What is this?
This check asks whether your headings are literally phrased as questions. AccessKnight sends every h1 through h4 on the page to a model with a strict instruction: award points only for headings that are actually interrogative, usually ending in a question mark. A heading that could be reworded as a question earns nothing.
The strictness is deliberate and it is written into the grading prompt, which names its own failing examples: action, feature, and topic labels like "Paste Your URL", "History", or "Pricing" are not questions, and a page where almost none of the headings are real questions scores 0 or 1 out of 8. There is no partial credit for intent. The words on the page are the whole of the evidence.
That makes this one of the few checks you can fix with certainty rather than guesswork, because the rewrite is mechanical: take the section that answers something and title it with the question a reader would actually type. "Shipping" becomes "How long does shipping take?". "Refunds" becomes "How do I get a refund?". The section below it does not have to change at all.
Who does this affect, and why does it matter?
Answer engines are affected first. ChatGPT search, Copilot, Perplexity, Google AI Overviews and AI Mode, and Claude all build a response from passages, and a heading is the label on the passage beneath it. People ask these systems questions, so a heading already in question form is the closest match a page can offer.
Human readers benefit from the same change, and screen reader users benefit most concretely. Assistive technology can list a page's headings and jump between them, so a heading list made of questions works as a table of contents someone can navigate by intent rather than by guessing what "Overview" contains. Skim-readers on a phone get the same advantage for the same reason.
Extractability is the 30-point pillar that measures how easily an engine can lift a clean, self-contained answer from your page, and AccessKnight scores this check at 8 of those points — tied with direct answers as the largest in it. The row passes at 7 or above, warns from 4 to 6, and fails below 4. It is not a gating check, so a weak score costs points without capping your headline score.
The mechanic behind the points is passage matching. Retrieval systems chunk a page by its structure and rank those chunks against a query, and the heading travels with its chunk as a label. Phrasing the heading as the question makes that label match the query's shape. It removes an obstacle to extraction; it does not promise that any engine will use the passage.
Two honest limits, because this check is easy to overcorrect. First, question headings on their own do not make a page extractable — a question heading with a meandering answer under it earns these 8 points and loses the 8 for direct answers, which the same scan grades separately. Second, the check rewards a ratio, not a purge. Rewriting a product name or a legal notice into an interrogative to farm a point makes the page worse to read.
How do I fix it?
Rewrite the headings that sit above an answer into the question a reader would type, leaving the section beneath each one untouched. "Shipping" becomes "How long does shipping take?". Leave the headings that genuinely are not questions alone — the check rewards a ratio, not a purge — then re-scan and read the row.
- List every h1, h2, h3, and h4 on the page — including the ones in your site header, navigation, and footer, because this check reads headings from the whole document, not just the main content area.
- Mark each heading that is a literal question ending in a question mark. Headings that only imply a question, such as "Shipping", "Pricing", or "History", score nothing.
- Rewrite each section heading that answers something into the question a reader would type, keeping the wording close to how they would phrase it out loud.
- Leave the headings that genuinely are not questions alone — a hero h1, a product name, a legal section label — since the check rewards a ratio rather than a clean sweep.
- Reduce the number of non-question headings your template emits sitewide if the footer or mega-menu contributes a long tail of one-word labels, since those are counted too.
- Re-scan with "Also grade AI extractability" ticked and confirm the "Question-style headings" row reports 7 or 8 of 8.
<h2>Shipping</h2>
<p>Standard orders arrive in three to five business days.</p>
<h2>Returns</h2>
<p>You have 30 days from delivery to start a return.</p><h2>How long does shipping take?</h2>
<p>Standard orders arrive in three to five business days.</p>
<h2>How do I return an item?</h2>
<p>You have 30 days from delivery to start a return.</p>How does AccessKnight detect this?
This check is graded by a model, not matched by a rule. AccessKnight collects every h1 through h4 in the document — including headings inside your site header, navigation, and footer, which are not filtered out — keeps the first 40, and sends them to Claude with a sample of the page text. The instruction is explicit: award points only for headings literally phrased as a question, with no credit for one that could merely be reworded into it. Temperature is 0, so the same page should score the same, but a model grade is a judgment and can move between runs. The row passes at 7 of 8 and warns from 4 to 6.
What strings mean this?
If one of these strings is in front of you right now — in a config file, an HTTP header, an AccessKnight report, or another checker’s output — it is describing the failure explained on this page.
- AccessKnight report
- “Question-style headings”
- AccessKnight report
- “Graded from page headings.”
- AccessKnight report
- “Extractability (AEO)”
- AccessKnight dashboard
- “Also grade AI extractability”
Frequently asked questions
Do all of my headings need to be questions?
Does a heading count if it implies a question but has no question mark?
Do headings in my site header and footer affect this score?
Do h5 and h6 headings count toward this check?
Will question headings on their own raise my Extractability score much?
Where does this come from?
Every external claim on this page traces to one of the documents below — the vendors’ and standards bodies’ own documentation. What AccessKnight adds is the scan, not the standard.