SecondDoor

Knowledge

Methodology

How a site is classified, how each door is walked, how the ten dimensions are scored, and what we refuse to guess at.

Site classes

The class describes the business, not the page. The scan first resolves who owns the domain: a blog or docs property is recognised from its own metadata, classified from its root business's homepage, and never ranked as a company of its own, so github.blog reads as GitHub's blog rather than a publisher missing a pricing page. The class then decides which funnel is walked. A retail storefront is walked to a payable cart. A software business is walked to an opened account. Dimensions that do not apply to the pages actually scanned, pricing on a blog for one, are marked not applicable rather than scored. When the evidence does not clear the bar, the class is unclassified, because a coin flip stated with confidence is worse than admitting the model was not clear.

Software businessScoring against the software funnel: trial, plans, API access.
Retail storefrontScoring against the shopper funnel: find a product, see its price, reach a checkout.
Software and storefrontBoth a product catalog and a software funnel. Scoring against both.
PublisherScoring as a publisher: the product is the content itself.
Service businessScoring as a service business: the funnel ends in a booking or a call.
UnclassifiedThe business model was not clear from the homepage. Scoring against the general rubric.

The two walks

Both walks start at the same URL and pursue the same goal. They differ only in what they are allowed to use.

Human door

  • /A real browser, javascript on, cold cache, 1280 by 800
  • /Clicks and typing only, at human speed, with human patience
  • /Every step screenshotted, every field recorded
  • /Scored on friction, not taste

Agent door

  • /HTTP fetches only, no javascript, no rendering
  • /Declared as an agent user agent, and it respects robots.txt
  • /Reads structured data, feeds, and any documented API
  • /Scored on whether the goal is reachable at all

Basic scan

Scored from deterministic checks: what the site publishes for machines, and one automated pass over the signup path.

Deep scan

Scored by two agents that walked the site, one over HTTP with no JavaScript and one in a real browser, with cited evidence for every dimension.

The ten dimensions

Six on the agent door, four on the human door. Each is scored 0 to 100. The human score is the mean of the four human dimensions, the agent score the mean of the six agent ones, the composite the mean of the two doors, and the Door Gap is the human score minus the agent score. None of those numbers is written by a language model.

What we refuse to guess at

Anything the scan could not reach is recorded as not measured, and not measured is never counted as zero. A wall is a fact worth reporting; a guess about what sits behind it is not.

Percentiles are computed in SQL against other scans of the same depth, and only once at least 12 exist. Below that a report says there is not enough to compare with, because a percentile drawn from three scans is noise wearing a statistic.

Computed, not judged

Eight of the ten dimensions are calculated directly from what the scan found. A file either exists or it does not, and asking a model to score that only adds noise. The remaining two are anchored to values computed from the same evidence, and the written score may move at most ten points from its anchor. Every issue sentence then passes a deterministic linter: a negative the probes did not establish, or a claim that contradicts a recorded pass, is rejected and replaced with one scoped to exactly what was checked. A refusal is never reported as an absence: a wall in front of a path is recorded as a wall.

The sample

The corpus is not yet large enough to publish sample statistics. They appear here, live, once at least 12 sites are ranked.

Scan a site