AI Insights turns recorded Accessibility Test findings into reviewable actions. Analytics remains the source of facts; AI contributes hypotheses and verification steps.
Create and review a report
- Open AI Insights with a Premium or Agency plan and a connected Freemius license.
- In Create new report, choose Desktop or Mobile and open Choose test period. Select the latest saved state, the last 7, 30, or 90 days, or a custom date range. Each page contributes only its latest saved result within that period, including scores of zero. Before generation, the dataset summary shows analyzed pages, matching results, the actual test dates, and pages without a matching result.
- Review the allowance beside the action, choose Create report, and confirm. An unchanged snapshot in the same language reuses its saved report without a new AI request. The optional checkbox explicitly requests a new interpretation and uses your allowance.
- Read the recommended action, hypothesis and verification step. View findings switches within the saved report to locally recorded pages, rule mappings, element examples and links to test results. Back to recommendation and the browser Back action restore the overview.
- Review the cause before changing code. Marking an action done is a browser-local checklist, not proof that a finding was fixed. Repeat the Accessibility Test and any necessary manual checks.
The UI uses your WordPress admin language (German, English, Spanish or French). Old reports remain readable during generation. You can select saved reports independently of a running request. Errors never turn an earlier report into a successful new report.
Evidence and limits
The snapshot covers at most 500 saved results in the selected viewport and test period and 50,000 rule-result outcomes. It uses no more than one result per page: the latest result inside the chosen period. It does not represent untested pages or certify compliance. The report discloses coverage, the actual included date range, and omitted data. It interprets at most 20 deterministic candidates; recurring groups are a bounded sample.
Priority follows recorded severity, affected pages and then element-rule occurrences. Every recommendation explicitly displays its rank as Priority 1, Priority 2, and so on; severity remains a separate fact. A failed rule-result pair and an affected element are different units. WCAG mappings come from the plugin registry, not AI. Recurrence does not prove a common template. Comparable rule history, dependencies and remediation effort are currently unknown. No score forecasts or time estimates are generated. The model may abstain; partial and empty interpretations are explicitly marked.
Saved evidence is compared with the current Analytics snapshot and analysis version when loaded. There is no fixed age threshold: a report is only marked as outdated when its underlying saved evidence or analysis contract no longer matches. The notice distinguishes changed saved test results from a changed analysis or rule version. A report’s creation date is not the date of its test snapshot, so a report created today can already use an earlier analysis state after a rule change. Legacy reports receive a separate notice because their claims lack versioned evidence links.
Privacy and allowance
Page addresses and result links stay in WordPress. AI receives opaque evidence, candidate and page identifiers, rules, severity and counts. It does not receive page HTML, DOM snippets, authors, comments or page text. The proxy retains the established authentication, queue and quota controls; site metadata needed for authentication is not included in the model prompt.
Allowance information is loaded separately. Refreshing, browsing reports and printing do not create AI requests. An intentional new interpretation uses the applicable plan allowance.
Print and retention
Print / Save as PDF uses your browser’s print dialog. It prints the same validated actions, full evidence and scope, followed by technical provenance. A4 or Letter is selected from the locale; review your browser’s paper settings. Content flows across pages without fixed-height clipping.
The ten latest successfully saved reports are retained. A new report must be inserted successfully before older reports are removed. Switch report opens the saved-report history over the current content without moving the page. Every history row has its own named delete action and requires confirmation. Export a report first if you need to keep it.
After confirmation, the creation button displays the current phase and remains locked against duplicate requests. A durable status card distinguishes preparation, AI analysis, and saving while the existing report remains readable. Toasters confirm start, reuse, successful completion, or a terminal error. Retry appears only after a repeatable processing failure. Missing test data, exhausted allowance, or unavailable cloud access instead links to the relevant corrective workflow. Reloading restores an active request without replaying its start notice.
The dismissible Quick start uses the familiar Dashboard pattern with a Rocket identifier, guiding heading, three connected step cards, and a highlighted Recommended start. It explains the path from saved test results through prioritization to human verification. Configure report collapses Quick start, moves the creation card into view, and focuses its heading. Its visibility preference applies only to your browser and can always be reversed. The white creation card uses a report icon and presents viewport, dataset, and report creation as an ordered three-step flow. Arrows point right between the steps in wide containers and down when the steps stack. Reserved helper areas prevent dataset and allowance updates from resizing the card. Choose test period opens an anchored dropdown on desktop and the shared filter sheet on phones. The separate saved-report surface labels the selected report and shows its creation time, stored viewport, stored test period and stored data status. Switch report opens the custom history popover on desktop and the shared filter sheet on phones; refresh and print remain available there. A stale-data notice links to a new Accessibility Test, while an outdated analysis links to report creation. Your next step highlights the first recommendation with a short rule-based title. Prioritized actions keeps every recommendation in one vertical work plan. Only one item opens at a time and shows Observation with the distinct AI hypothesis, Recommended action, How to verify, the findings action, and the browser-local completion checkbox. The progress indicator and All, Open, and Completed filters use that same local checklist state; an empty filter result is stated explicitly. Rank, severity, affected pages, and rules stay visible as separate facts on every row. The findings view remains part of the report, has its own URL, and changes into labelled cards on narrow surfaces instead of requiring a horizontally scrolling table. Help in the page header explains the difference from Analytics, the recommended review flow, privacy and limitations. Missing stored element details are explicitly shown as unavailable; this does not mean that the page has no affected elements.