a button that needs two lines of HTML, or a deep link in this format: https://www.google.com/preferences/source?q=example.com. You can use the link in newsletters and social posts. This works only for readers who already know you, so it builds on brand work from day 06b.

What matters most

Section titled “What matters most”
Priority Practice Why
1 Self-contained sections with clear headings Both Google and Microsoft support this, and it helps readers
2 Direct answers first, with specific facts A short passage can be lifted only if it holds the answer
3 Sources, quotes and numbers The strongest effect in the GEO research, and it builds trust
4 Presence on the sites that AI cites Many citations go to community, video and reference sites
5 TL;DRs, tables and Preferred Sources Useful where they fit the content and the audience

Do not write for machines at the cost of readers. Every practice in this lesson also makes a page easier for a person to use. That is the test to apply.

Key terms

Section titled “Key terms”
  • Extraction: a system lifts a short passage from a page and uses it in an answer.
  • Structuring for AI extraction: formatting content so that a system can lift a clean answer from it.
  • Self-contained passage: a passage that makes sense without the rest of the page.
  • Direct answer: a plain reply to one question, in one or two sentences, close to the question.
  • TL;DR: a short summary of the conclusions at the top of a piece of content.
  • Data table: information in rows and columns, with labelled headers.
  • Tabular data: data with the same attributes for several items.
  • Citation sources: the sources that a page cites, or the sites that AI systems cite most often.
  • Preferred sources: the sites that an AI system favors for a topic; also a Google feature where users choose sites to see more of.

Exercise

Section titled “Exercise”

Exercise · about 15 min

Rewrite one section for extraction

Pick a page from your own site, or a product or help page from a well-known site. Use Perplexity (free) for step 5.

  1. Find one section that answers a question that a reader may ask. Write the question down.
  2. Read only the heading and the first two sentences. Do they answer the question alone? Write yes or no.
  3. Rewrite the section with the five-step pattern: question heading, answer first, named subject, specific fact, details after.
  4. If the section compares several items, turn the facts into a table with <th> headers, units and a caption.
  5. Ask Perplexity the question from step 1. Write down the domains of the cited sources. Mark each one as community, video, reference, brand or publisher.
  6. Optional: enter your domain in Google’s source preferences tool at https://www.google.com/preferences/source. Does it appear?
Show a possible solution

A possible result for a dishwasher help page:

  • Question: “How loud is the Model Q dishwasher?”
  • Step 2: no. The section opened with “Our engineers spent years on this design”, and the noise level came in the fourth sentence.
  • Step 3, new heading: “How loud is the Model Q dishwasher?” New first sentence: “The Model Q dishwasher runs at 42 dB, which is quieter than most dishwashers.” The design story moves below it.
  • Step 4: a table with the columns “Model”, “Noise (dB)” and “Energy rating”, and the caption “Noise levels of our 2026 models, from lab tests”.
  • Step 5: expect a mix. Product questions often cite brand pages, review sites, Reddit and YouTube. Your results depend on the question, the date and your location.

If your domain is not in the source preferences tool, it may be too new or it may not publish content that Google shows in Top Stories. A folder of a site is never eligible.

Check yourself

Section titled “Check yourself”
A section starts with 'It is our best seller. Customers love it.' The price comes in the fifth sentence. Why is this section hard to use in an AI answer?
A GEO agency says: 'Google needs you to split every page into tiny chunks.' Microsoft says that Copilot parses pages into smaller pieces. What should you do?
In the GEO paper, which edit gave the largest visibility gain for a page at search rank 5?
A blog lives at example.com/blog. The owner wants readers to add it as a Google Preferred Source. What is the problem?
A software company finds that Perplexity often cites G2 and Reddit for its product category. How should it use this?

Sources

Section titled “Sources”

Next

Section titled “Next”

Day 07c covers AI agents and AI crawlers: how agents browse, MCP and structured access, and how to manage GPTBot, ClaudeBot, Google-Extended and PerplexityBot.