Articles / Getting found

How to write a page a machine will quote

Nielsen Norman Group watched test users open new web pages and found that 79% of them scanned the page while only 16% read it word by word. A chatbot that answers a question does the same. It grabs one short passage from your page and quotes it. Make the right passage easy to grab and you get quoted.

By Alex Yeskolski, Founder, VuseDesk. 1,373 words, about 7 min at 200 words a minute, 7 sources, 1 table.

Nobody reads the whole page

People read about a fifth of your words, so the answer has to sit near the top.

Weinreich, Obendorf, Herder and Mayer logged 45,237 page views from 25 people using tracking browsers, and Nielsen Norman Group's analysis of that data found users read at most 28% of the words on a page, more likely 20%. The same data showed each extra 100 words bought only 4.4 more seconds of attention. Readers got through half of a page only when it was 111 words or shorter.

That is a phone reader on a job site. It is also how a retrieval system works. It pulls out one chunk that matches the question and skips the rest.

Answer in the first line

Put the answer in the first sentence, then explain it.

No study I found measures whether a direct answer in the first sentence raises how often ChatGPT, Gemini or Google's AI Overviews cite a page. So treat this as a practice, not a proven result. It follows from the reading data above.

Say what the thing costs, how long it takes, or what the rule is, in the first sentence under each heading. Then give the basis for it. A reader in sunlight gets the answer without scrolling, and a machine gets a clean sentence to lift.

Keep each block short

A passage of about 50 to 100 words is easiest for a retrieval system to find.

Bhat, Rudat, Spiekermann and Flores-Herr at Fraunhofer IAIS tested 5 chunk sizes from 64 to 1024 tokens on six question datasets with two embedding models, and for fact questions chunks of 64 to 128 tokens found the right passage best. On the SQuAD dataset in that same study, the 64 token chunk hit the right passage first 64.1% of the time. A token is a piece of a word, and 64 to 128 tokens is roughly 50 to 100 words, though that conversion is my estimate and not one the paper gives.

So write in blocks of 3 or 4 sentences that each stand alone. Give every block one job. One block, one answer.

Use small words

Plain words sell better, and Unbounce's own numbers put it at about double.

Unbounce, which sells landing page software, reports from its own customers' pages that copy written at a 5th to 7th grade reading level had a median conversion rate of 11.1%, while copy at a professional reading level converted at 5.3%. That sample was 41,000 or more landing pages with 57 million or more conversions, by Unbounce's count. On that same sample Unbounce reports that hard words of three or more syllables had a 24.3% negative correlation with conversion, reading time 19.4% and word count 18.6%. Those are correlations from a vendor, not a controlled test.

Nielsen Norman Group ran a controlled test with five versions of one site and identical tasks, and found that concise copy improved usability by 58%, a scannable layout by 47%, objective language by 27%, and all three together by 124% over promotional copy. Cut the sales words. Say the thing.

Put it in text, not pictures

If your price list is a photo, the machine cannot read it.

Google Search Central lists 1 content format rule for its AI features: keep important content in text. Google's image documentation also says Google does not index CSS images, and names alt text as the most important image metadata. Google does not say text inside an image is unreadable, it just says plain text is the safe ground.

Google also gives 2 conditions for a page to be cited as a supporting link in AI answers: it is indexed and eligible for a snippet. It names 4 controls that limit that: nosnippet, data-nosnippet, max-snippet and noindex. Check your site for those tags, because a page with nosnippet on it will not be quoted.

Put your name on it

Google asks whether a page has a byline, so give it one.

Google's helpful content guidance asks 1 direct question, whether pages carry a byline where one might be expected, and says it strongly encourages accurate authorship. The same guidance asks 2 more questions: does the content have any easily verified factual errors, and was it written or reviewed by someone who demonstrably knows the topic well. No study I found links a named author to how often a page is cited, so the byline is Google's stated preference, not a proven lever.

Google also asks whether the use of automation, including AI generation, is self evident to visitors through disclosures. Its spam policy treats automation used mainly to manipulate rankings as a violation. So if a tool drafted it, say so, and check every fact yourself.

No secret markup needed

Google says you need nothing extra to show up in its AI answers.

Google Search Central states there are 0 additional requirements to appear in AI Overviews or AI Mode and 0 special schema.org types to add. Anyone selling a special tag for AI search is selling something Google says you do not need. The work is the writing.

What the evidence says

The whole method in one table, with who measured each number.

Page elementWhat the number saysWho measured it
Opening paragraphUsers read about 20% of words, and half a page only when it is 111 words or fewerWeinreich and colleagues, 25 users, 45,237 page views, via Nielsen Norman Group
Passage length64 to 128 tokens, about 50 to 100 words, best for fact questions, 512 to 1024 for technical contextFraunhofer IAIS, five chunk sizes, six datasets, two models
Reading level5th to 7th grade copy converted at 11.1% against 5.3% for professional levelUnbounce, vendor data from 41,000 or more customer pages
Plain text or imageImportant content must be in text, and CSS images are not indexedGoogle Search Central
Author lineGoogle asks whether pages carry a byline and strongly encourages accurate authorshipGoogle Search Central
Writing styleConcise +58%, scannable +47%, objective +27%, all three +124% usabilityNielsen Norman Group controlled test, five versions of one site
Extra markup0 additional requirements or schema for AI OverviewsGoogle Search Central

What we got wrong

Some of this evidence pushes back on writing for a machine at all.

I started this piece thinking a page had to be written for a machine. Google's own documentation says there are no additional requirements and no special optimization for AI Overviews or AI Mode, so that framing was wrong. What helps a person on a phone is the same thing that helps a retrieval system.

I also wanted to say keep every paragraph short. The Fraunhofer IAIS study does not support that everywhere, because on its TechQA dataset of technical questions the 1024 token chunk hit the right passage first 71.5% of the time against 61.9% for 512 tokens. The 2 embedding models in that study also disagreed, with Stella favouring larger chunks and Snowflake smaller ones.

The Unbounce figures are a vendor's own customer data, and Unbounce sells the product being measured. The Nielsen Norman Group reading tests come from browsing long before phones or chat assistants, and Google never claims a byline changes ranking or citation. What stays human is deciding what the true answer is, putting a real name on it and checking for the factual errors Google asks about.

Do it yourself

You can do this on your own site with no tools.

Take your busiest page, put the answer in the first sentence, cut every block to 3 or 4 plain sentences, keep prices in text and sign it. If you would rather not, VuseDesk can set up the website and the SEO for you, and a consult is $199. Everything on this page is the whole method, free and in full.

Alex Yeskolski Founder, VuseDesk. He writes the software these articles describe. More about the author
Walk the tour, then read the pages

Or call (252) 666-7217 and ask the 1 question this article did not answer. Email [email protected] if you would rather write it down.