Overwrite Protection & a Reliability Sweep

Features

• Overwrite Protection on WordPress Updates — If a post or page on your self-hosted WordPress site has been edited directly since we last published to it, publishing an update now stops instead of replacing those edits, and asks you to bring them into your draft first. In the Omni Agent, the agent explains the conflict and replaces the live edits only once you confirm; when the live page can't be checked, publishing works as before.

Improvements

• Agency Command Center Pages — The agency Command Center is now three pages under its own sidebar (Portfolio, Members & seats, and Portfolio reports) instead of one long scrolling page. The sidebar collapses to icons and remembers your choice.
• Prompt Dialogs — Creating or generating prompts opens a wide two-pane dialog, with the prompt or brief you are writing on the left and its settings on the right. The model count it shows matches the models your plan will run.
• Agent Composer — The Omni Agent's message box is one surface centered in the conversation, with a single send button that becomes a stop button while a reply is streaming.
• Security and Privacy Pass — Stricter browser security headers; links in agent results and sources are checked before they become clickable; source icons load from the source's own site instead of a third-party service; your WordPress application password no longer travels in a web address; app connection links work only once; and credentials are scrubbed from our error reports.

Fixes

• Gemini Citations Counted — Gemini returns its sources as redirect links that name the real site only in their title, so Gemini's citations of your site scored zero and Source Influence listed a Google redirect address as the source of every grounded Gemini answer. Citations are now matched on the site they name, and Share of Voice counts the citations a model returns alongside its answer, not only links written into the text.
• Category Overall Score — The Overall figure on Prompt Tracking's categories counted every category you don't use as a zero, so a workspace scoring 90 in its only category read 10. It now averages only the categories that have data.
• Share of Voice Accuracy — Brand names guessed from a domain no longer lose letters (shopify.com read as "Ify") or turn into ordinary words that then counted as mentions. A snapshot window now covers exactly its length instead of one extra day, the trend compares against the previous window of the same length and market, and "analysed across N prompts" counts prompts rather than prompts times models.
• Every Model Gets Measured — While a prompt's cached answer was still valid, rotation kept picking the same model, so a new prompt could measure one model for up to two weeks. Rotation now moves on after a cached or failed attempt.
• Fewer Timed-Out Answers — Time limits now match how long models actually take, for Gemini, where nearly a quarter of executions were timing out, for the open models, and for scheduled Grok runs with search. Claude executions, which failed for about a day on August 21–22 after a provider library update, run again.
• Scheduled Runs and Reports — Each deploy restarted the twelve-hour wait before the next scheduled prompt run, so on busy days prompts could go more than a day without running; the first run now comes fifteen minutes after start-up, and daily jobs such as the Google Analytics sync no longer skip a day when the server is briefly busy. The list of scheduled reports, which had failed to load since scheduling launched, now loads.
• Publishing Updates the Right Page — Updating an existing page could match the wrong one (/team/about matched /about), and on sites with more than about 100 posts, and on Shopify pages, it created a duplicate instead. Matching now uses the full path across the whole site, and publishing to self-hosted WordPress no longer identifies itself as a crawler, so host firewalls stop blocking it.
• Annual Billing Kept — Changing plan on an annual subscription moved the new plan onto monthly billing, and the billing page showed monthly prices to annual subscribers. Plan changes now keep your billing cycle, and the billing page shows your real price and cadence.
• Integration Tool-Call Allowance — Pro and Premium accounts were capped at 100 integration tool calls a month instead of 500 and 2,000. They now get the full allowance.
• Agency Organizations — Members of a paying agency organization were sent to the plan picker, shown upgrade prompts for Prompt Research, discovered domains and Sources, and charged a second time against their personal allowance for work the agency pool had already paid for. All three are fixed, and a plan change can no longer alter a different subscription from the one on screen.
• Usage and Checkout Screens — The upgrade comparison shows plan limits instead of dashes, and "Update card" on a locked feature opens billing instead of failing. Usage shows the cost actually recorded for each run and the limits actually enforced on your account, and no longer counts client-workspace work toward your personal usage. Checkout no longer promises a free trial that isn't there, and switching between sign-in and sign-up keeps the plan you picked.
• Long Lists Are Complete — My Prompts stopped at 100 prompts, Opportunities at 20, and the Sources leaderboard and a prompt's Opportunities tab at 50, while report history failed after its third page. Each now loads the rest, and My Prompts takes its prompt and slot counts from your account rather than from the rows it had loaded.
• Agent Documents Kept Intact — Cleanup of agent-written documents was removing code blocks, merging paragraphs, stripping indentation and deleting sentences that began "Here are" or "This is", and the trimmed text was what got saved. It now removes only a leading preamble and a trailing sign-off, and a similar filter that deleted sentences from chat answers was narrowed the same way.
• Agent Conversation Reliability — A deep-research turn or an unreadable tool result no longer blanks the conversation, and the end of a reply is no longer dropped when the stream closes. Research steps show the query being searched, code and diffs highlight without stray markup, and a failed rename, delete or archive now says so instead of quietly reverting.
• Uploads, Titles and Exports — Content uploads could fail during extraction; they process again, files are read by their actual type, and an interrupted upload is marked failed instead of staying "processing". Generated titles are no longer cut at 500 characters, and exported files keep non-English characters in their names.
• Switching Workspaces and Apps — Switching between organizations now clears the previous one's data immediately, and switching or leaving a workspace keeps you in the app instead of landing on the marketing site. An expired app connection can be reconnected from Settings as well as Apps. Google Analytics connections are no longer dropped by a temporary sign-in refresh error or broken by two simultaneous connect attempts, and a failed Google sign-in or disconnect now says so.
• Pages That Failed to Load — Brand Perception no longer errors when a model's answer is malformed, the score trend no longer errors when nothing in range was analysed, and a cancelled crawl marks the domain failed instead of leaving it on "analyzing" with recrawl disabled. Pages that failed silently now show an error, and a sign-in hiccup while accepting an invitation no longer signs you out.
• Dates and Smaller Measurement Fixes — Trend axes, Share of Voice dates and the Journeys chart showed the previous day to viewers west of UTC; they now show the day the data belongs to. Brand clarity ratings such as 8.5/10 are read whole, Google Analytics page correlation only matches pages on the same site, and fix measurements keep the prompt wording identical before and after the change.