Score only the landing page and you get 41% of funnels wrong

Almost every published landing page study measures one page. That is the reasonable thing to do, because one page is what a crawler can cheaply get. It is also, we can now show, the thing that changes the answer.

By Ben Peetermans · 2026-08-09 · UXXRAY Funnel Census, n=1450

We scored the same funnels two ways. Same sites, same detectors, same scorer. The only variable was how many steps the scorer was allowed to see.

The census: what the launch class of 2026 looks like

The corpus grew from 556 funnels to 1450 in about ninety minutes, seeded from nine startup and AI launch directories. These are not curated brands. They are the products that shipped this month.

Band	Funnels	Share
Sparse	708	48.8%
Thin	371	25.6%
Partial	253	17.4%
Strong	101	7.0%
Excellent	17	1.2%

Median 40/100. Average 39.9. Average walk depth 2.31 pages.

Read the top row carefully. Nearly half of everything shipping right now surfaces so few provable conversion tactics that it lands in the bottom band. Not "bad design", which we do not measure. Just: the things that reliably move a visitor toward a decision are, in most cases, absent from the page.

And 17 funnels out of 1450, 1.2%, do it well enough to reach the top band.

The finding: depth changes the grade

Here is the experiment, because the design matters more than the number.

Take every funnel we walked at least two pages deep. Score it normally. Then score the same funnel again using only the evidence recorded on its entry page, and compare it to itself. No second crawl, no second site, no estimate. The crawler already records which step every piece of evidence came from, so the entry-only score is derived from its own receipts.

41%

n = 255 funnels, each scored both ways.

	Full walk	Entry page only
Median score	45	30
Average score	46.2	32.3
Tactics detected	4.32	3.06

• 61% of funnels score lower when only the homepage is read.
• 41% (105 of 255) land in a different band.
• Median gap: 10 points.

Where the grades move:

Shift	Funnels
Sparse → Thin	45
Thin → Partial	18
Sparse → Partial	14
Sparse → Strong	8
Partial → Strong	7
Thin → Strong	6
Thin → Excellent	3
Strong → Excellent	2
Sparse → Excellent	1
Partial → Excellent	1

33 funnels move two or more bands. One goes from Sparse, the bottom, to Excellent, the top, entirely on the strength of pages a homepage crawler never opens.

This is not a rounding argument. A band is the number you would publish, quote, or put in a report. On four funnels in ten, reading only the front page gets it wrong.

Why: the tactics live past the homepage

The mechanism is unglamorous. Conversion tactics are not evenly distributed across a funnel, and the ones carrying the most weight tend to sit exactly where a single-page crawl cannot see them.

Share of funnels where the tactic appears only past the entry page:

Tactic	Invisible to a homepage-only crawl
Us-vs-them comparison	68%
Lead-capture form	64%
Proof above the fold	52%
Pricing tiers	45%
On-page review wall	37%
FAQ / objection handling	35%
Scarcity / urgency copy	35%
Anchor / strikethrough pricing	33%
Usage / adoption proof	22%
Free trial / freemium	16%

Forty-five percent of pricing tiers are not on the homepage. Pricing is the single most decision-shaped thing a funnel does, and nearly half of it sits one click out of frame.

For context, the tactics most commonly present anywhere in the corpus: FAQ / objection handling (181), Trusted-by logo wall (159), Free trial / freemium (153), Pricing tiers (103), Demo or docs entry (82), Press mentions (68), Lead-capture form (64), Proof above the fold (62).

Triangulation: two methods, one answer

The closest comparable work is roast.page's State of Landing Pages 2026: 1,000+ pages, scored with AI vision analysis, explicitly landing pages only. Their median is 44. Our full funnel median is 45.

Two studies with no shared code, no shared rubric, and opposite philosophies about whether a model belongs in the measurement path, landing one point apart on the same population. That is worth more than either number alone. It suggests the low-40s median is a property of the web right now, not an artifact of how either of us chose to measure.

One comparison we are deliberately not making. Our entry-only median is 30 against their 44, and it would be easy to present that as evidence that AI scoring runs generous. We will not, because it is not like-for-like: their rubric assigns 20% to First Impression and 10% to Visual Design, categories we do not score at all because we cannot extract them deterministically. Different denominators. The honest comparison is the full-funnel one, and it agrees.

What broke

Tripling a dataset is a stress test whether or not you meant it to be. Three things failed, and all three are more interesting than the growth.

A regex that could not read a comma. Our count-consistency gate checks that every published figure agrees across every page. It matched numbers with \d+. The moment the corpus crossed 1,000, pages began rendering "1,415" and the gate captured "415", compared it against the correctly parsed 1415, and declared a contradiction. Six consecutive deploys failed. The pages were right the whole time. The checker could not read them. That bug had existed for as long as the project had, and was unreachable until the corpus grew enough to trip it.

A missing structured-data field. Google Search Console flagged our Dataset markup for a missing description. The per-report Dataset node had always carried one. A second, corpus-level node beside it did not, because it was written independently in three files and only one copy got the field. Now there is one definition and three callers.

A scoring fix we wrote and threw away. Two detected tactics, scarcity copy and press mentions, were deliberately unscored: detected and displayed, worth zero points. That produced a real contradiction on some pages, "1 tactic detected" sitting beside "0/100". The obvious repair is to score them. We did, and the golden fixtures failed immediately: scores went down across the board, 95 to 85 on one fixture. The score is coverage of the tactics applicable to a funnel type, so adding a scorable tactic raises the denominator for every site of that type, and every site lacking it loses points for a change it had nothing to do with. Reverted. The contradiction is real and is a presentation problem to solve, not an arithmetic one.

And the first version of this study was wrong. The initial comparison put walked funnels against unwalked ones: 44.2 against 21.0, a 23-point gap. That is confounded, because a site the crawler cannot walk past may simply be a thin one-page site. The rebuilt experiment then reconstructed step data incorrectly and reproduced the crawler's own stored score on only 5% of funnels, so it was discarded before producing a single published number. The version above validates at 98%: recomputing each full score from stored evidence matches what the crawler recorded, on 255 of 259 funnels. The honest controlled gap is 10 points, not 23. Less than half what the first cut suggested.

What is biased about this data

The corpus is now roughly 74% SaaS, and e-commerce has fallen to about 11%. That is self-inflicted. Launch directories list AI and developer tools, so seeding from nine of them pulled the sample hard toward one vertical.

Every number here should be read as the launch class of 2026, skewed to SaaS and AI tooling, not as the web. The band distribution in particular would likely look different for a retail-weighted sample, where tactics like anchor pricing and review walls are near-universal rather than occasional.

We name it because a sampling bias you disclose is a caveat, and one you do not is a finding somebody else gets to publish about you.

Methodology

Deterministic extraction, no model in the loop. Every fact is DOM plus computed CSS plus string matching. If a claim is not provably on the page, it is not recorded. There is no LLM anywhere in the measurement path, which is what makes a crawl reproducible rather than merely repeatable.

Blocked is never zero. When a site refuses the crawler the row is labelled blocked. Inventing a zero for a site we could not read would be a fabricated measurement, and a fabricated zero for a well-known brand would discredit the instrument faster than an absent row ever could.

The score is coverage, not quality. It measures how many of the conversion tactics applicable to that funnel type are provably present, normalised to 100. A SaaS funnel is never penalised for lacking e-commerce tactics. It is not a judgment about whether a page is good.

Scores freeze at crawl time, stamped with the engine that produced them. This snapshot is Roentgen 1.20.1, 2026-08-09. This run: 796 previously unseen domains attempted, 709 scored (89%), 42 blocked (5%), 45 errored (6%).

Reproduce it

Every figure is derivable from public endpoints. No key, no signup.

• Corpus totals: GET https://api.uxxray.com/api/v1/corpus/stats
• The full ranked corpus: GET https://api.uxxray.com/api/v1/league?limit=5000
• One funnel, every measured fact: GET https://api.uxxray.com/api/v1/teardown/{slug}
• Machine-readable brief: GET https://api.uxxray.com/api/v1/teardown/{slug}/brief
• Tactic definitions: GET https://api.uxxray.com/api/v1/patterns

The depth experiment is reproducible from teardown/{slug} alone: each detected tactic carries evidence strings prefixed with the step it fired on, so filtering to the entry step reproduces the entry-only score without a second crawl.

FAQ

What does a score of 40 actually mean?
That roughly 40% of the conversion tactics applicable to that funnel type were provably present on the pages we walked. It is a coverage measurement, not a quality verdict, and a low score can be a deliberate choice.

Why is Excellent so rare?
The top band needs broad coverage across all six categories: offer clarity, action path, trust, friction, value and risk reversal. Most funnels are strong in two or three and empty in the rest. Seventeen of 1,450 covered enough of all six.

Does this mean landing page studies are wrong?
No. They measure the landing page, which is what they say they measure. The finding here is narrower: if the conclusion is about the funnel, reading one page changes the band 41% of the time.

Is any of this AI-generated or AI-scored?
No. The measurement path is deterministic: DOM plus computed CSS plus string matching, with no model in the loop. That constraint is the reason two independent methods can be compared at all.