What we check. What we can prove.

SiteProof inspects one public page in an automated browser at desktop 1440 × 900 and phone 390 × 844. Each report records the capture time, evidence, completed checks and gaps. A phone-sized browser is not a physical-device test.

Measurements first

We inspect page structure, visible links and images, accessibility, search and sharing metadata, structured-data syntax, public response headers, initial loading and uncaught script errors. We fetch the standard robots.txt and sitemap.xml paths and check up to 12 selected links on the same site. We do not crawl the entire site, sign in, submit forms or complete purchases.

Accessibility uses applicable automated axe checks tagged for WCAG A and AA through 2.2. Some checks need human judgment, and many requirements cannot be tested automatically. A decorative image with empty alternative text is not automatically an error.

Every outcome stays visible

• Passed: this specific automated check passed within the recorded scope.
• Confirmed issue: a measured failure with supporting evidence.
• Needs review: an observation whose meaning depends on the page’s intent, such as an intentional search exclusion or alternate canonical address.
• Unavailable: the check was blocked, inconclusive or missing data. It is not a pass.
• Not applicable: the rule did not apply to the captured page.

Only passed, failed and measured review outcomes count as completed checks. These counts are not a website quality score, a certification, or a prediction of sales.

Loading measurements with context

We record an initial-load sample before injecting accessibility analysis or scrolling. The sample includes largest visible content paint, layout shifts, observed long-task time and network transfer. Measurements use our browser and network proxy without phone CPU or network throttling. They describe this visit, not your real visitors’ experience.

A large download can deserve review while still rendering quickly. Missing measurements remain unavailable. We do not claim a Lighthouse score, real-user Core Web Vitals or interaction latency from this capture. The 2.5-second paint and 0.1 layout-shift references come from Google’s loading guidance and layout stability guidance; a single lab visit cannot establish the field result.

AI review is not included in the current report

Current inspections run automated technical checks. AI design and messaging review is off. The following describes the separate review process if it becomes available.

The AI review uses sampled top, middle and bottom screenshots, visible copy, actions, forms and measured findings. It reviews offer clarity, calls to action, trust, visual hierarchy, phone usability, forms and missing decision information. A second pass checks its draft against the same evidence. Findings must cite recorded evidence; exact quotations are checked against captured text. Unknown evidence references and mismatched quotations are rejected; a valid citation alone does not prove an interpretation is correct.

AI judgments can still be wrong. The report labels interpretations, confidence and missing coverage. We do not invent business facts, testimonials, guaranteed results or revenue-loss estimates. If an inspection requires AI and that review cannot finish, its paid checkout stays closed.

Design opinions stay separate

We do not grade a website’s taste. AI concerns must identify a visitor task, an observable obstacle and supporting evidence. Optional style suggestions appear separately and never increase the issue count or enter the repair queue. A dark theme, unusual typography or missing public prices is not automatically a defect.

Customer AI reviews require a qualified model and a current release approval. Qualification includes repeated reviews of good layouts and known defects, a human check of the actual outputs, and separate real-site checks. A model’s confidence and a second pass are not independent proof of accuracy. When the approval is missing or expired, new AI sales stay closed.

Fixes you can verify

The repair queue sorts by severity, then puts measured failures before context-dependent suggestions at the same severity. Visitor access and usability come before optional metadata at equal severity. Unavailable checks become manual verification tasks. Every action includes evidence, a proposed repair and a verification step. The developer brief carries the same limits. Rechecking the issue after a repair is separate from claiming a business outcome.

Reports and screenshots are private to anyone holding their access link and expire after 30 days. Read the privacy and access details.