- AI engines work with passages, not whole pages. Write each section so it can be lifted out and still be correct and clear.
- Lead with the answer, then give the reason, the evidence and the limit. A reader who stops after two sentences should still leave with the right idea.
- Nobody can promise a citation. Clear writing removes reasons to be skipped. It does not force a mention.
What writing for AI search actually means
Writing content for AI search means writing so that a system that reads thousands of pages can find your answer fast, trust it and quote it correctly. It does not mean writing for robots. The test is simple: if a person reads only the first two sentences under a heading, do they leave with the right answer?
This matters because answer engines such as ChatGPT, Perplexity, Google AI Overviews and Microsoft Copilot build responses from retrieved passages. We do not know the exact weighting any engine uses, and they do not publish it. What we can observe is that a page which states its answer plainly, near the question, gives the system less work to do. Our post on how AI search engines choose sources covers what is known and what is not.
Google says in its guidance on AI features that the same helpful, reliable, people first content principles apply and that no special optimization is required (Google Search Central, checked October 2026). We agree with that as a starting point. Good writing is the base. What we add is a set of habits that make good writing easier to extract.
Start with the question a buyer would really type
Most weak pages answer a question nobody asked. They describe the company, then the features, then eventually the thing the buyer wanted. Reverse it. Begin from the prompt.
Write down the full sentence a buyer would give an AI tool, such as "which IT AMC provider in Dubai is suitable for a 50 person office?" Notice the details: the service, the place, the size and the choice. These details tell you what the page must cover. Our guide to researching buyer prompts shows how to collect them from sales calls, support tickets and search data.
Also remember that one question often becomes many. Engines may split a prompt into sub questions and search for each. That behavior is explained in query fan out explained. For writing, the lesson is to cover the neighbouring questions on the same page or a closely linked one: cost, timing, who it suits, who it does not suit, and how it compares to the alternative.
- Cost of an AMC
- How long it takes to start
- Who the service suits
- Who it does not suit
- How it compares to the alternative
The answer first pattern, step by step
- Heading as a question or a clear topic. Use the words a buyer would use, not a clever title.
- The answer in one to three sentences. Name the subject in the sentence. Do not start with "It" or "This" because the passage may be read without the heading.
- The reason. One short paragraph on why the answer is true or how it works.
- The evidence. A specific fact, number, example or source, with who and when.
- The limit. When the answer does not apply. This is what separates an honest page from a sales page.
- The next step. A link to the page that goes deeper, or the action the reader can take.
Here is the same content written two ways, as an illustration only, not as pricing advice. The first is a typical page opening: "We are a leading provider with years of experience, offering tailored solutions that fit your needs." It says nothing that can be checked. The second is answer first: "A standard annual maintenance contract for a 50 person office usually covers helpdesk support, preventive checks and hardware repair for a fixed yearly fee. The price depends on device count and response time." The second can be quoted. It is also easier for a human to use.
Make every passage stand on its own
Engines often take a passage from the middle of a page. If that passage depends on a pronoun, an earlier sentence or a diagram, it breaks when lifted out. So write with a few habits.
- Repeat the subject by name at the start of a section instead of saying "it" or "they".
- Keep one idea per paragraph, and keep paragraphs short.
- Define a term the first time it appears in a section, in plain words.
- Put units, currency, dates and places next to every number.
- Avoid "as mentioned above" and "see below". Link to the section or restate the point.
- Use lists for parallel items and tables for comparisons, with clear column labels.
Read each section aloud with the heading hidden. If a stranger could not tell what it is about, rewrite the first sentence.
Make each section survive being lifted out
- ✓Name the subject at the start of a section
- ✓One idea per paragraph
- ✓Define a term the first time it appears
- ✓Put units, currency and dates next to numbers
- Start with It or This
- Write as mentioned above
Use specifics, and show where they come from
Specific content is easier to verify, and verified content is safer to quote. Replace vague claims with named facts: what you do, for whom, in which place, at what scale, under which conditions.
Be careful with numbers. A number is useful only if the reader can see where it came from. Say who measured it and when, and link to the source if you can. If you cannot source a figure, write the point in words instead. Invented or rounded up statistics damage trust with readers and, as far as we can tell, make your page a worse source.
The 2023 GEO research paper by Aggarwal and colleagues, presented at KDD 2024, is relevant here. In their simulated engine, adding quotations, statistics and cited sources to content improved measured visibility compared with an unchanged baseline, while repeating keywords did not help. That supports the idea that evidence rich writing helps. But the test used a research benchmark and a model setup built by the authors, so it is not a promise for any live engine. We discuss it more in our post on original data.
Structure the page so the answer is easy to find
| Element | What to do | Why it helps |
|---|---|---|
| Title and H1 | Name the topic in plain words | Tells the system and the reader what the page covers |
| First paragraph | Give the main answer in two or three sentences | The most quotable passage on the page |
| H2 and H3 headings | Match real questions or clear subtopics | Lets a section be matched to a sub question |
| Lists and tables | Use for steps, options and comparisons | Compact, easy to extract and check |
| FAQ block | Real questions with 2 to 4 sentence answers | Covers neighbouring questions on one page |
| Last updated date | Visible and honest | Helps readers judge whether it is current |
| Author and review line | Name the person and their role | Gives context on who stands behind the claim |
If you add an FAQ, keep the questions visible on the page. The FAQ schema generator can produce matching markup, and the principle from schema markup for AI search applies: markup must describe what a reader can actually see.
Write for people first, and avoid the common traps
- Keyword stuffing. It reads badly, and in the GEO paper it did worse than the unchanged baseline.
- Hype and filler. Phrases with no checkable meaning take space and add nothing a system can use.
- Hiding the answer. Long introductions push the useful sentence down. Cut them.
- Copying competitor wording. You add nothing new, so there is no reason to cite you.
- Claiming what you cannot prove. Engines and readers both compare claims across sources. A claim that no one else supports stands out in the wrong way.
- Writing one page for everything. One page per main question is easier to match than one page that tries to answer ten.
How to check whether the writing works
Pick the five prompts that matter most. Before you change anything, ask them in each engine you care about, several times, and note whether your brand is mentioned, whether a page of yours is cited and whether the description is correct. Date the record. Then rewrite the matching pages and repeat the same prompts after a few weeks.
Expect noise. Answers differ by user, place and day, and an engine may not recrawl a page quickly. A change in one run proves little. A pattern across many runs says more. Our post on how to measure GEO performance goes deeper on method. Remember also that being named in an answer without a link is different from being cited. See brand mentions vs citations.
A before and after rewrite of a service page section
Example, hypothetical: a US cybersecurity consultancy has a section on its penetration testing page. The original reads: Our award winning team delivers best in class security solutions tailored to your business. Contact us today to learn more. Nothing in it can be checked or quoted, and an engine looking for an answer to what a penetration test includes would pass over it.
The rewrite applies the pattern. The heading becomes: What does a penetration test include? The first sentence answers: A penetration test is an authorized, time limited attempt to break into your systems, performed by a security professional, followed by a written report of what was found and how to fix it. Then comes a short paragraph on scope, one on what the client must provide, a plain statement of what is not covered, and a link to the scoping page. The rewrite promises no outcome, makes no claim it cannot support, and answers a real question in the first two sentences.
Notice what changed. The facts were already known to the consultancy. They simply had not been written down in a form that a reader, or a retrieval system, could use. Most content work for GEO is exactly this: surfacing what the business already knows.
Before
- Award winning team, best in class solutions
- Nothing in it can be checked
- An engine would pass over it
After
- Heading asks what a test includes
- The answer comes in the first two sentences
- Scope, client duties and limits follow
Page types and what each one needs to be cited
Different page types serve different prompts. If you write the same way for all of them, you leave gaps. Use the table to decide what each type must contain.
| Page type | Prompt it serves | Must contain | Common gap |
|---|---|---|---|
| Service page | Who offers X for Y? | Definition, who it suits, scope, process, price logic, proof, location | No plain definition, no scope limits |
| Comparison page | A versus B | Fair criteria, a table, a decision rule, honest weaknesses | One sided sales copy |
| Guide or explainer | How does X work? | Direct answer, steps, examples, limits, dates | Long introduction, no clear answer |
| Pricing page | How much does X cost? | A price, range or method, and what changes it | Contact us only |
| Case study | Does X work for companies like mine? | Context, actions, measured results with source and limits | Vague outcomes, no numbers or dates |
| FAQ block | Specific buyer questions | Real questions, 2 to 5 sentence standalone answers | Marketing answers that dodge the question |
| Location page | X near me or in a city | Real address or service area, hours, local facts | Duplicate text with only the city name changed |
If you want the strategic layer behind this table, read our page on GEO content strategy. For service businesses with several offers, a consistent structure across pages also helps the connection between pages, which we discuss in entity SEO for GEO.
Writing comparison and best of content honestly
Buyers ask engines to compare options and to name the best. Content that answers those questions fairly has a real job to do, and content that is pure promotion is easy to spot. The rules we follow are simple.
- State the criteria first. Say what you are comparing on, such as scope, response time, cost model or fit for company size.
- Be fair about alternatives. Describe what a competing approach does well. A page that admits trade offs is more believable.
- Say who you are not right for. This builds trust and helps the right buyer self select.
- Date the page and say what you checked. Products and prices change.
- Never make claims you cannot support, and never run down a named company without evidence.
A good decision rule beats a verdict. Write: choose option A if you need these conditions, choose option B if you need those. An engine can quote a rule like that, and a reader can apply it.
Editing checklist you can run on any page today
Use this checklist when you review a page before publishing. It takes about fifteen minutes per page.
- Does the page have one clear main question it answers, and does the title say so in plain words?
- Does the first paragraph answer that question in one to three sentences?
- Is every H2 a question or a clear topic, in the words a buyer uses?
- Does each section begin with its answer, with the subject named in the first sentence?
- Is every number sourced, dated and attached to a unit?
- Is there at least one honest limit or exception in each major section?
- Are lists and tables used where items are parallel or compared?
- Are the visible FAQ questions real ones from buyers, with standalone answers?
- Do internal links point to the next logical page, with descriptive anchor text?
- Does everything on the page match your other public profiles, such as name, address and services?
- Is the page readable without scripts, and can crawlers fetch it? Check with the AI crawler access checker.
Do formatting choices really change what gets cited?
Writers hear many rules: short paragraphs, bullet lists, specific word counts. The honest position is that structure helps readers and probably helps extraction, but no one can prove a specific format causes citations. Engines read text, so a clear sentence is the key unit. A table helps when the content is a comparison. A list helps when the items are parallel. Forcing everything into bullets can strip the reasoning out of a page and make it less useful.
What we do recommend, because it helps people and machines alike, is to keep answers close to their questions, keep one idea per paragraph, and keep important facts in the page text itself rather than inside images, tabs that do not load or scripts that render late. If your site relies on heavy scripting, see GEO for JavaScript and headless sites. Freshness is another variable that depends on the topic, covered in content freshness for AI search.
Using AI to draft content without losing quality
Many teams draft with AI tools. That is not a problem by itself. Google's guidance, as we read it in October 2026, focuses on whether content is helpful and original, not on how it was produced. The risk is volume without substance: pages that restate what is already online, contain no firsthand detail and add no reason to be cited.
- Use AI for structure and first drafts, never for facts. Every claim needs a human check against a real source.
- Add what only you have: your process, your data, your photos, your pricing logic, your customer questions.
- Have a subject expert read each page and correct it.
- Do not publish dozens of near identical pages. Fewer, stronger pages are better.
- Keep a changelog with the date and the reviewer, so you can show how a page was checked.
If you want original evidence to add to your pages, original data and statistics for AI citations shows how to produce it honestly.
What to do next
Take one important page and apply the pattern: rewrite the headings as questions, put a direct answer under each, add one sourced fact per section and state one limit. Then link it from related pages. For broader planning, see our GEO content strategy page.
In the GEO Pilot we do this work for you. In week 3 we rewrite and restructure up to 10 priority pages to be answer ready, with direct answers, headings that match real questions, FAQs and schema. The Pilot is $999 fixed, one time, delivered inside 30 days, and we retest the same prompts at the end. We cannot promise any particular result.




