---
name: answer-pages
description: Build and test web pages that AI models can answer from. Use when creating or rewriting a page that answers one buyer question (an answer page) or helps a buyer choose between options (a decision page), when asked to make a page "AI-ready", "LLM-friendly" or easier for AI search to answer from, or when asked to test whether a model can answer buyer questions from a page.
---

# Answer pages and decision pages

Method by NRLC.ai (https://nrlc.ai/en-us/insights/page-anatomy-for-ai-answers/), version 2026-11-09.

Use this skill whenever you create or rewrite a page that either answers one buyer question (an **answer page**) or helps a buyer choose (a **decision page**). Before starting, collect from the user:

- **Page:** the URL or planned URL
- **Type:** answer or decision
- **Main question:** the buyer's question, in the buyer's words
- **Options compared** (decision pages only)
- **Does the page owner sell one of the options?** If yes, which one

The rules come from NRLC.ai's token-efficiency benchmark (https://nrlc.ai/research/token-efficiency-benchmark/). In that benchmark, a 7B model answered 1,000 questions about 50 Wikipedia articles. From the 4 best-matching passages of about 200 words it scored 98.5%, against 95.1% from the whole article, using 89% fewer input tokens. Model-written summaries scored 9.4%, and rewriting the text as one-line facts lost 4.9 points. The benchmark tested factual questions only, so the decision-page rules below are house rules, and the test in Step 4 is what holds them to account.

## Hard rules

- Never invent facts, figures, prices, clients, quotes or dates. If a needed fact is missing, leave a visible `[NEEDS SOURCE: …]` note and report it. Don't guess.
- Every figure gets a source link and an as-of date.
- Competitor prices come only from the competitor's own pricing page.
- Never promise or imply that a page will be cited by ChatGPT, Perplexity, Google or Copilot.
- No "it's not X, it's Y" sentences, no em-dash chains, no buzzwords (comprehensive, innovative, robust, leverage, unlock, seamless, cutting-edge), no unsourced numbers.

## Step 1: write the questions first

Before drafting, write the page's questions and save them with today's date. Don't change them after testing starts.

| Question type | Answer page | Decision page | Correct answer |
|---|---|---|---|
| Buyer questions (facts) | 10 | 6 | The fact, copied from the page |
| Comparison ("which is cheaper for …") | – | 3 | The option and figure the page states |
| Recommendation ("I have a 5-person team and need …; which?") | – | 2 | The option the verdict passage assigns |
| Concession ("where is the other option better?") | – | 1 | The point the page concedes |
| Traps (plausible, not answered on the page) | 3 | 3 | "not stated" |

Phrase each question the way a buyer would ask it. The buyer questions become the page's section headings. Save them as JSON, for example:

```json
{"page": "https://example.com/pricing/", "type": "answer", "saved": "2026-11-09",
 "questions": [
  {"id": "q01", "q": "How much does a kitchen remodel cost in Austin?", "gold": ["$38,000 to $65,000"]},
  {"id": "t01", "q": "Do you offer financing?", "unanswerable": true}
 ]}
```

`gold` lists acceptable answers; `"A & B"` means both parts must appear.

## Step 2: build the page

### Answer page structure

1. **Title and H1:** the main question, in the buyer's words. Add the year if the answer changes over time.
2. **Answer passage, first on the page:** 150–250 words that answer the H1 with the specifics: the number and its unit, the place, the date it was true, and its source. It must make sense as the only text a model sees.
3. **One section per buyer question.** The heading is the question. The first two or three sentences answer it, then the detail follows. Each section names the subject again.
4. **Conditions and exceptions:** when the answer changes, for whom, and what to check.
5. **Sources and as-of dates:** next to the facts, plus a short list at the end.
6. **Key takeaways (optional):** for skimming readers only. Nothing may appear only here.
7. **One next step,** named plainly ("Request a feasibility brief"). Never "Learn more".

### Decision page structure

1. **Title and H1:** the decision, in the buyer's words ("Tool A vs Tool B for inference cost").
2. **Verdict passage, first on the page:** 150–250 words saying who should choose which, and why: "Choose A if …; choose B if …". Name every option, the two or three criteria that decide it, and the numbers behind them. If the page owner sells one option, disclose it here.
3. **One section per criterion** (price, setup time, data access, where it runs, support). Each compares every option in sentences, with figures and dates. A table may follow, but every value in it must also appear in a sentence.
4. **One section per option, "Who X is for":** the option's name, best use and main limit, readable on its own.
5. **Where the other option wins,** stated plainly.
6. **Prices:** one per option, dated, linked to the vendor's own pricing page. Say "custom" if that's the truth.
7. **One next step,** named plainly.

An "alternatives" page lists real alternatives with the same criteria for each. The owner's own product may be one of them, disclosed, and not automatically first.

## Step 3: writing rules for every section

- Write each section as if it will be read alone, by someone who hasn't seen the rest of the page.
- Name the subject every time. Never open a section with "this", "it", "as mentioned above" or "the above".
- Answer first, in the first two or three sentences, then add detail.
- Put numbers with units, places, dates and sources in the sentence that makes the claim, not in sidebars, images, tooltips, tabs, accordions or stat labels.
- Write plain prose in short paragraphs. Use lists only for steps or genuinely parallel items. Don't rewrite the page as a list of one-line facts.
- State each figure once, in one place.
- Make sure every answer is in the server-sent HTML. Check the page with JavaScript off.
- Length follows the questions. Never pad.

## Step 4: test before publishing

1. **Prepare two inputs per question:**
   - **Whole page:** the visible text of the page's main content from the server-sent HTML, fetched with JavaScript off.
   - **Passages only:** the same text split into passages of about 200 words (whole paragraphs, added until a passage reaches 200 words), keeping the two best matches for the question by BM25 keyword ranking.
2. **Use two models:**
   - `qwen2.5:7b-instruct` via Ollama, temperature 0, seed 0, run once.
   - One current frontier model, run three times per question, each call seeing only the prompt.
3. **Use the same prompt for every run:**
   ```
   Answer the question using only the text below. If the answer is not in the text, reply exactly: not stated.

   TEXT:
   {{input}}

   QUESTION:
   {{question}}
   ```
4. **Score each answer:**
   - Facts: correct if the answer contains a gold answer as a whole word or phrase after normalizing case, accents, dashes and thousands separators.
   - Recommendations: correct if the answer names the option the verdict assigns.
   - Traps: correct only if the answer is "not stated".
5. **Log every miss** as one of three kinds, and fix the section it points to:
   - **Not stated:** the fact is missing or hard to find.
   - **Wrong:** often a nearby fact from another section.
   - **Made up:** an answer to a trap.
6. **Run INGEST** (https://ingest.nrlc.ai/test?focus=density) on the URL and record two numbers: the share of text in specific sentences, and visible text as a share of the HTML.

For a rewrite of an existing page, run the same saved questions on the old version first and report before and after.

## Step 5: publish only if every bar passes

The bars apply to both models.

| Check | Answer page | Decision page |
|---|---|---|
| Buyer questions, whole page | ≥ 9 of 10 | ≥ 5 of 6 |
| Buyer questions, passages only | ≥ 9 of 10 | ≥ 5 of 6 |
| Comparison questions | – | 3 of 3, every run |
| Recommendation questions | – | 2 of 2, every run |
| Concession question | – | 1 of 1 |
| Made-up answers to traps | 0 | 0 |
| Text in specific sentences (INGEST) | ≥ 50% (house target) | ≥ 50% (house target) |
| Visible text as share of HTML (INGEST) | ≥ 10% (house target) | ≥ 10% (house target) |

Pass/fail gates:

- [ ] Every figure has a source link and an as-of date
- [ ] No `[NEEDS SOURCE]` notes left
- [ ] The page reads correctly with JavaScript off
- [ ] Decision pages disclose when the owner sells an option
- [ ] Competitor prices come from the competitor's own page
- [ ] The writing rules above are clean
- [ ] The questions, answers and scores are saved with the page

## Step 6: measure after publishing

1. **Before launch:** ask the page's buyer questions in ChatGPT, Perplexity, Google AI Overviews and Microsoft Copilot, three runs each. For each answer, record whether it was correct against the page, whether the page was cited, which sources were cited instead, and for decision pages the option recommended and the price quoted.
2. **Repeat** weekly for 4 weeks, then monthly. Report answer pages and decision pages separately.
3. **Re-run Step 4** when a price, rule or figure on the page changes, when a live answer gets a fact wrong or quotes an old price, or when a major model update ships. Update each fact's as-of date when you re-verify it.

## Report

When the page is done, report to the page owner:

- The page URL and type
- The saved questions
- A score table for both models and both inputs, with every miss and the fix made
- The two INGEST numbers
- Any `[NEEDS SOURCE]` items or decisions the owner needs to make
- The baseline from the live AI products