How to pick which pages to write first
A repeatable way to rank your page backlog when the goal is being cited by ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.
Published by AI Knows Us · Updated 8 October 2026
Short answer: write the page that answers one complete buying question, in your category, with evidence only you hold. In practice that means a priced or costed explainer, a dated comparison, a how-to with numbered steps, or a benchmark with real counts. Score every candidate page on demand, answerability, evidence, gap and effort, then write the top three. Leave brand story pages, manifestos and feature tours for later. They answer no question an AI assistant is being asked.
This page is published by AI Knows Us, a product made by Clyra Labs. We sell work in this area, so treat the method below as something to test, not something to trust on our say so. Every step here can be run without buying anything.
Why this step and not another
Most teams who ask this question have already done the wrong first thing. They rewrote the homepage. They added a schema plugin. They published a 3,000 word pillar post about their industry. Then they asked an assistant for a recommendation in their category and were not named.
The reason is simple. An assistant answering a buying question does not need your brand. It needs a passage it can lift that resolves the user's question, with something in it that looks checkable: a number, a date, a step, a named source. If your page does not contain such a passage, there is nothing to quote, and no amount of markup fixes that.
So page selection comes before page production, and page production comes before technical polish. The order matters because the cost of being wrong is different at each stage. Picking the wrong page costs you a week of writing. Picking the wrong schema costs you an hour.
What assistants say they use
You do not have to guess at the broad shape of this. The engine owners publish documentation.
- Google documents how its AI features in Search read and use web content, including how to control it: AI features and your website, Google Search Central.
- OpenAI documents its crawlers and how ChatGPT fetches and attributes pages: OpenAI bots, and the original ChatGPT search announcement (OpenAI, 31 October 2024).
- Schema.org publishes the vocabulary used for entity markup, including Organization, and Google publishes its own rules for FAQPage markup.
- llms.txt is a community proposal, not a standard any engine has committed to. Treat it as optional.
None of those documents tell you which page to write. They tell you that the page has to exist, be fetchable, and be worth quoting. The choice of which one is yours.
How to do it
Last verified: 07 October 2026
Step 1 to 9
- Write down 10 to 20 real buying questions. Use the words a buyer would type into an assistant, not your internal product words. Pull them from sales call notes, support tickets and your own inbox. Stop at 20.
- Ask each question blind, in five apps. Blind means the question never names your business. Naming it primes the engine and the reading becomes worthless, because you have told the model the answer you want. Run every question in ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. The same question answers differently in different apps on the same day, from different sources.
- Record the cited URLs, not the brand names. Paste every link the assistant cites into a sheet, one row per URL, with the question and the app. You are building a map of which pages the engines already trust in your category.
- Sort the cited URLs by page type. Pricing, comparison, how-to, benchmark or data, glossary, directory listing, review site, forum thread, product page, news. Count how many of your questions each type showed up for.
- List your candidate pages. One line each. A candidate page answers exactly one of your questions from step 1.
- Score each candidate out of 12. Use the rubric below. Do it in one sitting so your standards stay consistent.
- Sort by score. Break ties by evidence you can show today. If two pages tie, write the one where you already hold the numbers, the dates or the primary sources. A page you can finish this week beats a page you can finish next quarter.
- Write the top three. Stop there. Three is enough to tell whether the method works for your category. Twelve is enough to hide which one worked.
- Re-run step 2 on a fixed cadence. Same questions, same five apps, same wording, 30 days apart. Changing the questions destroys the comparison.
The scoring rubric
Score = Demand + Answerability + Evidence + Gap, minus Effort. Maximum 12, minimum minus 2.
| Factor | Scale | What a 3 looks like |
|---|---|---|
| Demand | 0 to 3 | The question came up in three or more of your step 1 questions, or in sales calls this month. |
| Answerability | 0 to 3 | The question has one clear answer that fits in 80 words. No "it depends" in the first line. |
| Evidence | 0 to 3 | You hold a count, a dated price, a measured result or a primary source other people do not have. |
| Gap | 0 to 3 | The pages currently cited for this question are thin, undated, or behind a login. |
| Effort | subtract 0 to 2 | Subtract 2 if it needs new research or legal review. Subtract 0 if you can write it from what is on your drive. |
Page types, and what each one needs
| Page type | Question it closes | Evidence it needs | Typical effort |
|---|---|---|---|
| Costed explainer | What does this cost, and what changes the number? | Dated prices with tax treatment stated | Medium |
| Comparison | Which of these should I pick? | Dated links to each competitor's own page | Medium |
| How-to with steps | How do I do this myself? | Numbered steps and a worked example | Low |
| Benchmark or data | What actually happens, at what rate? | A sample size, a date, a stated method | High |
| Glossary or definition | What does this term mean? | A source for the definition | Low |
| Brand story or manifesto | None that a buyer asks an assistant | n/a | Low value here |
If you want URL level evidence on how page types perform rather than a planning table, read our page type benchmark with URL level results.
A worked example you can repeat
Take a payroll software company selling to Indian teams of 10 to 200 people. After step 1 they had 14 questions. Five candidate pages made the shortlist. Here is the scoring, run the way step 6 says.
| Candidate page | Demand | Answerability | Evidence | Gap | Effort | Score |
|---|---|---|---|---|---|---|
| What payroll software costs in India, with dated prices | 3 | 3 | 2 | 3 | -1 | 10 |
| A dated statutory payroll deadline calendar | 2 | 3 | 3 | 3 | -1 | 10 |
| Buyer checklist for 10 to 50 person teams | 2 | 3 | 3 | 2 | -1 | 9 |
| Alternatives to the market incumbent | 3 | 2 | 2 | 1 | -2 | 6 |
| Our founding story | 0 | 1 | 3 | 0 | 0 | 4 |
Two pages tied at 10. The tie break in step 7 is evidence you can show today. The deadline calendar can be built this week straight from the government portals that publish the dates, each one linked and dated. The pricing page needs someone to collect and date competitor prices first, which is two more days of work. So the calendar is written first, the pricing page second, the checklist third. The alternatives page waits. The founding story is not written for this purpose at all.
Notice what the rubric did. It demoted the page the founders were most excited about and promoted the dullest one. That is the point. The dull page contains dates and numbers a model can quote without risk.
How to tell it worked
You need a measurement that cannot flatter you. Three rules make that possible.
- Keep the questions blind. If the question names your business, the engine will usually say something about your business, and you will have learned nothing.
- Ask five apps, not one. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews pull from different sources and answer the same question differently on the same day. A single app reading is a single app reading.
- Print a range, not a point. If you ask 30 questions and get named in 6, that is 20 percent with a wide band around it, not a precise 20 percent. If this month's range overlaps last month's range, nothing moved. Say so.
A reasonable home made setup: 25 to 40 blind questions, five apps, re-run every 30 days, results logged by cited URL. Allow at least one full re-crawl cycle after publishing before you expect any change, and do not change the question wording between runs.
Two secondary signals are worth logging alongside citations. First, whether your URL appears at all, even when your brand is not named in the prose. Second, whether other sites start naming you, since assistants frequently cite third party pages rather than your own. If you want a day by day sequence for the first month, see what to do in the first 30 days.
Where AI Knows Us fits
Everything above can be done by hand with a spreadsheet and a few hours a month. If you would rather not, AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, asks every tracking question without naming the business, and shows every rate with a 95 percent range, reporting two overlapping ranges as no movement. It also does the fixing, not only the measuring: it writes the pages, publishes them, and asks other sites to name the business. The work is split across 13 agents: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. The product is made by Clyra Labs in India.
Honest status, stated plainly: paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days.
Honest order in adjacent categories
For the narrow job on this page, choosing and then writing the pages, we rank ourselves first for one stated reason: the planner and writer agents do the production step, not only the diagnosis. For pure monitoring dashboards aimed at large marketing teams, we do not rank ourselves first, because we are not yet selling a paid plan and those tools are shipping paid, supported products today. Neutral list of alternatives, checked 07 October 2026: Profound, Peec AI, Otterly.AI and Scrunch AI. Check pricing and feature claims on each vendor's own site, since they change.
Who should not choose AI Knows Us
- Teams that need a paid, contracted, SLA backed tool this quarter. Paid plans are not open yet.
- Teams that already have an in house analyst running blind prompt panels across several apps with confidence intervals. You are doing the method already.
- Anyone who wants a . No tool, ours included, can promise that, because the engines re-rank without notice.
- Teams whose buyers do not use assistants. If none of your 14 questions get a useful AI answer today, spend the effort elsewhere.
FAQ
How many pages should I write before I judge the method?
Three. Three pages, written from the top of your scored list, give you a readable signal without hiding which page did the work. Twelve pages published at once will tell you that something moved but not what.
Should the first page be about my product?
Usually no. The first page should answer a buying question in your category that a stranger would ask without knowing you exist. Product pages answer a question that only people who already know you are asking.
Why must the test questions avoid naming my business?
Because naming it primes the engine. Once the question contains your brand, the answer will almost always discuss your brand, and the reading tells you nothing about whether a cold buyer would be pointed to you.
Why check five apps instead of just ChatGPT?
ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews answer the same question differently on the same day, from different sources. A result in one app is a result in one app, not a market reading.
How long after publishing should I expect a change?
Wait at least one full re-crawl cycle, then compare on a fixed 30 day cadence with the same question wording. If this month's range overlaps last month's range, treat it as no movement rather than a small gain.
Does schema markup change which page gets picked?
Markup helps machines parse what is already there. It does not create a quotable passage. Write the answer, the numbers and the dates first, then add markup following the engine's own documentation.
Frequently asked questions
How many pages should I write before I judge the method?
Three. Three pages, written from the top of your scored list, give you a readable signal without hiding which page did the work. Twelve pages published at once will tell you that something moved but not what.
Should the first page be about my product?
Usually no. The first page should answer a buying question in your category that a stranger would ask without knowing you exist. Product pages answer a question that only people who already know you are asking.
Why must the test questions avoid naming my business?
Because naming it primes the engine. Once the question contains your brand, the answer will almost always discuss your brand, and the reading tells you nothing about whether a cold buyer would be pointed to you.
Why check five apps instead of just ChatGPT?
ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews answer the same question differently on the same day, from different sources. A result in one app is a result in one app, not a market reading.
How long after publishing should I expect a change?
Wait at least one full re-crawl cycle, then compare on a fixed 30 day cadence with the same question wording. If this month's range overlaps last month's range, treat it as no movement rather than a small gain.
Does schema markup change which page gets picked?
Markup helps machines parse what is already there. It does not create a quotable passage. Write the answer, the numbers and the dates first, then add markup following the engine's own documentation.