What questions should content answer for GEO?

Five, and they run from the product outward to the decision. What makes your product special, what sets your brand apart, what customer need you are fulfilling, what is keeping a new customer from shopping with you rather than a competitor, and what is keeping an existing customer loyal. The first three describe the brand, and most companies already publish some version of them. The last two describe the decision a person is making, and they answer the comparative and objection-shaped prompts people bring to a model. A catalog that covers the first three and skips the last two gets read and rarely reached for. The questions being asked are about choosing, not about describing.

How is writing content for AI different from writing for SEO?

The unit changes. Search retrieves pages and a person decides what to read, so an SEO page can build an argument across several sections and still work. A model retrieves passages and assembles an answer, so any sentence you want repeated has to stand alone once it is separated from everything around it. In practice that means naming the subject inside the sentence rather than relying on the heading, stating one claim per passage, and putting the qualifier in the same sentence as the claim. The keyword research, the internal linking and the technical foundations still matter. What is new is that a page can rank perfectly well and still contain nothing a model can quote.

Why does my content get crawled but never cited?

A crawl and a citation test different things. A crawl means the page was reachable and the content was readable, so robots rules, rendering and response time have already cleared. A citation means the model found a passage it was willing to stand behind while answering a real question. The usual gap between the two is framing: the page describes rather than claims, hedges the sentence that mattered, or spreads its answer across three paragraphs so no single passage carries it. That makes crawled but not cited the most useful diagnostic a brand can watch. It has removed every explanation except the writing.

How do you write a passage a model can quote?

Make it survive removal. Take the sentence out of the page, put it in front of someone who has never seen your site, and check that it still says something specific and true. That takes three things in one sentence: the subject named explicitly rather than as it or the product, a single claim rather than a list of attributes, and the qualifier that bounds the claim. A sentence like it is designed for demanding conditions carries nothing once it is lifted. A sentence naming the product, the condition and the limit carries the whole claim, and a model repeats it without inventing the missing half.

Should a brand write content about its competitors?

Yes, and the reluctance to do it is the reason the gap exists. People ask models comparative questions constantly, and a model answering one assembles it from whatever sources describe both options. Publish nothing on where you fit and the comparison still gets made, entirely from other people's descriptions of you. Writing it yourself does not mean attacking anyone. It means stating plainly who your product is for, who it is not for, and on what basis someone would reasonably choose the alternative. Naming the case where you are the wrong answer makes the rest of the passage credible enough to repeat.

Framing content for AI training: the five questions a brand has to answer.

GEO content fails at the framing stage, before a word is written. It starts from a keyword list instead of a question. Content written purely for marketing fails the same way, and so does content written purely for a model. WISLR frames for the customer. These are the five questions a brand has to answer for a model to have anything to say about it, how to turn each one into a passage worth quoting, and how to tell from server logs whether the framing worked.

Available in:

Deutsch Español Français हिन्दी Italiano Bahasa Melayu Nederlands Polski Português Türkçe Українська

Author: Tony Castillo

Published August 9, 2026

← All articles How we research

A stop-motion clay notepad holding five modeled question marks, three in green and two in gold, with a clay pencil resting across it

The short version

Framing is a separate step, and it comes before writing. A keyword tells you what a page is about. A question tells you what claim the page has to make, who it has to name, and what has to be true for a model to repeat it. Most GEO content is briefed from the first and then judged by the second.
Five questions cover the whole decision. What makes the product special, what sets the brand apart, what need is being met, what stops a new customer from choosing you, and what keeps an existing one. The first three describe you. The last two describe a decision somebody is making about you, and they are the two almost nobody publishes.
A model quotes passages, not pages. Whatever you write has to survive being lifted off the page and dropped into an answer with no surrounding context. That means naming the subject in the sentence, stating the claim plainly, and carrying the qualifier that makes it true.
The logs tell you whether the framing worked. A page that gets crawled and never fetched into an answer has already cleared access and retrieval. What is left is the writing. Crawled but not cited is the closest thing to a direct score on your framing.

Start With the Question, Not the Keyword

Most GEO content is briefed from a keyword list, and that is the wrong input. A model is not matching a string against a page. It is assembling an answer to a question somebody asked in their own words, from passages it can stand behind.

A keyword tells you what a page is about. A question tells you three harder things: what claim the page has to make, who it has to name, and what has to be true for a model to repeat it without hedging.

Two habits produce the same failure. Content written purely for marketing says what the brand wants said. Content written purely for a model says what a model is assumed to reward. Neither one answers the question a person asked. Write for the customer. The model is working on that same question.

The failure looks the same every time. A brand publishes forty pages, every one of them gets crawled, and almost none get reached for during a real conversation. The pages are readable. They are indexed. They contain no answer to anything anybody asked.

Framing is a separate step, and it comes first. Decide what your brand is claiming before deciding what to publish.

The Five Questions

Five questions cover everything a model needs in order to recommend a brand. They run outward, from the product to the company to the person choosing.

What makes your product special?

The attribute-level claim. What it does, what it is made of, what it does better and by how much.

What sets your brand apart?

The company-level claim. Who you are, how you operate, what you refuse to do that others will.

What customer need are you fulfilling?

The problem in the customer's language, not the category name you file yourself under.

What is keeping a new customer from shopping with you rather than a competitor?

The objection. The reason someone comparing you to an alternative picks the alternative.

What is keeping an existing customer loyal?

The reason for the second purchase, which is rarely the reason for the first.

The first three describe you, and most brands already publish some version of them. The last two describe a decision somebody is making about you, and almost nobody publishes those at all. That asymmetry is the single most common finding when we audit a catalog for AI visibility.

Why These Five and Not a Keyword List

People do not talk to a model in keywords. They describe a situation and ask for a recommendation, and the shape of those prompts is consistent enough to plan against. Every one of them resolves to one of the five.

How people actually askWhat they are really askingQuestion

"Best waterproof boot for a wet commute"Which product has the attribute I need?01

"Is this brand any good, or is it just marketing?"Can I trust the company behind it?02

"My feet get cold walking the dog at 6am"What solves the thing I am dealing with?03

"Brand A or Brand B for a narrow foot?"On what basis do I choose between them?04

"Are they worth it long term?"What happens after I buy?05

Questions four and five decide the purchase, and they are the two the catalog cannot answer. A model asked a comparative question still produces a comparison. It builds that comparison out of reviews, forums, and whatever your competitors wrote about the category. Those are the only sources that engaged with the question.

Declining to write about the competitive decision does not remove your brand from the comparison. It removes your voice from it.

Working Through Each Question

For each question below: what a model needs to find, what most brands publish instead, and the shape of the passage that fixes it.

What makes your product special

What a model needs: A specific, bounded attribute claim it can repeat without adding anything. Named subject, one property, and the number or condition that makes it checkable.
What gets published: Adjective stacks. Premium construction, thoughtfully designed, built to last. None of it can be repeated. None of it says anything a model could be wrong about.
What to write: Replace each adjective with the fact underneath it. "Built to last" becomes the material, the tested cycle count, and the length of the warranty that backs it.
The test: Put a competitor's name on the sentence. A sentence that stays true was never a claim.

What sets your brand apart

What a model needs: Facts about the company that hold across the whole catalog: how long you have operated, how you make or source, what you will not do, who stands behind the product when it fails.
What gets published: An About page written as autobiography. Founding story, values, a photograph of the team. It reads well and contains almost no repeatable fact.
What to write: The operating facts, stated plainly and dated. A model recommending a brand needs a reason it can attribute, and "family owned since 1974, all repairs done in house" is a reason. "We care about quality" is not.
The test: Remove your logo from the page. A reader who cannot name the company from what is left has read nothing a model can use either.

What customer need you are fulfilling

What a model needs: The problem described the way the customer describes it, sitting in the same passage as the product that solves it. This is the bridge between a situation somebody types and a thing you sell.
What gets published: Category language. Brands write "technical outerwear" while customers write "I get soaked walking in from the parking lot." Those two phrases never meet on the page, and the model never connects them.
What to write: Take the ten situations your support inbox and reviews actually describe. Write each one in the customer's words, then answer it in the same passage. Support tickets are the highest-value content brief most brands already own and never read.
The test: Search your own site for the phrase a customer would use. Zero results means you have written about the category and never about the need.

What is keeping a new customer from choosing you

What a model needs: An honest account of where you fit against the alternatives, including the cases where you are the wrong answer. Comparative prompts are among the most common purchase questions people bring to a model.
What gets published: Nothing, usually. Where a comparison does exist it is a table where the brand wins every row, which a model will treat as marketing and weight accordingly.
What to write: Name the objection and answer it in the same passage. Price, switching cost, sizing risk, delivery time, whatever the real one is. Then name the customer you are genuinely not right for.
The test: Find the concession. A comparison that concedes nothing carries no information, and it reads as promotional to a model exactly as it does to a person.

What is keeping an existing customer loyal

What a model needs: Evidence about what happens after the purchase: durability over time, service, repairs, replacement, how the second and third order differ from the first.
What gets published: Acquisition content only. The catalog is written entirely for someone who has never bought, and the entire post-purchase relationship lives in a help center excluded from crawling.
What to write: The long-run facts. Reorder rates, repair policy, how the product ages, what owners say after two years. "Worth it" is a question about year three, and it is asked constantly.
The test: Check whether your help center and policy pages are crawlable. This is the one failure on the list that is technical rather than editorial, and it is common.

The Passage Has to Survive Being Lifted

Search retrieves pages, and a person decides what to read once they arrive. A model retrieves passages and assembles them, so every sentence you want repeated has to work with everything around it removed.

That constraint separates content that gets cited from content that does not. Three things have to sit inside the sentence rather than around it.

Name the subjectNot "it" or "this product." The heading does not travel with the passage, so the sentence has to carry the name itself.

One claim per passageA sentence carrying four attributes gets quoted for none of them. A model reaching for a specific answer needs the passage to be about that answer.

Qualifier in the same sentenceThe condition that makes the claim true has to sit beside it. Split across two sentences, the qualifier gets left behind and the claim becomes one a model will not repeat.

One product line, written both ways.

Cannot be lifted

It is designed for demanding conditions and built with premium materials for lasting performance.

No subject, no measurable claim, nothing a model could be held to.

Can be lifted

The Harrow parka uses a 20,000mm waterproof membrane and is rated for sustained rain down to minus 10 Celsius, though it is too warm for anything above 12 Celsius.

Named, specific, bounded, and it concedes a limit.

Turning the Five Into a Content Map

The five questions do not each become one page. They distribute across page types you already own. This is a framing exercise before it is a production one.

Page typeQuestions it should carryWhat usually has to change

Product detail01, 03Replace adjective stacks with bounded attribute claims, and add the situation the product is for.

Category and collection03, 04Add selection guidance. Most collection pages are a grid with a paragraph of category language above it.

Comparison and alternatives04Usually does not exist. This is the highest-value gap in most catalogs.

About and company02Convert the founding story into dated operating facts.

Help, policy and support05Make it crawlable first, then write the durability and service facts into it.

Reviews and owner content05Surface long-term owner language in server-rendered HTML rather than a JavaScript widget.

Two of those rows are the ones that move the number. Comparison content, because it answers the question that decides the purchase and almost nobody has written it. Help and policy content, because it answers question five and is frequently blocked from crawling by a robots rule nobody has reviewed in years.

Before you write anything

Check which of your existing pages a model has already read and never used. That list is a framing audit you do not have to commission, and it is sitting in your server logs right now.

The seven metrics that describe the channel →

Knowing Whether the Framing Worked

One measurement judges framing. A page that AI crawlers have read and never fetched into an answer has cleared every other explanation. It was reachable, so robots rules and rendering are fine. It was read, so the model has the content. What is left is the writing.

Crawled but not cited is the closest thing to a direct score on framing. Watched over a few weeks after a rewrite, it settles the question content work usually cannot: whether the change was an improvement or an opinion.

Fix access first. A page a crawler cannot reach will never be cited no matter how well it is written. Once access clears and pages are being read, every remaining gap is a framing gap, and the five questions close it.

It's Now Safer to Put Cloudflare in Front of Shopify: What O2O Fixed, and What the Warnings Still Mean