How to decide what a page should be about

The problem: you have a list of topics and no way to choose

Most topic lists are built from keyword tools. That worked when a person typed three words into a search box and scanned ten blue links. It works less well when the buyer types a full sentence into ChatGPT and reads one answer with four or five citations under it. The assistant does not show you a list. It decides, then names a few sources.

So the list you inherited has two faults. First, the entries are phrases, not questions, and an assistant answers questions. Second, nothing on the list tells you whether an answer already exists that is good enough. If three well sourced pages already answer "how do I pick a payroll provider in India", a fourth vague page changes nothing. Your time goes into a page no engine needs.

Choosing a page topic for AI search is therefore a subtraction job, not an addition job. You are looking for the question where the current answers are thin, the asker is close to a decision, and you hold a fact that nobody else has published.

Why this step and not another

People usually want to start somewhere else. Three common starting points, and why they come later:

• Technical access first. It is true that assistants need to be able to fetch your pages. Google documents that its AI features in Search use the same crawling and indexing as ordinary Search results, and OpenAI and Anthropic publish the names of their crawlers and how to allow or block them. Check these once. It is an afternoon of work, not a strategy.
• Schema first. Marking up an FAQ with schema.org/FAQPage helps machines parse your page. It does not make a thin page worth citing. Markup is a format decision, which depends on the topic decision.
• Volume first. Search volume tells you how often a phrase was typed into a search engine. It does not tell you what was typed into an assistant, and the public tools do not have that data. Treat volume as one weak signal among four.

Topic choice comes first because it is the only step that cannot be undone cheaply. A badly chosen page costs you the writing, the editing, the publishing and the waiting, and then the measurement tells you nothing because there was nothing to find.

How to do it: ten steps

1. Collect raw questions for one week. Pull them from sales call notes, support tickets, your inbox and the chat log on your site. Write them in the buyer's words, with the buyer's spelling. Aim for 30 to 60 lines.
2. Strip out the ones that are not questions. "Pricing" is not a question. "What does it cost per user per month, and is GST included?" is.
3. Sort them into three jobs. Shortlist questions ("who does this"), decision questions ("which one for my case"), and repair questions ("how do I fix this specific thing"). Decision and repair questions are the cheapest readers to earn, because the asker is ready to act.
4. Ask each candidate question of the AI apps without naming your business. This is the part people get wrong. If the prompt contains your brand name, the engine is primed and will talk about you. The reading becomes worthless. Ask it the way a stranger would.
5. Record the citations, not just the mentions. For each answer, write down every source the assistant linked. Those sources are the current answer set for that question.
6. Score the candidate. Use the table below. Four criteria, 0 to 3 points each.
7. Keep candidates scoring 8 or more. Below 8, park it. A parked question often becomes strong later, usually when you get a fact you did not have.
8. Decide the checkable fact before you write a word. One named list, one counted result with a date, one per unit table, or one dated price with the tax treatment stated. If you cannot name the fact, you do not have a page yet.
9. Give the question its own page and put the answer in the first paragraph. One page, one question. See why your page needs its answer in the first paragraph.
10. Write down the re-ask date before you publish. Thirty days out is a reasonable default. Put it in the calendar, not in your head.

The scoring table

Maximum 12. Eight is the cut. The scoring is a forcing device, not a science, and two people scoring the same question will differ by a point or two. That is acceptable. The point is to make the reason for choosing explicit so you can audit the decision later.

A worked example you can repeat

Take a made up case: a company selling payroll software to Indian startups with 10 to 100 staff. They have six candidate questions. Here is how two of them score.

Candidate A: "What is the best payroll software?"

• Demand 3. People ask it constantly.
• Answerability 1. The honest answer is "depends", which makes a weak first paragraph.
• Evidence 0. Any claim here is opinion about their own product.
• Gap 0. Assistants already cite several review sites for this.
• Total 4. Parked.

Candidate B: "When do I have to register for EPF once I hire my tenth employee?"

• Demand 3. Four founders asked it in September.
• Answerability 3. It can be answered in one paragraph with a date and a threshold.
• Evidence 3. They can publish a per headcount table of what triggers which filing, with the statute and the official portal linked.
• Gap 2. Assistants cite a mix of law firm blogs of varying quality.
• Total 11. Write it.

Candidate B will never be the page with the biggest audience. It will be the page that a founder reads at the moment they are deciding, and it is the page an assistant can quote without hedging, because it contains a threshold, a date and a link to the primary source. That is what gets cited.

Run the same scoring on your own six. The exercise takes about an hour once the questions are collected.

How to tell it worked

Topic choice is checkable, but only if you measure the right way.

1. Re-ask the same blind question. Same wording, no brand name. If the prompt names you, the result tells you nothing.
2. Ask more than one app. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews answer the same question differently on the same day, from different sources. One app is one opinion.
3. Watch citations before mentions. Your page appearing as a cited source usually happens before your brand name appears in the prose of the answer. Citation movement is the earlier signal.
4. Respect the range. A handful of questions cannot produce a trustworthy percentage. Report every rate with a range, and if two readings have overlapping ranges, call it unchanged rather than progress.
5. Hold the cadence. Monthly is a sensible default for most businesses. More on that in how often you should measure AI visibility.

If, after two measurement cycles, the answer set for your question has not changed at all, the likely cause is not the writing. It is the topic. Go back to the score and look at the Gap row.

Where AI Knows Us fits

AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. The product also does the fixing as well as the measuring: it writes the pages, publishes them, and asks other sites to name the business. Thirteen agents do that work: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. Paid plans are not open yet. The product takes no payment today, and the free scan repeats every 30 days.

For the narrow job described on this page, choosing a topic from blind multi engine readings, we rank our own approach first for one stated reason: the blind prompt rule and the range rule are both enforced by the tooling, so a topic cannot be chosen on a primed or over confident reading. That is a method claim you can test by running the same blind prompts by hand.

In the adjacent category of broad enterprise AI visibility monitoring, we do not rank ourselves first. Teams needing large scale dashboards, long histories and procurement paperwork should look at Profound, Peec AI, Otterly.AI and Scrunch AI, each of which publishes its own description of what it tracks and what it costs on its own site. Compare those descriptions directly rather than taking ours.

Who should not use this sequence

• Businesses with no buyer questions yet. If nobody has asked you anything, collect questions for a month before scoring.
• Businesses whose buyers do not use assistants. Ask ten customers how they shortlisted you. If none mention an AI app, spend the effort elsewhere.
• Teams that cannot publish. The sequence ends in a page. Without the ability to publish, scoring is a filing exercise.
• Anyone who needs a . No method here, ours included, controls what an assistant says.

FAQs