Most headings are short and not phrased as questions

The check takes the headings in the page content (H1 to H6, not just subheadings) and asks one thing about each: is it shaped like a topic? A heading counts as topic-shaped if it either ends in a question mark or has at least 3 words. Headings in navigation, header and footer have not been counted since 2026-09-19: a footer link directory (“Explore”, “Products”, “Company”) is not a section heading, and on production data it decided the verdict for six findings out of twelve. What counts as site chrome lives in config/content-scope.json; an H1 is never treated as chrome, because it is the page title wherever it sits.

The finding is raised when the share of the rest exceeds 50 %, and only if the page has at least 3 content headings and at least 3 of them fall below the word threshold. All four numbers live in config/thresholds.json under rule_engine.STR-02 as min_words_for_topic_shaped_heading: 3, max_generic_heading_ratio: 0.5, min_content_headings: 3 and min_generic_headings: 3, and they are marked "_status": "PROPOSAL" (our proposal pending revision, not numbers from the product specification). The last two were added on 2026-09-19: a ratio over a single heading says nothing (one heading flips it from 0 to 1), and the finding was talking about “headings” in the plural anyway.

This is the crudest heuristic in the entire audit, and we will be straight with you: it counts words, not quality. A heading reading “How this all works” passes; “Pricing” does not, even though the second may be entirely right where it stands. The actual quality of your content is judged by a language model elsewhere in the report, not by this check.

What the check does not do:

• It does not understand meaning. “Blue cat sings” has three words and passes. “Pricing” has one and does not.
• It ignores the type of page. On a contact page, short headings (“Address”, “Phone”) are correct and a finding there is a false alarm.
• It counts neither characters nor language. In German, a compound that would be three words in English counts as one, so a page may get the finding undeservedly. We measured this on 2026-09-19 across six sets of the same torumata.com pages in seven languages (68 headings per language): the share of headings below the threshold ranged from 0.21 in Portuguese to 0.35 in German. The threshold therefore stays the same for every language and the finding text admits the spread; tuning seven thresholds from a single site would be fitting noise.
• It does not judge whether a heading matches what sits underneath it.

The finding is page-level and informational, not a warning. It costs you no score at all. That is not an accident: for a check that can only count words, deducting score would not be honest.

We have no evidence that a particular heading shape improves your chances of being cited. No operator of a language model declares it, and we know of no measurement comparing a site with question-shaped headings against an otherwise identical one without them.

The reasoning behind the recommendation is sensible and worth spelling out, so you know what we actually think: a user asks an assistant in a full sentence (“how much does X cost”, “how do I do Y”), and a page whose heading names that sentence literally pairs with it more readily than a page headed “Pricing”. It is a hypothesis that makes sense. It is not proof.

What is documented in this area concerns content rather than headings. The field's only peer-reviewed experiment (Princeton, KDD 2024, 10,000 queries) measured what actually lifted visibility in generated answers: quotations in the text +41 %, concrete statistics +33 %, links to sources +28 %. Heading shape was not among the factors measured.

The practical conclusion, even though it undercuts our own recommendation: if you have an hour to spend on your chances of being cited, rewriting headings is a weaker bet than adding a concrete number and a link to a source in the text beneath them. That second one is measured by AUTH-01, and there is an experiment behind it.

And once more about the heuristic itself: the 50 % threshold and the three-word line are our estimate, recorded in configuration as a proposal pending revision. If the finding strikes you as unfair, there is a fair chance you are right. That is why the finding text prints the measured ratio next to the threshold, so you can recompute it yourself.

First look at what kind of page it is. On a contact page, a price list or a technical specification, short headings are correct and this finding has nothing to fix there. Ignore it.

Where it does help is content that answers questions: articles, guides, FAQs, service descriptions. There, rewording a heading is cheap work with a visible result: a reader can tell from the outline where to find what.

Three rules that hold up in practice:

• A heading should answer a question, not merely hint at it. “Delivery” is a label. “How much is delivery” is a question. “Delivery is free over £50” is an answer, and that is the best of the three.
• Do not turn every heading into a question. A page where every subheading is a question reads worse than one with a couple. The goal is clarity, not satisfying our heuristic.
• Do not overdo the length. A twelve-word heading fits neither your navigation, nor a table of contents, nor a search result. Three to seven words is a sensible range.

And if you conclude that the headings stay as they are, that is a legitimate outcome. This finding is informational precisely because we expect you to override it sometimes.