OmniAgent Memory, Repository Reading & Honest Metrics

Features

• OmniAgent Memory — A new settings panel shows what OmniAgent knows about your brand, where each memory came from, and lets you review or delete them individually.
• Repository Reading — With GitHub connected, OmniAgent can read files from your repository, so a recommendation can reference the code actually behind a page instead of inferring it from the rendered output.
• Redesigned Agent Chat — The conversation now reads as bubbles, tool calls appear as cards you can expand, and search results and answer sources are listed properly instead of being buried in the reply.
• Background Share of Voice Analysis — Run Analysis now queues immediately and reports progress as it goes, instead of holding the page open until it finishes.
• Visibility Trend Charts — Reports plot daily and rolling visibility as two separate lines across the week, so a single bad day is distinguishable from a real trend.

Improvements

• Reorganized Sidebar — Visibility navigation is grouped around how the work actually flows rather than by internal structure.
• Faster Public Site — omniimpact.net now downloads less than half the code it previously needed before it can render, and the homepage artwork loads immediately rather than waiting on the rest of the page.
• Agent Status Indicator — A live orb replaces the old brain icon and settles when a run ends; a shimmer replaces the bouncing dots while the agent is thinking.
• Screen Reader Support in Chat — Messages announce who is speaking, and the conversation is a single scroll region instead of several nested ones.
• FAQ Editing — FAQ detail is readable at a glance, the edit dialog scrolls on short or zoomed-in windows, and the page picker only offers pages on your own domain.
• Journey Run Clarity — Every run states the task it was graded against, and a withheld overall score now explains why it was withheld instead of showing a bare dash.
• Agent Memory Reliability — Memory updates are written in order, so two changes arriving at once can no longer overwrite each other, and long memories no longer crowd out the rest of the agent's context.

Fixes

• Honest Metrics Across the Board — A broad pass through visibility scoring. Head-to-head comparisons, per-model rates, and weighted statistics now reflect the runs that actually completed. A measured zero is no longer presented the same way as missing data, stale windows are no longer averaged into current figures, and estimated stand-ins that could appear alongside real measurements have been removed.
• Prompt Tracking Rates — Visibility rates now count completed executions only, so a run still in progress no longer drags the number down.
• FAQ Generation Scope — FAQ generation and publishing are restricted to your own domain, including sites on multi-level and private domain suffixes where the owning domain was previously resolved incorrectly.
• FAQ Edits — Fixed expanded rows showing a question as it read before your edit, and fixed a regenerated FAQ being discarded instead of replacing the previous version.
• Publishing Confirmation — The content library now honors the write-confirmation gate before anything reaches a live site.
• Google Analytics Sync — Fixed a sync failure on fractional conversion counts, days with genuine zero traffic being dropped from the series, and reporting windows that did not line up with calendar dates.
• Share of Voice Runs — Interrupted analyses now refund credits exactly once, runs orphaned by a restart are cleaned up automatically, and a run that finished with only partial data is labeled as degraded rather than presented as complete.
• Journey Runs — Cancelling a run now cleans up reliably even if cancelled twice, and a database error during a run can no longer leave it stuck in a running state.
• Journey Scores — A score that cannot be verified against the criteria it was graded on is no longer shown, and score aggregates only combine runs graded under the same version of the weights.
• Agent Answer Reuse — Fixed a case where a cached answer could be reused for a question it did not actually match.
• Agent Sources — Only http and https sources become clickable links.
• Agent Chat Layout — The chat now sizes itself to its container rather than the browser window, fixing a cut-off composer on some screens.
• Prompt Tracking Model — Replaced a model that the provider retired, so tracked prompts continue executing without gaps.
• Integration Errors — A failure coming from a connected provider is now reported as a provider problem rather than as an expired session.
• Source Coverage — Grounding and coverage figures count only completed executions, and no longer include averages from windows that have since gone stale.