Three days ago I added a fifth child page to every vehicle on our car site. The pattern is simple: a model-year page gets five children — specs, safety, maintenance, owner’s manual, problems — and each one gets its own URL, title, and FAQ block.
Then I sampled forty cars and measured how much those children actually differ from each other. Two of the five page types came back as the same page, over and over, with the model name swapped.
The shape of the site
The sitemap index has 96 files. All 96 fetched, 24,333 URLs, none missing.
| Depth | URLs | What lives there |
|---|---|---|
| 1 | 103 | home, brand index, OBD index, legal |
| 2 | 1,213 | brand and model roots |
| 3 | 3,862 | model years, 3,831 of them with a 4-digit year |
| 4 | 19,155 | 5 sub-pages × 3,831 |
So 94% of the site is one tree. Brand, model, year, five children. Nothing else.
What I measured
Forty model-year pages at a fixed stride through the sorted URL list, so brands scatter instead of clustering. For each, all six URLs — parent plus five children — one request at a time, 240 total, every one 200.
For each page: the block, tags stripped, inline JSON-LD removed, then 5-gram Jaccard against every other page of the same type. 780 pairs per type. Word counts after the same cleaning.
| Page type | Mean words | Most common exact length | Similarity to another car’s same page |
|---|---|---|---|
| Model year (parent) | 686 | 699 (4 of 40) | 35.7% |
| specs | 386 | 361 (2 of 40) | 33.0% |
| safety | 362 | 356 (28 of 40) | 73.0% |
| maintenance | 565 | 573 (3 of 40) | 47.2% |
| owner’s manual | 326 | 321 (29 of 40) | 59.7% |
| problems | 834 | 313 (3 of 40) | 20.0% |
Twenty-eight of forty safety pages are exactly 356 words. Twenty-nine of forty owner’s manual pages are exactly 321 words. Forty vehicles spanning four decades and two continents produce the same length, because the length was never derived from the vehicle.
Three findings, one of them a retraction
The children are not the parent cut into pieces. I was wrong about that.
The obvious suspicion with page multiplication is that the sub-pages are excerpts of the parent, re-served to fill the index. It’s cheap and it’s common and it’s easy to prove.
5-gram overlap between each child and its parent: safety 1.4%, problems 2.1%, maintenance 2.9%, specs 5.4%, owner’s manual 6.9%. Under a twentieth of shared phrasing. These are separate pages written about the same car, not one page sliced five ways.
That kills my first hypothesis. It also means the duplication, where it exists, came from somewhere less obvious.
Stripping the JSON-LD is what exposed it.
My first word count included everything inside , and contains an inline JSON-LD breadcrumb block — a couple of hundred words of URLs and schema that differ on every page.
The number I first wrote for safety pages was 568 words. Clean is 356. And clean is where 28 of 40 pages collapse onto the same integer.
The extra 200 words had been acting as camouflage: identical bodies with different breadcrumbs produce 40 different word counts, and a 40-page duplicate looks like a 40-page variety pack. The padding was not there to hide anything. It hid the sameness anyway.
The site admits the gap in one place and sells it in another.
Every parent page carries a link with the exact words “maintenance schedule and fluid capacities”. Every maintenance page contains zero figures in quarts or litres.
| Check | Count |
|---|---|
| Parent pages linking to “…fluid capacities” | 40 / 40 |
| Pages of 240 containing a quarts/litres figure | 0 |
| Parent pages with a “this table is empty on purpose” section | 40 / 40 |
| Safety pages with any NCAP rating table | 1 / 40 |
| Safety pages naming a star value | 1 / 40 |
| Owner’s manual pages with a link to the manufacturer | 40 / 40 |
| Problems pages with at least one recall row | 34 / 40 |
The parent page is candid. There’s a section headed “What is still missing on this page” saying fluid capacities and torque values aren’t verified and are left blank rather than guessed. That is the right policy and I’d defend it.
But the anchor text above it still says capacities, three thousand eight hundred and thirty-one times, and the destination has none. Good disclosure does not cancel a misleading label two inches above it.
The safety pages have the same structure. One of forty had a table. The rest explain that NHTSA returned no record — plausible for low-volume cars, and equally plausible if the data join silently returns nothing.
Edges of this measurement
I never called the NHTSA API, so “no rating exists” and “we didn’t find the rating” are indistinguishable in my data. Thirty-nine of forty supports both stories and I won’t pick the nicer one.
I didn’t measure crawl behaviour or impressions. I can see 73% similarity; I cannot see what a search engine does with 19,155 pages at that similarity, and that number is not in this dataset.
The problems pages are the only child type whose length moves with the data — 313 to 2,113 words, recall and complaint counts differ per vehicle — and that’s the one I’d keep if someone made me pick. That’s a hunch about usefulness, not a measurement, and I’m labelling it as one.
Disclosure
carspecdata.com is ours: specs, EPA fuel economy, NHTSA recalls, and manufacturer manual links, with the source named on the page. The 240 requests were made against it from a laptop in Luoyang, one at a time, over about twenty minutes.
The thing that gets worse
Six days ago our ledger recorded 3,198 model-year pages. Today it’s 3,831. Six hundred thirty-three new pages in six days, each one immediately growing five children.
Two of the five templates don’t change. The number of pages using them goes up every day we leave it alone.