Most SEO reporting fails quietly. Someone exports a dashboard, writes twenty recommendations, and a month later none of them has shipped. The weekly loop fixes this by keeping the output small: three actions, shipped before the next review, each with a number you will come back to.
- Look at the numbersSearch Console, analytics, last week's log
- Pick three actionsThe ones most likely to move a number
- Ship themBefore Friday; done beats perfect
- Check what movedKeep what worked, drop what didn't
The method comes from Noah Kagan, who runs AppSumo. In a thread on X in November 2026, Kagan described how Claude runs AppSumo's SEO every Friday, and why it's set up that way:
"I stopped reading the report. It was too much. So now it picks 3 actions a week, max."
The same thread reported $64,320 from Google search in the previous week, and that ChatGPT named AppSumo in 8 of the 10 questions the agent checks. You don't need that much traffic for the loop to work. On a site with 30 clicks a day it works the same way; you just need the full three weeks before a reading means much, because small numbers jump around more.
A few terms first, since the rest of the module uses them. Google Search Console is Google's free tool that shows how your site appears in Google Search. An impression is one time a link to your site was shown in a result. A click is one visit from that result. Click-through rate (CTR) is clicks divided by impressions. Average position is roughly how high your site appeared, where 1 is the top. Read the growth curve explains how Google calculates each of them.
Why three actions beat a long report
A report describes what happened. The loop ends every week with exactly three things someone will do, and a week later it asks whether they worked.
The limit of three is the point. With no limit, a list grows until nothing on it gets done, and the cheap, low-value items crowd out the ones that matter. With a limit you have to rank, and Kagan's rule for ranking is revenue: pick next week's three actions "ranked by revenue." If your site sells nothing directly, rank by sign-ups, trials or whatever number the business lives on.
The rules of the loop
Most of these rules come from Kagan's thread. The explanations of why they matter are ours.
- Three actions a week, maximum. In Kagan's words: "A report nobody reads ships nothing." The template also caps new rules at two a week.
- Write the number down before you ship. Each action names the page or pages, the metric you expect to move (clicks, CTR, sign-ups) and the direction. If you can't name a number, treat it as a chore rather than a test. Chores are fine; they just don't need three weeks of tracking.
- Every test gets three weeks. "3 readings the same way: PROVEN, it becomes a rule. Split: DROPPED." Weekly numbers move on their own: holidays, news, a Google update, a competitor's launch. One good week proves very little. Three weeks pointing the same way is a pattern you can build on.
- Compare against a control. A control is a similar group of pages you did not change. If the whole site rose 20% this month, your changed pages rising 20% means nothing. Kagan's test shows why this matters. Kagan was sure that titles with price math ("Lifetime Deal: pay once vs Competitor's yearly price") would win. Two weeks in a row, the test pages lost: clicks on the test pages fell 7.5% while clicks on pages with normal titles rose 23.7%. That is a gap of 31.2 percentage points against the change. Looking at the test pages alone, you would have seen a small dip and kept waiting.
- Let the numbers kill your ideas. "It kills my ideas too," Kagan wrote of that test. "One more week and it's dead." The loop is only useful if you accept its verdicts on your favorite ideas.
- Check AI answers too. Kagan's agent asks the same ten questions in ChatGPT every week, such as "Sites like AppSumo," "Is AppSumo legit" and "Best Mailchimp alternative," and records whether AppSumo is named in the answer and whether appsumo.com is cited as a source. That week's reading: named in 8 of 10, cited in 6. "The misses become next week's actions." Get cited by AI assistants covers what to do about a miss.
Google now gives you part of the AI picture directly. Search Console's Generative AI performance report shows impressions for your pages in AI Overviews and AI Mode, and Google says it was rolled out to all websites as of August 31, 2026 (a site with very few AI impressions may not see it yet). Add that impression count to your weekly numbers next to your ten questions.
A test that is proven becomes a rule you apply everywhere without testing again, for example "product titles name the product once, then say what it does." A rule that later stops working goes back into testing.
A sample week
Kagan's agent runs the whole loop on Friday. If you do it by hand, it is easier to spread it out. Here is what a week looks like for the hypothetical invoicing app above, at about 60 clicks a day from Google.
| Day | What you do | Time |
|---|---|---|
| Monday | Read the playbook file. Pull the numbers. Score last week's three actions and update every open test. Pick this week's three. | 45 minutes |
| Tuesday | Ship action 1: the new pricing page title. Add an annotation in Search Console. | 20 minutes |
| Wednesday | Ship action 2: the 301 redirect. Load the old URL yourself to confirm it lands on the right page. | 20 minutes |
| Thursday | Ship action 3: five internal links to the template. Ask the ten AI questions and record named and cited. | 40 minutes |
| Friday | Write one dated line in the change log and rewrite the playbook file. | 15 minutes |
Monday's numbers come from three places. In Search Console, open Performance, then Search results, click the date filter, choose the Compare tab, and set the last 7 days against the 7 days before. Search Console data usually arrives two to three days late, so "last 7 days" means the latest seven complete days the report has. The table then shows a Difference column, and you can sort the Queries and Pages tabs by it to see what moved. When you compare longer ranges, switch the chart to weekly granularity; Google's help notes that this removes the day-of-the-week effect (a Monday compared with a Saturday).
Revenue comes from your analytics: in GA4, look at the Organic Search channel in the Traffic acquisition report. AI visibility comes from your ten questions and the Generative AI report.
The annotation on Tuesday matters more than it looks. In the Performance report you can right-click the chart on a date and add a note of up to 120 characters. A property can hold up to 200 annotations, and they stay visible whatever filters you apply (they are hidden in comparison mode). Three weeks later, when a line moves, you will know what you shipped and when.
An action that didn't ship by Friday does not roll over automatically. It competes again on Monday with everything else.
The playbook file and the tracking table
The loop keeps all of its memory in one file. Kagan calls it "the brain": "Every Friday the agent reads it, pulls Search Console and GA4, does the work, then rewrites the file. Next Friday it starts where it left off. Nothing gets lost between reports." This is the skeleton from the thread:
# SEO playbook (the agent reads and rewrites this every week)
## How this file works
Run every Friday. Pull Search Console + GA4 (last 7 days vs the 7 before).
Score last week's 3 actions. Max 3 new actions. Max 2 new rules.
A test moves to PROVEN or DROPPED after 3 weeks pointing the same way.
## This week's 3 actions
## Last week (clicks, impressions, CTR, organic revenue, AI assistant revenue)
## Open tests (one line per reading, dated)
## PROVEN rules
## TRIAL rules
## Change log (one dated line per run)The prompt in the thread keeps it short: rewrite the file and "keep it under 8,000 characters." A file that grows without limit turns back into the report nobody reads. Old change-log lines and dropped tests are the first things to cut.
Here is the same file filled in for the invoicing example. The numbers are made up to show the format.
## This week's 3 actions
1. Pricing title: "Invoicing app pricing: free for 5 clients | Acme Invoices" (CTR, /pricing)
2. 301 /integrations-old to /integrations (clicks, /integrations)
3. Link /templates/late-payment-reminder from 5 posts (clicks + sign-ups, template)
## Last week (Oct 26 to Nov 1)
Google clicks 412 (+6%) · impressions 21,300 (+9%) · CTR 1.9% · organic sign-ups 31
AI Overviews/AI Mode impressions 840 · AI questions: named 4/10, cited 2/10
## Open tests
A. Product name once in titles: test +9% vs control +2% (Oct 18), +11% vs +4% (Oct 25), ...
## PROVEN rules
(none yet)
## Change log
Oct 30: 3/3 shipped. FAQ block on 8 docs pages, sitemap fix, footer link cleanup.When a test runs across weeks, a table makes the verdict easy to read. Copy this one into the file or a spreadsheet. The rows are made up to show how readings look:
| Test | Pages changed | Control | Week 1 | Week 2 | Week 3 | Verdict |
|---|---|---|---|---|---|---|
| Product name once in titles | 12 product pages | 12 product pages left alone | +9% vs +2% | +11% vs +4% | +6% vs +1% | PROVEN |
| Year in titles ("2026") | 15 guide pages | 15 guide pages left alone | -3% vs +1% | -1% vs +2% | -4% vs 0% | DROPPED |
| FAQ block on docs pages | 8 docs pages | 8 docs pages left alone | +4% vs +1% | -2% vs +3% | +3% vs +2% | DROPPED (split) |
Pick the control before you ship, and make it as similar as you can: the same page type, a similar amount of traffic, the same part of the site. Choosing it afterwards invites you to pick whichever group makes your change look good.
Let an agent run it
The loop suits an AI agent well. The work repeats every week with exact rules, and the file carries the memory from one run to the next. Kagan's setup is Claude with access to Search Console and GA4. Any capable agent works.
- Give it read-only access. In Search Console, go to Settings, then Users and permissions, and add the agent's account (or a Google Cloud service account) as a restricted user, which has view rights on most data. That is enough to read the Performance report through the API.
- Connect the data. Several open-source MCP servers (small connectors that give an agent tools) wrap the Search Console API, and Google publishes an official MCP server for Google Analytics. An agent can also call the Search Console API directly.
- Give it the file and the prompt. Kagan's weekly prompt, word for word: "Read seo-playbook.md. Pull last week from Search Console and GA4. Score the 3 actions. Update every open test with this week's number. Pick next week's 3 actions, ranked by revenue. Rewrite the file. Keep it under 8,000 characters." If you track AI answers, add one line asking it to check your ten questions.
- Decide what it may change. A sensible start is that the agent drafts the actions and a person ships anything that touches the live site, or the agent opens a pull request that someone reviews. Widen that once its suggestions have earned trust.
- Keep it off Google's results pages. Google's spam policies list "scraping results for rank-checking purposes" as machine-generated traffic, which violates the policies. Positions come from the API, not from automated searches.
What the loop catches
What stands out in Kagan's thread is how ordinary the actions are. The three actions in the week the thread describes were fixing product titles that repeated the product's name, a new homepage title, and a redirect for a dead product page. Then this: "A new product launched with its name in the Google title twice. Clicks: 0. It flagged it the next Friday."
A 40-page report rarely gets read closely enough to find that. A loop finds it because it looks at the same few numbers every week, and a new page with impressions but zero clicks stands out. The same goes for a redirect that broke during a deploy, a template change that wiped the titles on a whole section, or an important page that dropped out of Google's index. Each shows up as a line in the Pages table that moved the wrong way since last week.
Use this checklist at the end of every run: