It's Monday, Search Console is open, and forty minutes later you've looked at every report and decided nothing. That is the normal outcome of reading data without a fixed set of questions. A weekly review works when the inputs are the same every week, the questions are the same every week, and the output is always the same shape: three actions, each tied to a page and to a number you'll check next Monday.
Why weekly, and why only three actions
Rankings move every day, and most of that movement is noise. Google runs experiments on its results, personalises by location, and shows or withholds an AI Overview for the same query from one day to the next. Look daily and you'll chase ghosts. Look monthly and you'll miss a page that dropped out of the index weeks ago. A week is long enough for real movement to show and short enough to act on it.
Three actions rather than a list, for two reasons. A list of twelve gets none done. And three changes in a week are few enough that next week's data can tell you which one worked, which is how you learn what moves rankings on your own site. A review that ends with "everything needs work" has failed.
The time cap exists for the same reason: the review decides the work, it isn't the work.
Step one: pull last week's numbers from Search Console
Search Console's Performance report is the only source that tells you which queries showed your pages and which of those impressions became clicks. Google's own help for the Performance report explains its compare mode; use it to set the last seven days against the seven before. Its data runs a little behind real time, so compare complete weeks rather than ending on today.
Export the Queries table and the Pages table, then paste both into Claude with the same three questions every week:
- Which queries gained or lost the most clicks, and which pages those clicks landed on?
- Which pages gained impressions without gaining clicks?
- Which queries with a healthy number of impressions sit in a window just off page one, say average positions six to twenty?
Ask it to cite the rows it used and to say what the data can't tell you. The connector can't read Search Console, so pasting is how this data reaches the assistant.
Each question targets a different mechanism. Clicks are the money. Impressions rising while clicks stay flat means Google is showing the page more and searchers are choosing someone else: a title-and-snippet problem. Queries just off page one sit on pages Google already half-trusts, where a small change moves more than a new page would.
Step two: verify with live lookups before you trust a trend
Search Console's average position is averaged across every impression, every location and every variant of the query, so a reported position of nine can be a mix of threes and fifteens. Before you act on it, check the live position for the queries that matter. With the connector connected, ask Claude to run a rank check for your domain, the keyword and your market. Pick five, not fifty: the connector is rate limited like the free tools, and five tells you whether the trend is real. Without the connector, the free Google rank checker does the same thing one keyword at a time.
For the pages from question two, ask for each page's metadata: title, meta description, H1 and headings. More often than not the title says nothing a searcher would choose over the competing results.
For your two or three most commercial queries, ask whether an AI Overview or AI Mode answer appears and whether your domain is cited. Log it weekly. "No AI surface for this query" is a real reading, and a query that flips to a surface citing a competitor is worth knowing the week it happens.
The rule that keeps you honest: a Search Console move confirmed by a live lookup is a signal, and a move without confirmation is a note for next week.
Step three: turn it into three actions
Now give Claude the job it's good at, with a prompt that forces specifics:
"Based on these exports and the live lookups, give me three actions for this week. For each one: the page, the exact change, why you expect it to help, and the number I should look at next Monday to know whether it did. Rank them by expected impact against effort. Don't give me more than three."
Read the answer as a proposal, not a verdict. Reject any action that doesn't name a page. Reject "improve the content" unless it says what to add and why the competing pages have it. Reject anything built on a number the assistant didn't see. What survives goes into a running document with a check date.
Next Monday's review starts by reading last week's three and checking their numbers before you look at anything new. After a couple of months that document is the most honest record you'll own of what moves rankings on your site, because every change was one of three and you saw what followed.
What to take away
- Weekly is the cadence that separates real movement from daily noise while still leaving time to act on it.
- Search Console tells you what was shown and clicked; a live rank check tells you where a page stands right now, and you need both before you trust a trend.
- The review ends with exactly three actions, each naming a page, a change and the number that will show whether it worked.
- A running log of past actions and their results becomes your own evidence of what moves rankings on your site.
Next
When a week's data shows more problems than three actions can hold, use The 30-minute quick-wins audit to rank them.