How do I build a page AI can answer from?

Build a page that AI can answer from as a set of short answers that each stand on their own. Open with a passage of 150 to 250 words that answers the page's main question by itself. Give every buyer question its own section, with the question as the heading and the answer in the first 2 or 3 sentences. Write the prices, numbers with units, dates, places and names into those sentences, keep the prose as prose, and make sure the words are in the HTML your server sends. Then test the page before publishing: give a model the page as its only source, ask the buyer questions, and fix every section it misses.

The evidence comes from NRLC's token-efficiency benchmark, published on 2026-11-09. Given only the 4 best-matching passages of an article, the model answered 98.5% of the answerable questions correctly, against 95.1% when it read the whole article. The 4 best-matching passages used 89% fewer input tokens than the whole article. Short, self-contained passages beat the whole page.

The skill gives an AI agent the method on this page: write 10 buyer questions and 3 traps first, build the page, test it with 2 models, and publish only if it passes. For Claude Code, save it as ~/.claude/skills/answer-pages/SKILL.md. The Markdown is also at /downloads/page-anatomy-for-ai-answers.md.

The answer passage at the top of a page should run 150 to 250 words. That is long enough to carry the specifics and short enough to be retrieved and read as one passage; the passages in NRLC's benchmark were about 200 words each. The answer passage has to make sense as the only text a model sees, so it names the subject, gives the answer with its number, unit, place and date, and says where the figure comes from.

Each buyer question gets its own section because a model often answers from one retrieved passage, and in NRLC's benchmark the whole page was the harder source: the model got 39 answerable questions wrong when it read the whole article, and answered 36 of those 39 correctly from the 4 best-matching passages. Put the buyer's question in the heading, answer it in the first 2 or 3 sentences, then add the detail. Name the subject again in every section instead of writing "as mentioned above": a kitchen remodel page says "a kitchen remodel in Austin" in each section, not "it".

Prices, numbers with units, dates, places and names belong in the sentence that makes the claim, not in a sidebar, image, tooltip, tab or accordion: in the first run of NRLC's benchmark, labelled exploratory, the sentence holding the answer carried a name, number, date or other concrete detail for 76 of the 80 answerable questions. Give each figure once, in one place, with its source and the date it was true; the figures on this page are as of 2026-11-09.

A summary or key-takeaways box helps people skim, but it will not carry a page's facts. In NRLC's benchmark, the model's own summaries kept the answer for only 128 of the 800 answerable questions. Answers from those summaries were correct 9.4% of the time. The summaries ran 122 to 493 words. Keep a takeaways box if readers like it, but never put a fact only there.

Rewriting a page as a list of one-line facts saved little and cost accuracy in NRLC's benchmark: the full fact-list rewrite of each article was only 6% smaller than the original prose, and answers from it were 4.9 points less accurate than answers from the whole article. The original Wikipedia prose did better, so spend the effort on clarity and specifics rather than on reformatting.

On most web pages, visible text is a small share of the HTML. On 12 pages sampled in the first run of NRLC's benchmark, from Wikipedia, BBC News, Shopify, Stripe and others, visible text was a median 4.9% of the HTML by characters, and markup used more tokens per character than prose: about 2.5 characters per token against about 4. Any system that reads raw HTML pays for the rest, so the words that answer buyer questions must be in the HTML the server sends, not only in images, carousels or text that a script loads later.

Test an answer page with 10 buyer questions plus 3 trap questions that the page does not answer. A decision page, one that helps a buyer choose between options, gets 6 buyer questions, 3 comparison questions, 2 recommendation questions, 1 question about where the other option is better, and 3 traps. Write the questions before drafting, save them with the date, and don't change them once testing starts; this page's questions were saved on 2026-11-09.

An answer page passes when a model answers at least 9 of the 10 buyer questions correctly twice: once from the whole page and once from only the 2 best-matching passages of about 200 words. The model must also answer "not stated" to all 3 traps instead of inventing an answer. NRLC runs the test with two models: qwen2.5:7b-instruct on a laptop and a current frontier model three times per question. The INGEST fact-density check adds two house targets: at least 50% of the text in specific sentences and visible text at least 10% of the HTML.

NRLC's token-efficiency benchmark had 1,000 questions in total, about 50 English Wikipedia articles. The model got each article seven ways: whole, as the 2 or 4 best-matching passages, as a summary, and as fact-list rewrites of different sizes. Some of the 1,000 questions were deliberately unanswerable: 200 could not be answered from the article. The model was qwen2.5:7b-instruct, run on one laptop at temperature 0, and the whole run took about 3 hours of model time. The protocol was committed on 2026-11-08, before the run, and the questions, all 7,000 answers and the code are published with the study.

These page-building rules come from a 7-billion-parameter model answering factual questions about pages it was handed. Models larger than 7 billion parameters read long pages better, so the gap between a whole page and its best passages may be smaller for them. The benchmark did not test opinions, comparisons or recommendations, so NRLC's rules for decision pages are house rules, held to account by the question test above. No page layout guarantees a citation: ChatGPT, Perplexity and Google AI Overviews retrieve and rank in ways the benchmark did not measure.

We tested this page with its own method on 2026-11-09, using 13 questions saved before the first run: 10 buyer questions and 3 traps. Each model answered every question twice, once from the whole page and once from only the page's 2 best-matching passages. Claude Sonnet ran 3 times per question; qwen2.5:7b-instruct ran once, at temperature 0.

Before the rewrite, this page kept its key figures in stat labels rather than in sentences, and both models answered only 8 of the 10 buyer questions from passages, below the pass mark of 9. After the rewrite, Claude Sonnet answered 10 of 10 from the whole page and 9 of 10 from passages in all 3 runs, and qwen2.5:7b-instruct answered 10 of 10 and 9 of 10. Neither model invented an answer to a trap, before or after. The INGEST check put 35.8% of the old page's text in specific sentences and 66.4% of the new page's, and visible text rose from 7.9% to 12.9% of the HTML.

The first two drafts of the rewrite failed with qwen2.5:7b-instruct, at 8 of 10. Where a sentence carried several figures, the model took the wrong one or none: 800 answerable questions instead of 1,000 in total, 498 instead of 128 for the summaries, and "not stated" for the token saving and for summary accuracy. Rewriting those sentences to one figure each fixed all four. One miss remains: given only passages, both models missed the question about passage accuracy, because keyword ranking picked the section on buyer questions over the opening answer that holds those figures.

Download the answer-page skill (SKILL.md)