We Measured 1,650 Sections of Our Own Blog. Structure Alone Was Not Enough.

We parsed all 170 published posts on this blog and measured every one of their 1,650 H2 sections against our own writing standard for AI retrieval. The corpus passes almost all of it: 73.6 percent of sections fall inside the 100 to 300 word band we aim for, 1,648 of 1,650 open with a paragraph rather than a list, and 97.5 percent of headings are question-shaped.

Over the same period the site averaged position 55.7 in Google Search. Those two facts coexist. That does not tell you structure is worthless, and it does not tell you what structure contributed, because a single corpus with no unstructured arm cannot answer that. What it does tell you is that structural compliance is not sufficient on its own, which is worth knowing before you spend a quarter on it. Here is the full measurement, the three places the corpus fails its own rules, and how to run the same audit on your site.

What exactly did we measure?

We pulled every published post from the CMS on 29 October 2026 and split each post's HTML on its H2 boundaries. For each section we recorded four things: the word count of the section body, the element type of the first block after the heading, the word count of the first sentence, and whether the heading begins with an interrogative word or ends with a question mark. For each post we also recorded whether it carries a structured FAQ.

The standard those four map to is ours, not an industry consensus. Our internal writing standard asks for sections of roughly 100 to 300 words that answer their heading in the first sentence or two, and our published guide to making a passage quotable sets a tighter 40 to 90 words for an individual answer block. We are not aware of a study establishing any of those numbers, which is part of the reason for measuring what following them actually produces.

What does full structural compliance look like?

It looks like a tight distribution with a long right tail. Across 1,650 sections the mean is 242.7 words and the median is 217, with the quartiles at 166 and 296.

Section length	Sections	Share
Under 100 words	39	2.4%
100 to 300 words	1,214	73.6%
301 to 500 words	349	21.2%
Over 500 words	48	2.9%

The other measures came back close to saturated. Only 2 sections out of 1,650 open with something other than a paragraph, one with a table and one with a list. The median first sentence under a heading is 14 words, the 90th percentile is 25, and only 5 sections anywhere in the corpus open with a sentence longer than 40 words. 1,609 headings are question-shaped. All 170 posts carry FAQs in the CMS field that generates FAQPage markup, 902 question and answer pairs in total.

Citation habits are similar. The corpus carries 550 external links, with a median of 3 per post and only 2 posts carrying none. 286 of those links point at developers.google.com and another 60 at support.google.com, so 63 percent of all outbound citations go to Google's own documentation.

Where does the corpus fail its own rules?

Three places, and two of them are the same failure: a section that grew past the point where it can be lifted as a single answer.

48 sections run past 500 words. The longest is 1,204 words, under the heading "Which internal link checker tool fits your site?" in our internal link checker comparison. Seven of the twelve longest sections are tool or agency comparisons, where the heading asks one question and the body answers it nine times. That is a useful page shape for a reader scrolling a comparison and a poor one for a retriever looking for a single passage.

753 of 902 FAQ answers run past 60 words. The median structured FAQ answer here is 69 words and the longest is 102. Our own standard says answer each FAQ in under 60 words. The corpus misses that on 83.5 percent of its FAQ entries, which is the largest compliance gap we found.

97.5 percent question-shaped headings is a template, not a style. At that rate the pattern is no longer a choice made per section. Google's guidance on helpful content says that "using generative AI to produce large amounts of text without manual oversight or curation represents little to no effort", and while Google says nothing about heading uniformity, reading a 97.5 percent rate as a sign that the template rather than the writer chose the heading is our inference and a fair one. The fix is not to stop writing question headings. It is to notice when a heading is a question because the section answers one.

What did the structure coincide with?

An average position of 55.7. In the 28 days from 28 September to 25 October 2026, Google Search Console reported 6,213 impressions, 18 clicks and a click-through rate of 0.29 percent for growthhasten.com.

Read that against the measurements above and the honest conclusion is narrow: a corpus that satisfies essentially every structural rule we set for it sits, on average, five pages deep in Google. Structure is not sufficient. We cannot tell you what it contributed, positively or negatively, because we have no version of this blog written without it.

We can tell you what we did not measure. We did not measure authority, so the common explanation that links and domain strength are the real constraint is a hypothesis here and not a finding. Testing it would mean comparing referring domains against the pages that outrank us, which is a different study. We also did not measure AI citations, so nothing here says what the structure bought in AI engines. That is a separate exercise with its own method, covered in how to track and measure AI search visibility. The related constraint on the publishing side is in what 149 posts earned us.

So is structural work worth doing at all?

Yes, and it is important to be precise about which part is evidence and which is reasoning.

The evidence: the most-cited primary research in this area, GEO: Generative Engine Optimization by Aggarwal and colleagues, accepted to KDD 2024, reports measurable lifts in a source's share of a generated answer from adding quotations, statistics and citations, in the range of roughly 28 to 41 percent depending on the modification. Note what that is about: what is in the text, not how the page is sliced.

The reasoning: engines retrieve passages, not pages, so a section that only makes sense after the three above it is harder to lift intact. That follows from how retrieval-augmented systems work. It is not a measured result for this site or any other, and you should treat anyone presenting it as one with suspicion.

The honest summary is that structural work is cheap, low-risk, good for readers and defensible on first principles. It is table stakes. On the evidence available it is not a lever by itself.

How do you run this audit on your own site?

You need your posts as HTML and about an hour. The procedure, with no tool required beyond whatever can read your CMS:

1. Split each post on its H2 boundaries and count the words between each heading and the next.
2. Record the first element type after each heading. Anything that is not a paragraph means the section opens with a list, table or sub-heading and has no standalone answer at the top.
3. Count the words in the first sentence of each section. Above roughly 35 words, the opening sentence is carrying a caveat rather than an answer.
4. Count your structured FAQ answers and flag anything over 60 words.
5. Plot the section-length distribution rather than the mean. The mean here, 242.7, looks healthy and hides the 48 sections over 500 words that are the actual problem.

Then fix the tail and leave the middle alone. Rewriting sections that are already between 100 and 300 words is motion, not work.

What this study cannot tell you

It is one site, measured once, with no control corpus. Compliance and poor position coexist here. Neither caused the other as far as this data can show, and any reading stronger than that is not supported.

We measured the HTML our CMS stores, not everything a crawler receives. Two spot checks on 29 October 2026 cover part of that gap. Requesting /blog/audit-false-positives with GPTBot, ClaudeBot and PerplexityBot user agents returned HTTP 200 and an identical 174,047 byte payload in each case, so AI crawlers can fetch these pages. The same URL returns a FAQPage block with four Question entities in the served HTML, which confirms the FAQ field renders as markup on that page. We did not verify the rendered output of all 170. If you run this audit on a site that renders content client-side, measure what the crawler gets rather than what the CMS holds.

How this analysis was produced

GrowthHasten's SEO agent, credited in the byline above, ran this analysis. It fetched all 170 published posts through the CMS API on 29 October 2026, parsed the stored HTML of each, and computed the section and FAQ statistics above. The post count was read from the CMS on the same day; our own content inventory recorded 166 on 28 October, before four posts published since. Search Console figures are from the growthhasten.com property for 28 September to 25 October 2026, the window Search Console returns for a 28-day request after its reporting delay. Every number here is a count or a percentile from those sources.