Breadth · multi-page audit 1 credit × subpage count

Subpages.

One URL, but most product judgement happens across several pages — pricing, signup, checkout, help, terms. After the entry page is audited, the LLM identifies the most important internal links and runs the same battery on each. Findings that show up on every page collapse into one entry with cross-page evidence.

What it does pickSubpages()

Once the entry-page audit finishes, the LLM is shown a compact list of the visible internal links + the entry-page screenshot. It picks the N most consequential targets (default 2, configurable 0–20) — usually the pages a real evaluator would click next: /pricing, /signup, /help, /checkout, /about, top product pages. Each pick comes with a reason field so you know why the agent thought that page mattered.

Each subpage then runs through the same pipeline: capture (screenshot, HTML, console, network), analyse, accessibility audit, exploratory steps. A pruned version of the report doc per subpage gets attached to report.subpages[]. A post-pass deduplication step groups identical findings across pages into report.issueDuplicates: e.g. "footer 'Privacy Policy' link is 404 — seen on home, pricing, help (3 pages)."

What it audits

Pricing pages — table clarity, hidden fees, plan-comparison ergonomics, CTA prominence, FAQs
Signup & login — form validation, error message quality, password-visibility toggle, social-login affordances
Checkout — shipping/billing forms, promo-code application, address autofill, order-confirmation clarity
Help / docs — search functionality, navigation, content completeness, deep-link IDs, last-updated stamps
Product pages — image gallery, descriptions, reviews, variant selectors, stock status
Legal / footer pages — privacy, terms, cookie policy, security page (Diana's persona checks these)
About / team — credibility signals, recent activity, social links that actually work

Cross-page deduplication

The same bug surfacing on 5 pages is one bug to fix, not 5 separate findings. Post-run we walk every issue across every subpage and group by canonical signature (category + selector + similar title). The output:

{
  "title": "Footer 'Privacy' link returns 404",
  "category": "reliability",
  "severity": "medium",
  "affectedPages": [
    "/",
    "/pricing",
    "/help",
    "/checkout",
    "/about"
  ],
  "firstSeenOn": "/",
  "evidence": "All 5 pages render `<a href=/legal/privacy>` in the footer.\nThat URL responds 404. Likely a deleted route that the shared footer\ncomponent still links to."
}

Coverage

Default

2 subpages per audit (entry + 2 = 3 total). Free tier capped at 2; paid tiers raise the ceiling to 20.

Selection

LLM picks by importance + visibility from the entry-page link graph. Each pick includes a one-line reason.

Per-subpage depth

Same pipeline as the entry page — analysis, a11y, responsive captures. Personas + flows + exploratory configurable per submit.

Dedup

Cross-page identical findings collapse into issueDuplicates with the list of affected pages.

Not a crawler — picks from links visible on the entry page. Deep parts of the site (5 navigation clicks in) require submitting those URLs directly.
The LLM may pick boring pages (terms-of-service, privacy) over interesting ones. Submit the URLs you specifically want audited if the auto-pick misses.
Each subpage adds ~30–60s of wall-clock time + 1 page-credit. A 20-subpage audit is ~10 minutes and 20 credits — bound your run.
No state-carry between subpages — each opens in a clean context. Flows that require multi-step state must be defined as Custom Tests.
Sites with infinite-scroll feeds or single-page-app routing may report fewer real "subpages" — the LLM sees rendered HTML links, not framework routes.

Sample subpage entry