Locul went from 0 clicks to 197 in three months. Here are the loops behind it.

There is no old screenshot to show here, because there was nothing to screenshot. Search Console has rows for locul.ai from 14 June 2026. The earliest complete 28 days, 14 Jun to 11 Jul 2026, cover the 23 days before the first article and the first five days after it. That window is the “then”.

From 14 June 2026 to 6 October 2026, 115 days, locul.ai earned 267 clicks from 35,151 impressions in Google Search. The first click was recorded on 15 July 2026, eight days after the first article. The best single day so far was 22 September 2026, with 15 clicks.

Switch the chart to average position and you will see the part most new-site stories leave out. Through August the position got worse, not better: 28.5 in July, then 44.6 in August. Impressions were climbing at the same time, from 1,252 to 9,382. The articles were being shown for many more searches, most of them far down the results.

Then it turned. In the week of 31 August the weekly average was 17, and in September the month averaged 9.7. Clicks followed: 6 in July, 42 in August, 174 in September.

One article shows the same curve on its own. This is the page that now earns the most clicks on the site, published on 28 July 2026, on four single days.

If you had judged that article on 21 August you would have called it a failure. Two weeks later it was on page one. If you run a new blog, the practical lesson is to wait before you decide anything about an article. This one needed about five weeks. Our study of 633 new articles has the wider numbers on that wait.

Locul sits in a crowded space: note-taking apps, second brain methods, and AI memory. The obvious searches to want are the category ones, such as “second brain app”. Three months in, those are not where the clicks come from.

What ranks on page one is narrower. People who already use a named tool and are choosing between two features, or two tools, are the readers the site reaches today.

Two more things in the data surprised us. The article with the most clicks is not about notes at all. It is a guide to choosing a local language model, a neighbouring topic that the same readers care about. And the third most clicked page on the whole site is a German translation.

The German versions are worth a line of their own. 51 of the 103 English articles have one. In the last 28 days, 43 German pages earned 46 clicks from 1,852 impressions at an average position of 7.3. That is a click-through rate of about 2.5%, against 0.9% for the site as a whole. Fewer people search in German, and far fewer pages compete for them.

Locul's Search Console is connected to Murkuz once. After that the work happens inside Claude. Murkuz hands Claude the real rows for the site, with the window and the source attached, and Claude does the reading: which queries, which pages, what changed, what to do next. We draft the articles and the edits in the same conversation.

The drafts then go into the Murkuz CMS, which renders Locul's blog. On 9 October 2026 it held 103 published English articles for the site. That part of Murkuz is an add-on we enable on request, and it is covered further down. The loops themselves need only the connector.

The method is the same one we use on LigoSocial, a site with about ten times the clicks. Three questions, asked again and again of fresh data. On a new site the weight sits on the first loop, because there is little to merge and little to tighten until pages rank. We show all three below, including the one where we have no result to report yet.

A new site has no queries to read, so this loop cannot start from Search Console. It starts from keyword research. As rows arrive, it switches to the site's own data: searches the site is shown for with no page that answers them directly.

Between 7 July 2026 and 28 September 2026 we published 103 English articles: 57 in July, 43 in August and 3 in September. Figure 2 shows what followed: impressions first, clicks about two months behind.

So 145 of the site's 203 clicks in that window landed on an article. Of the rest, 34 landed on the home page, and searches for the name Locul account for 29 clicks. The most recent article is “Best MCP Servers in 2026: Picks by Job for Claude Code, Cursor and Codex”, 3,362 words, published on 28 September 2026. It is too new to judge, for the reason Figure 4 gives.

Publish a hundred articles in two months and some of them will answer the same search. Google then has to choose between your own pages, and often ranks neither as well as it would rank one clear page. The second loop finds those collisions.

Here is one we have already acted on. Two articles covered AI plugins for Obsidian, with nearly the same address. In the last 28 days both were on page one, a place apart.

The second address now redirects to the first. The last day Search Console recorded an impression for the old address was 3 October 2026. In the four days after that, the surviving page was shown 54, 37, 58, 54 times a day at average positions of 6.4, 4.8, 4.4, 5.2.

That is four days of data, and it proves nothing yet. In the week before, the same page averaged between 5.1 and 6.0 a day, so the early readings are close to where it already was. We will read this pair again at the next pass, and that later number is the one that counts.

The loop also has open work. On 9 October 2026 the connector found 3 queries where two or more Locul pages appeared for the same search in the last 28 days, counting queries with at least 50 combined impressions. One of the three is the site's own name, where several pages showing together is normal. This is one of the other two.

Two list articles about second brain apps, both far down, and Google shows the weaker one more often. This is the category search from Figure 6. One strong page has a better chance there than two thin ones, so this pair is next on the list.

The third loop takes a page that already ranks, lists every query it ranks for, and looks for the ones it is weak on. Here is the site's top article. Over the last 28 days it earned 36 clicks from 5,480 impressions at an average position of 6.8:

/blog/best-local-llm

The pattern is specific. The page is on page one for a named amount of memory: 8GB for coding, 32GB of RAM. It is at position 45.9 for 16GB. That tier is a search of its own, and the page does not answer it well enough yet. It is also on page three for the two shortest phrasings, “local llm” and “best local llm”, which carry the most impressions in the list.

The edit is small. Give 16GB a section of its own, and answer the short question in the first paragraph. Leave the rest alone, because the rest is working.

We do not have a finished example of this loop on Locul. No page here has yet had its weak queries worked in and then been measured two to four weeks later, so there is no before-and-after position to show. The 16GB edit above is the first candidate. Site-wide, the connector found 3 queries at position 15 or better with at least 100 impressions in 28 days. On a site this young, that list is short.

None of this depends on a person remembering to do it. Murkuz gives you the skill file for it: a prompt that Claude runs as a loop of its own, on your machine or wherever you run your agents. Connect your sites, hand Claude the prompt, and it works through all three loops and comes back with one prioritised list.

Here is a starter version you can paste into Claude today, with the Murkuz connector attached.