Search queries used to be keywords. Increasingly they are full questions, and the pages that get quoted — in featured snippets and in AI answers alike — are the ones that answer the question as it was asked. Most sites cover the topic. Far fewer pose the question.
There is a perfectly good reason most sites are built this way. When you know a subject well, you stop thinking in questions. You write the page a professional would write: a service page describing what you do, a specification listing what is included, a process page walking through the steps. All of it is accurate, and none of it is wrong. It simply is not shaped like the thing being typed into the search box, which is "how long does a treatment take", "what does it cost to replace a roof", "can I cancel before the notice period".
That shape mismatch used to cost little, because a keyword-matching search engine could bridge it. It matters more now, because the systems assembling answers — Google's snippets, the AI assistants — work by lifting a passage that already answers the question. If no passage on your site does, there is nothing to lift.
Why do question searches matter for AI visibility?
A growing share of searches — and nearly everything typed into an AI assistant — arrives as a full question: how long does, what does it cost, can I. Featured snippets and assistant answers are assembled from pages that answer the question as it was asked, plainly and early, so a page that covers the topic without ever posing the question is competing against pages that do.
This is not a new ranking trick; it is how these systems have to work. A snippet is a quoted passage, and an assistant citation is a quoted source. Both need a self-contained piece of text that stands on its own as an answer. A paragraph buried in the middle of a service description rarely does, even when the information is technically in there, because the passage only makes sense in context. A heading that poses the question, followed by a direct answer, makes sense lifted out — which is precisely why it gets lifted.
It is worth saying that the two surfaces reward the same writing. You do not need one version of a page for Google and another for the assistants. The mechanics of why assistants pick the sources they do — and why their answers vary between sessions — are covered in how to get cited by AI search and why AI answers differ every time; the writing advice underneath both is this article.
Where do I find the questions people actually ask?
In your own Search Console query data. It lists the searches your site was actually shown for, and the question-formed queries among them — the ones beginning with how, what, why, can or does — are the questions your subject genuinely gets asked. That list already belongs to you; no keyword tool has to guess it.
This matters because the alternative — inventing questions from your own head or from a keyword tool's suggestions — produces the questions a professional would ask, in the vocabulary a professional uses. Your query data contains the phrasing that was really typed, including the imprecise, everyday versions. Those are the versions worth answering, because those are the ones a snippet or an assistant will be asked to resolve. If Search Console is unfamiliar territory, the Search Console numbers that matter walks through the report this data lives in.
One caution when you go looking: read the queries raw. Many analysis tools "clean" or normalise text before showing it to you — lowercasing, stemming, dropping short common words — and the words most likely to be dropped as noise are exactly how, what and can. A normaliser can quietly turn "how long does an mot take" into "long mot take", and at that point the question has become a keyword and the intent is gone. The first word of the raw query is the signal. If your export or your tool cannot show you that first word untouched, work from the raw export instead.
How do I check whether a page answers a question as asked?
Take each real question and look for two things: a heading somewhere on your site that literally poses that question, and a first paragraph under it that answers in one or two sentences before any preamble. If both are there, the page answers as asked. That pairing — the question in a heading, the answer in the opening lines — is exactly what featured snippets and assistant citations select for.
The audit is mechanical enough to do with a spreadsheet. One column of real questions from your query data; one column for the page, if any, whose heading poses each question; one column for whether the paragraph under that heading answers it directly. Three outcomes fall out:
- No page poses the question. The most common outcome, and the easiest to fix: you usually already have the answer somewhere, written as prose. Give it a heading in the question's own words and lead with the answer.
- A heading poses it, but the answer arrives late. The section opens with background, caveats or history and gets to the point in paragraph three. Move the answer to the first sentence and let the nuance follow it — the nuance is still valuable, it just should not stand in front.
- A heading poses one question and the text answers another. A heading asks "what does it cost" and the paragraph explains why pricing varies. Understandable, and honest in its way — but a system quoting that passage as the answer to the cost question will make you look evasive. Either answer the question posed, ranges and all, or change the heading to the question you are actually answering.
"One or two sentences before any preamble" is the part people resist, because it feels blunt. It is not a rule against depth; it is a rule about order. Answer first, then explain, qualify and expand as much as the subject deserves. The reader who wanted the short answer has it, the reader who wanted the full picture keeps reading, and the machine looking for a quotable passage finds one.
Should I add FAQ markup to every question I answer?
Only mark up questions that are visibly on the page with the same answer a reader sees. FAQ markup that is hidden, or that gives a different answer from the visible text, is against Google's structured data rules — and, more practically, it promises something the page does not deliver, which is the fastest way to lose the trust the markup was meant to earn.
The honest version is boring and effective: the question is a heading on the page, the answer is the paragraph under it, and the markup repeats both word for word. Structured data is a label on content, not a substitute for it. If you find yourself putting a question in the markup that has no matching heading, the fix is not to remove the markup — it is to add the heading, because the visible version is the one doing the real work anyway.
Three ways this goes wrong
The failure modes are all versions of the same mistake: optimising for the format instead of the reader.
Manufactured questions. Once a site adopts the question-and-answer shape, it is tempting to fill it with questions nobody asks — "why choose us for your roofing needs?" is a heading, and it is shaped like a question, but it appears in no one's search box. Manufactured questions dilute the real ones and read as advertising. The discipline is to let your query data write the list: if a question never appears there and you cannot honestly imagine anyone typing it, it does not earn a heading.
The buried answer. Four hundred words of introduction before the answer is the traditional essay structure, and for search it is upside down. The passage a snippet wants is at the top of the section, not the bottom. If your sections habitually open with throat-clearing, the single highest-value edit on the whole site is moving each section's conclusion to its first line.
The swapped answer. Answering a different question than the heading poses — usually the question you wish had been asked. The heading says "how long does it take" and the text explains how thorough the process is. Readers notice, and so does anything quoting you.
What to do this week
Export your Search Console queries and read them raw, first word intact. Pull out the ones that are questions — there will be fewer than you expect, and they will be more specific than you expect. For each one, check whether any heading on your site poses it and whether the first paragraph underneath answers it in a sentence or two. Fix the gaps in order of the traffic behind them, and only add FAQ markup where the question and answer are already visible on the page.
TrustCtrl runs this audit continuously: it reads the question-formed queries from your own search data, checks them against the headings your pages actually render, and reports each question as answered, answered late or not posed at all — alongside the rest of the picture, from whether AI search can cite you to whether the crawlers can reach you in the first place.