The short version
- 44.2% of AI citations come from the first 30% of a page (GEO study, Princeton, Georgia Tech, Allen Institute for AI and IIT Delhi, KDD 2024). Write that block as the whole argument.
- A skim lands on headings, numbers, table rows and the pull quote. An audit of 200 essays here found one table across the entire library.
- Adding statistics lifts AI-answer visibility by 40%, adding quotations by 40% (GEO study, KDD 2024). Both are things a skimming reader stops on.
- Verify on the rendered page. Count the words, tables and links in the built output rather than trusting what the draft claims.
Write the first screen as though it is the only screen. 44.2% of AI citations come from the first 30% of a page (GEO study, Princeton, Georgia Tech, Allen Institute for AI and IIT Delhi, KDD 2024). Put the claim in the first 40 to 60 words, then spend the rest of the page proving it to whoever stays.
What does a skimming reader stop on?
Headings, numbers, table rows, bold text and the pull quote. Prose gets read after one of those has earned it.
Which makes the heading a promise the paragraph under it has to pay. A heading naming a topic gives a skimmer nothing to stop for, which is why question headings do more work than noun phrases.
Measured lifts in AI-answer visibility, and what each looks like to a reader (GEO study, KDD 2024):
| What you add | Measured lift | What a skimmer sees |
|---|---|---|
| A statistic | +40% | A number they can repeat to somebody else |
| A quotation | +40% | A named voice, pulled out of the column |
| A cited source | +30 to 40% | A publisher and a year next to the claim |
| Fluency optimisation | +15 to 30% | The same claim with the throat-clearing gone |
| Schema markup | 2.5x chance of appearing | Nothing. It sits in the head of the document |
How do you write a first screen that stands alone?
State the claim inside the first 40 to 60 words, without hedging. Every qualifier up there is a reason to keep skimming.
Then run the extraction test. Lift the opening paragraph out of the page and read it alone. If it needs the next three paragraphs to make sense, what you have written is an introduction, and introductions do not get lifted.
Why does a table do more work than a paragraph?
An audit of 200 published essays on this site found one table in the whole library, no links between essays and no structured data. A skimmer landing on any of them had nothing to stop on and nowhere to go next.
A table holds a comparison in a form that survives being read out of order, which a paragraph does not. It serves the reader who is scanning and the engine that is extracting without being rewritten for either.
What does a skimmable page look like when it is wrong?
A batch of twenty AI-written posts for this site came back looking right and measuring wrong: around 595 words against a bar of 1,000 to 1,400, no first-party data, no tables, no FAQs, no schema. Clean structure with nothing inside it skims beautifully and says nothing.
The second failure in that batch was quieter. After a repair pass an adversarial audit still found invented supporting detail, because the source material held one-line entries and the writers closed the gaps themselves.
A draft that says it has a table and a page that actually has one are indistinguishable until somebody counts the rendered output.
How do you check the skim works?
Measure the rendered page rather than the draft. Count words, tables, FAQs and internal links in the built output. On this site that is a script, and a post under the word count, or with fewer than five FAQs, or with no table, does not publish.
The client version is a QA gate on a newsletter: nine standing editorial rules plus a script that runs before an issue is queued, and eleven issues shipped through it. The rules were written at issue 20 rather than issue 1, because that is when the same notes started being given twice.
The mechanics either side of this one are measuring the output instead of the promise and writing for search without sounding like it.
Frequently asked questions
How do you write for readers who skim?
Put the claim in the first 40 to 60 words and make every heading state what sits under it. Add a table, because rows survive being read out of order. 44.2% of AI citations come from the first 30% of a page (GEO study, KDD 2024), and that block is what a skimming reader reads too.
Where should the most important content go on a page?
In the first 30% of it. 44.2% of AI citations pull from there (GEO study, KDD 2024). Anything saved for a conclusion is written for the small share of readers who reach one.
What is the right length for a blog post?
Length follows structure. Our bar is 1,000 to 1,400 words with a table, first-party numbers, internal links and five or six FAQs. A batch of twenty AI-written posts that came in around 595 words failed on missing tables and missing data before length was the problem.
Do tables help with SEO and AI search?
A table puts a comparison into rows that can be lifted whole, which prose cannot. An audit of a 200-post library found one table across the entire thing, and nothing else there was structured for lifting either.
How do you stop a page from being skimmable but empty?
Check the content, not the shape. A batch of twenty AI-written posts came back structurally clean with no first-party data and no tables. After the repair pass an adversarial audit still found invented supporting detail, because the source material held one-line entries and the writers filled the gaps themselves.
How do you verify a post meets the standard before publishing?
Measure the built page rather than the draft. Count rendered words, tables, FAQs and internal links in the output, and run it as a script so it cannot be skipped.
Sources
- Chua Network delivery data across 8 client accounts (internal fact bank)
- Chua Network engagement records, anonymized (internal experience bank)
- GEO study, Princeton, Georgia Tech, Allen Institute for AI and IIT Delhi, KDD 2024
