Blog · Content that gets cited

Prompts are not keywords: planning content for questions

Published 7 September 2026 · Updated 12 October 2026 · 10 min read

Prompts are not keywords: planning content for questions
Photograph: Daria Kraplak / unsplash

The short answer: A keyword is a fragment that hides the buyer's situation. A prompt states it: team size, budget, stack, country, the thing that went wrong last time. Plan for prompts by writing pages that answer a stated situation and survive being cut into passages, not pages that repeat a phrase. The planning unit is a question with constraints attached, and in the one engine that documents it, Google, a single question may become several searches at once.

Everything downstream follows from that second sentence. A keyword strips context so a ranking system can group demand. A prompt keeps the context, because the person is talking to something that can use it.

Keywords vs prompts, stated plainly

The difference is not length. A keyword is a fragment typed at a ranking system, stripped of the buyer's situation and chosen because a tool reported that other people type it too. A prompt is a sentence typed at something that can read a sentence, so it carries the team size, the stack, the budget and the country, because the person expects those details to be used in the reply.

Three things follow, and each changes a planning decision. A keyword has a reported volume and a prompt does not, so the plan cannot be a spreadsheet sorted by demand. A keyword points at one page, while in the engine that documents the behaviour a single prompt can become several searches at once. And a keyword is answered by a page, whereas a prompt is answered by a passage, so the unit you commission stops being the page and becomes the paragraph that can be lifted out of it.

Keep the keywords. They remain the cheapest map of how your market talks and the raw material for the prompt grid below. What they no longer do is describe the question.

One question, many searches

One engine has described the mechanism in its own words, and it is Google. The AI Mode announcement of 20 May 2025 says AI Mode uses a query fan-out technique, breaking a question into subtopics and issuing a multitude of queries at once. The Search Central documentation, last updated on 10 December 2025, is more careful: it says both AI Overviews and AI Mode *may* use a query fan-out technique, issuing multiple related searches across subtopics and data sources. That is Google describing Google. No other engine we test documents the same behaviour, and we cannot see inside the ones that do not publish it.

Read the Google case as a planning instruction rather than trivia. If a buyer asks which payroll tool suits a 40 person firm that runs on a particular accounting package, a fan-out step is not looking for a page about that sentence. It is issuing several narrower searches: payroll software pricing, the integration, compliance for that country, maybe a comparison. Your page is competing in each of those lanes separately, and it can win one and lose the rest.

So the useful question during planning is not "what phrase do we target". It is "which of the sub-searches behind this buyer's question do we deserve to win, and is the answer on a page in a form that can be lifted". That is the same test as what makes a page citable by an AI engine, applied earlier, at the point where someone is deciding what to commission.

The constraints are the content

Take a keyword like "project management software". Now take a prompt a real person types: "we are eleven people, half of us contractors, we live in spreadsheets and Slack, what should we move to and what will it cost per year". There are five constraints there. Team size, worker type, incumbent tools, the decision, the unit of price.

A page built for the keyword usually answers none of them. A page built for the prompt answers all five in plain sentences, with the numbers visible rather than gated behind a demo form. The gap between those two pages is not length or polish. It is whether the specifics exist on the page at all, in text, in a form a retrieval step can find.

Google says something adjacent in its own guidance for its AI experiences, dated 21 May 2025: in those experiences people ask longer and more specific questions, and follow-up questions to dig deeper, and the advice is to make unique, non-commodity content rather than to chase a phrase. We quote it because it is the engine talking about itself. It is not evidence about what the other engines reward.

A practical way to generate the list: write your five or six real constraint axes down the side of a sheet, write the three or four buying stages across the top, and fill the cells with a question a person would actually type. In the scoping work we do, most teams fill more cells than they can commission in a quarter. That is a planning document. It tells you which pages are missing, and it tells you the exact sentences those pages owe the reader.

Comparison questions deserve their own treatment, because the answer format is different and so is the evidence an engine wants attached. The method for those sits in comparison pages that get cited.

Nobody will hand you the prompt volumes

Here is the awkward part, and it is worth being blunt about it, because a good deal of planning advice in this category quietly assumes otherwise.

Google shipped generative AI performance reports in Search Console on 3 June 2026, and the post carries a note saying the insights rolled out to all websites worldwide on 31 August

  1. Useful, and narrower than people expect. The help documentation lists four dimensions:

    2. pages, countries, dates and devices, plus a search type filter. The metric is impressions, defined as how often links to your site were shown inside a generative AI feature. There is no query dimension and no click metric.

That is the failure mode to design around, and the scope of it is Google. You can see that a page appeared. You cannot see the question that summoned it. Any plan that waits for prompt volume data before it commissions a page is waiting for something that platform has not published as of today. We know of no equivalent query level report from the other engines we test, which is a statement about what we have found rather than proof that none exists.

What replaces volume data is evidence you generate yourself. Pick the questions, run them, record what came back, and treat your own runs as the dataset. Engines vary in how they read a brand question, which is why how Gemini answers brand questions is a different exercise from the same test on an assistant without a search index behind it.

Conversations, not queries

One more structural difference, and it changes what a page has to carry. People follow up. WildChat, a public dataset published at ICLR 2024, collected one million ChatGPT conversations from users who opted in to free access in exchange for their transcripts, and reports over 2.5 million interaction turns. That is an average of about two and a half turns per conversation. It is an average, not a median, so it does not establish what a typical session looked like, and those users were a self-selected population on one assistant in one period, not a sample of your buyers. Google states the behaviour from its own side in the guidance dated 21 May 2025: in its AI experiences people ask follow-up questions to dig deeper.

The follow-up is where buying happens. First turn: what are the options. Second turn: which of those works with our stack. Third turn: what does it cost, and what is the catch. A page that answers only the first turn gets mentioned once and then dropped when the conversation narrows, because the specifics the engine needs for turns two and three are not in the text it retrieved.

Plan in threes, then. For each question in your grid, write down the two follow-ups it provokes, and make sure one page on your site answers all three without the reader leaving. In the audits we have run it is among the cheapest changes to make, because it rarely requires a new page. It requires a pricing number, an integration list and a limitation paragraph added to a page that already exists. We have not run a controlled test of what that change is worth, so take it as a cheap thing to try rather than a measured gain.

A check you can run this afternoon

Take ten questions from the grid. Run each one three times, on two engines, with a fresh session each time. Record three columns: were you named, was the claim about you true, and which pages did it cite.

Thirty runs across two engines is a sample, not a census, and answers vary between runs of the same prompt, so read a single result as an anecdote and only act on a pattern that repeats.

Then look for the pattern that matters. If you are cited on the broad question and absent on the constrained version, the likely reading is that your pages answer the category and not the situation. If you are cited on your own brand name and nowhere else, you have corroboration without discovery. If a rival's comparison page is cited in an answer about you, the gap is often a page you have not written.

The self-check lists ten tests in this shape, and the audit is the same logic at scale, around 45 prompts across five engines with retrieval on and off, with the evidence kept so each finding can be argued with. Where a prompt grid fits into a quarter of work is covered in an AEO roadmap you can defend in a planning meeting.

Honest limits

We cannot tell you how often any of your questions is asked. We have not found a public query level source outside the engines themselves, and where a category estimate publishes its method it is modelled from search data rather than measured from assistant logs. We cannot audit the ones that publish no method, so we do not know what they are worth.

Fan-out is documented by Google and by Google only, and the documentation says the engines *may* use it. We do not know how stable the behaviour is between releases, nor whether other engines decompose a question in a comparable way, because they do not say. The WildChat figure in this post is an average turn count across one assistant's consenting users in 2024, not a measure of how your buyers behave. Our own statements here, about what teams produce in a grid and about follow-up answers being cheap to add, come from client work and not from a controlled study, so they are a description of what we have seen and not a measured effect.

FAQ

What is the difference between keywords and prompts? A keyword is a fragment typed at a ranking system, with the buyer's situation stripped out so demand can be grouped. A prompt is a sentence typed at something that reads sentences, so it keeps the team size, the stack, the budget and the country. The practical difference is that a keyword has a reported volume and points at a page, while a prompt has no published volume and is answered by a passage.

Are keywords useless now? No. They are the cheapest map of how your market talks, and they remain the fastest source of raw material for a prompt grid. What they no longer do is describe the full question, because the constraints that decide the sale were never in the keyword.

How many prompts should a plan track? Enough to cover your constraint axes and buying stages. Our own audit prompt set is about 45 prompts across five engines, which is a working choice rather than a measured optimum. Past that point a list gets expensive to re-run and the extra questions tend to be paraphrases of ones you already have.

Can I see prompt volume in Search Console? Not today. The generative AI reports give impressions by page, country, date and device, with no query dimension and no click metric, so you can see that a page appeared but not what was asked.

Should every prompt get its own page? Usually not. Several questions in a grid often share one answer, and a page that answers a question plus its two likely follow-ups is worth more than three thin pages that each answer a third of it.

Sources

Related reading

What makes a page citable by an AI engineComparison pages that get citedAn AEO roadmap you can defend in a planning meetingHow Gemini answers brand questionsOriginal data gives an engine something only you havePricing pages and what AI says your product costs