What Do Cited Pages Actually Look Like?
Netsujo Inc. ·
We re-fetched all 137 citation URLs from our earlier study of who AI search cites, and machine-extracted URL depth, title pattern, structured data, H1 count, and meta description. 43% of citations were a company's top page. The dominant title pattern was "what we offer + company name," not "what is X." Structured data (JSON-LD) was present on 60% of the 131 pages successfully fetched — a majority, but 4 in 10 had none at all. Research date: 2026-07-05. Engine: OpenAI Web Search (gpt-4o) cited URLs, observed 2026-07-04. Version: 2026-07-05.s1.
01 / Background and question
Our earlier study found that 94% of AI-search citations went to companies' official domains. The next practical question is: what do those cited official pages actually look like?
Discussion of AIO/AI-search measures often includes claims like "an FAQ schema is essential" or "'what is X' articles get cited" — but there is very little published data that actually counts the structure of pages that were cited.
This study re-fetched every one of the 137 citations observed earlier and machine-extracted URL depth, title pattern, presence and type of structured data, H1 count, and meta description, then classified the results. The extraction script and raw data are published in our repository.
02 / Research design
- Population — all 137 citation URLs observed in our earlier study (r2). No specific company is evaluated by name; only the distribution of structure is examined.
- Execution — re-fetched every URL (2026-07-05) using an explicit user agent, following redirects, with a 12-second timeout and 8-way parallelism.
- Extraction — parsed every application/ld+json block (including @graph) and collected @type values; extracted H1 count, title length, meta description, and html lang with regular expressions.
03 / Main findings
- Looking at URL depth across the 137 citations, 59 (43%) were a company's top page. Deep pages (3 levels or more) numbered only 15 — citations concentrated on "the front door" of a site rather than an article buried deep inside it.
- Classifying the 137 saved page titles, the dominant pattern (110 in the residual bucket, manually confirmed) was a naming style of "what we offer + company name," for example "Business-systems development | XX Corporation." The "what is X" pattern, often recommended for SEO, appeared in only 2.
- Of the 131 pages successfully fetched, 78 (60%) carried JSON-LD structured data. FAQPage schema appeared in only 17 (13%) — 4 in 10 cited pages had no structured data at all.
- Structured-data prevalence varied by page type: 88% (28/32) for article-type pages versus 52% (50/97) for service/company pages.
- Basic on-page hygiene was generally strong: exactly one H1 on 81% of pages (106/131), a meta description present on 92% (121/131), and a median title length of 35.5 characters.
- 6 URLs could not be fetched from our environment (explicit user agent, fetched 2026-07-05): 2 returned 403, 1 returned 429, and 3 failed on certificate issues. Pages we could not fetch under these conditions had nonetheless appeared in AI citations.
04 / Breakdown
| URL depth (all 137) | Count | Share |
|---|---|---|
| Top page (/) | 59 | 43% |
| 1 level (e.g. /service) | 22 | 16% |
| 2 levels | 41 | 30% |
| 3 levels | 11 | 8% |
| 4+ levels | 4 | 3% |
| Page-title pattern (137 saved titles) | Count | Example |
|---|---|---|
| "What we offer + company name" (residual bucket) | 110 | "Business-systems development | XX Corporation" |
| Comparison / recommendation / how-to-choose | 16 | "XX comparison" / "Top picks for XX" |
| Cost / pricing | 9 | "Typical cost of XX" |
| Case study / track record | 3 | "Implementation case studies" |
| "What is X" | 2 | "What is XX" |
Titles matching multiple patterns were counted in each, so the total exceeds 137.
05 / Data
| Fetch results (all 137) | Count |
|---|---|
| Fetched successfully (HTTP 200) | 131 |
| Access denied (403/429) | 3 |
| Failed on certificate issues | 3 |
| Structured data (of 131 successfully fetched) | Count |
|---|---|
| Has JSON-LD | 78 (60%) |
| BreadcrumbList | 54 |
| Organization-type | 51 |
| Article-type | 25 |
| FAQPage | 17 (13%) |
| Product/Service-type | 8 |
This study observes the structural distribution of cited pages; it does not judge whether any listed domain is better or worse.
06 / Analysis
The most common shape AI cites is a top page with a "name and offer" title. 43% of citations going to a company's top page, and the dominant title pattern being "what we offer + company name," pushes back on the idea that "citation measures = mass-producing explainer articles." In this observation, AI referenced pages that plainly state, in a BtoB ordering context, "who is offering what." It is also notable that only 2 of 137 titles used the "what is X" pattern; definitional explainer content does get cited in some contexts (our earlier study found 32 article-type citations), but for ordering/procurement questions, a business's primary information was the centre of gravity.
Structured data is common, but not required. 60% of the 131 successfully fetched pages carried JSON-LD — a majority, but 4 in 10 were cited with no structured data at all. Against the claim that "an FAQ schema is essential," we found FAQPage present on only 13% of pages. This study observes the structural distribution of cited pages and cannot establish whether structured data causes citation, but at minimum, the claim that "no citation without structured data or an FAQ schema" does not fit this data. Article-type pages carried structured data far more often (88%), showing the practice differs by page type.
Basic hygiene was consistently strong — the fundamentals as a foundation. Exactly one H1 (81%), a meta description present (92%), and a title with a median length of 35.5 characters — cited pages were, on the whole, well maintained on the basics. We cannot establish causation, but regardless of any "special trick," this basic-element hygiene was a common thread across the pages we observed. Pages that returned 403 or 429 from our environment had still been cited; fetch access depends on our own user agent and the point in time we fetched, so we do not draw conclusions about AI search's crawl paths — we record the observation as is.
07 / Limitations of this study
- Fetching happened on 2026-07-05, while citation was observed on 2026-07-04; page content may have differed at the time it was cited.
- Client-side-injected structured data cannot be detected by this method.
- This is an observation of structural distribution; it does not establish causation with citation.
08 / Practical implications
- A reasonable starting point for AI citation is making a top page clearly convey, at a glance, "who is offering what." Getting the front door's introduction right may matter more than deep articles.
- The dominant citation-winning title pattern was "what we offer + company name." Article-title conventions common for search-traffic content do not necessarily need to be applied to every page.
- Structured data is a common practice worth implementing, but treating an FAQ schema's presence as the deciding factor for citation does not fit this observation.
- H1, meta description, and a well-formed title were nearly universal among the cited pages we observed — fundamentals first is the suggestion from the data.
09 / Reproduction
- Target — our earlier study's classified data (all 137 citations, URLs and titles already saved).
- Fetch — an explicit user agent, following redirects, 12-second timeout, 8-way parallel, fetched every URL over HTTP.
- Extraction — parsed every application/ld+json block for @type (including @graph); extracted H1 count, title length, meta description, and html lang by regular expression.
- Title-pattern classification — regular-expression rules ("what is X" / cost-pricing / comparison-recommendation / case-study) plus manual review of the residual bucket; titles matching multiple patterns were counted more than once.
- Limitation — fetching (2026-07-05) postdates the citation observation (2026-07-04); this is a structural-distribution observation, not a causal one.
Related: "Who Does AI Search Actually Cite?" and "How Does AI Interpret the Word 'AIO'?"