Xavier is the agent built into PageOptimized. It is not a chatbot bolted onto a dashboard: it holds the same tools the app runs on, works inside one project at a time, and can change things — so it is worth knowing where the edges are before you rely on it.
The short version
Xavier can look things up, diagnose why something happened, and write into your project. It cannot browse the open web, act on more than one project at a time, or spend money without you.
You can decide which Xavier skills are allowed for you and for each project in Project settings -> Xavier. See Xavier voice input and skill controls.
What it can do
Answer from your own data
It reads your Search Console history, rank tracker, site audits, saved keywords, content tasks, backlinks and AI visibility tracking directly. Ask "which pages lost the most clicks last month" and it queries the stored history rather than guessing.
Diagnose, not just describe
This is the part most worth using. Four questions have a dedicated chain behind them, each joining several sources rather than reading one and inferring:
- "Why did traffic drop?" — decomposed per query into ranking loss, demand loss, CTR loss, or pages lost entirely. A drop because fewer people searched is a different problem from a drop because you slid down, and they look identical on a chart.
- "Which pages are decaying?" — decay is measured as a slope over months, not a before-and-after comparison. It separates you slid down while demand held (worth a refresh) from the topic is dying (a rewrite cannot bring back searches that stopped happening).
- "Are pages cannibalizing each other?" — checks whether the winning page actually changes over time, whether your internal anchor text points the phrase at the wrong URL, and whether two titles target the same term.
- "Is this site worth a link from?" — intersects what a domain ranks for and publishes against what you rank for, and tells you which of their pages to ask for a link on.
Ask these as questions ("why did X drop", not "show me X's numbers") and you get a conclusion with its evidence attached.
Run research that costs credits
Keyword research, SERP checks, domain and backlink lookups, local SERPs, brand visibility. It tells you what a run will cost and confirms before large batches.
Change your project
Save keywords, create and update content tasks, add rank-tracking keywords and tracked AI prompts, start a site audit, manage the link outreach queue.
These abilities can be switched off by skill group. If a project blocks a workflow, Xavier is stopped server-side and records the blocked attempt in the activity log.
Take voice input
In supported browsers, the microphone button beside the Xavier composer turns speech into a draft message. You can edit the transcript before sending it. Voice input only changes how you ask; it does not give Xavier any extra permissions.
Read screenshots and PDFs
Paste or drag an image into the chat. A screenshot of a SERP, a chart, a competitor's page or a Search Console view is evidence Xavier can read directly — usually faster than describing it. PDFs work too.
Remember the project between sessions
Xavier keeps two notes per project: durable facts about the site, and a dated research log. That is why it stops re-running research you already paid for and can say "we looked at this three weeks ago, here is what we found".
What it cannot do
Browse the open web
Xavier has no general web browser. It can fetch pages on your own domain, and every other fact comes from your connected data or a paid data provider. If you ask what a blog post published this morning on someone else's site says, it cannot open it. Paste the text or a screenshot instead.
Work across several projects at once
A conversation is bound to one project. It can list your projects and create a new one, but "compare all five of my clients" is not a question it can answer in a single pass — ask per project.
Spend without you
Read-only checks run freely. Paid research is confirmed first, and anything that publishes, deletes, or is visible to a client has to be approved before it runs. Every credit it spends appears in Billing → Usage attributed to the feature it ran.
Read spreadsheets
CSV and Excel attachments are refused rather than accepted and ignored. Xavier reads images and PDFs; a spreadsheet it cannot open would silently become a question about a file it never saw. Paste the rows as text.
Send email or contact anyone
There is no outreach sending, no email, no posting. It will draft the message; a person sends it.
Invent a number
If the data is not there, Xavier is built to say so rather than estimate. "No page has earned a click for this query yet, so which page should own it is a commercial decision, not something the data can settle" is a real answer it gives. Treat a confident number with no source attached as a bug worth reporting.
Getting better answers
Ask why, not what. "Why did /pricing lose traffic" reaches a diagnosis chain. "Show me /pricing's clicks" reaches a table. Both are useful; only one tells you what to do.
Say which page or keyword. Xavier will pick a scope if you do not give one, and it may not be the one you meant.
Let it finish a chain. A real diagnosis runs several tools in sequence. Interrupting after the first one gets you the raw rows without the conclusion.
Push back on vagueness. If an answer reads like general SEO advice, ask "what evidence is that based on?" A diagnosis will name the pages, dates and counts behind it. Advice that cannot will admit it.
When something looks wrong
If Xavier claims it made a change, check the relevant screen — it is instructed never to claim an action a tool did not actually perform, and a mismatch is worth reporting. Use Help & Support → Contact and include the conversation.
Related
- Xavier voice input and skill controls
- People and access