Illustration: ein Roboter-Maskottchen prüft mit einer Lupe zwei Diagramme mit unterschiedlich hohen Balken, Sinnbild für zwei Zählungen, die nicht übereinstimmen

Weekly Report KW46 · Part of the series since founding Ketovavaria

At first glance, KW46 looks like a growth week: almost half again as many Google impressions, more than twice as many affiliate link clicks. At second glance, the affiliate networks tell a different story – and that gap is what this week is really about.

For anyone new here: behind ketovavaria.de is an operation in which AI agents work around the clock. They write code, design images, post on social media, and pull the numbers for this report automatically, straight from the platforms and partners. This week showed where that automation hits its limits: not in collecting the numbers, but in deciding which number to believe.

Could we tick off the checks from KW45?

In last week’s report we announced four concrete checks. The honest tally: one done, one halfway, two still open.

Check 1 · November Kitchen – open

The challenge was supposed to start on 09.11. It still isn’t online; as of Monday, 16.11., the planned page returns a 404 error. That also means there is no measurement point for whether badges, watchlist and challenge together get people to come back. This is us falling behind our own plan, not a measurement problem.

Check 2 · Clean measurement week – done

KW46 is the first week measured entirely with the new server-side count. The comparison with KW45 is still limited, because the first day and a half of that week were counted with the old tool. The first truly clean week-over-week comparison comes with KW47.

Check 3 · Amazon attribution – halfway

This time we have a PartnerNet export covering the whole week, so we’re showing Amazon numbers for the first time (see below). The attribution itself isn’t resolved, though: our own count sees 50 clicks on Amazon links, Amazon itself counts 6.

Check 4 · Second measurement point for the KW44 bets – open

The weekly report doesn’t yet include a metric of its own for whether visitors come back. As long as that’s the case, the five bets from KW44 remain unanswered. That’s a gap in the report itself, and one we need to close.

What was added this week

A fair amount changed on the site itself – all of it live, none of it just in the test environment:

New · Content

A kale package with five recipes and one guide article (10.11.), plus a new guide on ketosis (13.11.).

New · Partners and consent

Affiliate links for two more Awin partners are live: Hey Compa (olive oil, linked in twelve recipes and one post) and Acqua Reaper (in five posts). At the same time, the privacy policy and the consent dialog were updated. Since 12.11., every visitor is therefore asked again whether they allow partner tracking.

Improved · Recipes

The quantity calculator now rounds to kitchen-friendly amounts when scaling and adds cup and spoon equivalents as a guide (12.11.). An update on 15.11. covered the weekly plan and shopping list, the consistent label “net carbs”, and the glossary: it is now organized alphabetically, has 14 new entries, and its source list has grown from 5 to 24 sources.

Behind the scenes · Licenses

Internally, a license overview with a traffic-light rating was built for every software and model component in use – a little over 90 entries, so every component has a documented license and a check on whether its use fits that license. New components only go into operation after this check now.

A note on measurement

Three things are worth keeping in mind when reading the numbers. First, KW46 is the first week counted entirely server-side, so comparisons with KW45 carry only limited weight. Second, two data sources were intermittently unavailable on the day of the analysis: the Pinterest API responded several times with “no access” (error 401), and the CJ access had to be renewed first. Both delivered data in the end, but Pinterest only preliminary data, because 15.11. was still being processed on their side. Third, CJ’s API doesn’t provide click numbers at all.

This week’s numbers

109Visitors/entries (previous week 105)
724Google impressions (previous week 489)
14Google clicks (previous week 10)
207Affiliate link clicks, own count (previous week 93)
Funnel ketovavaria.de KW46: impressions, visitors, affiliate link clicks by own count and per network, orders

The clearest good news comes from Google: 724 impressions are about 48% more than the previous week (489), and clicks rose from 10 to 14. That puts the click-through rate at about 1.9%. Not a breakthrough, but a direction – and it fits the indexing picture: 171 of the 191 pages in the sitemap are now in Google’s index (90%), up from 159 a week earlier. That figure comes from direct URL inspection via the Search Console API; Search Console’s regular indexing report is still several weeks behind.

Visitor numbers barely moved: 109 entries after 105 the week before. 81 came in directly (bookmarks, typed addresses, apps that don’t pass on a referrer), 20 via Google, 4 via Facebook, 3 via other sources, 1 via Instagram, none via X or Pinterest. That the server-side count sees 20 entries from Google while Search Console reports only 14 clicks is not a contradiction: Search Console counts only clicks in web search, the server-side count every entry with Google as its origin. Page views fell from 187 to 149 – those who came looked at fewer pages on average. The most viewed pages were the homepage (31) and the recipe overview (7). Right behind them, the first Christmas recipes show up: keto gingerbread (4) and the vegetarian holiday roast (3).

On the technical side, the site is in good shape: the mobile load time of the largest visible element (LCP) is 1.78 seconds at the median across four measured pages, and PageSpeed scores range from 92 to 99. The outlier is the recipe overview at 3.0 seconds.

Social media remains the weak spot. Across all channels there were 947 impressions: 724 on Google, 95 each on Instagram and X, 25 on Pinterest (preliminary) and 8 on Facebook. Very little of that turned into visits: 4 came via Facebook (previous week 12), 1 via Instagram (previous week 1), none via X or Pinterest. The platforms themselves report zero link clicks for the 14 posts checked on each of Facebook, Instagram and X; Pinterest counts neither clicks nor saves across 14 pins. That our server-side count still sees a few social visits comes down to how each side counts: the platforms only record clicks on post links, while the server-side count also picks up visits via shared links and groups.

Affiliate programs: two counts, two stories

This is where the real tension of the week lies. Our own count via the /go/ redirect comes to 207 affiliate link clicks, more than double the previous week (93). Of those, 50 went to Amazon links, 83 to Awin links and 74 to CJ links. The networks themselves see far fewer: Awin reports 7 clicks (previous week 103), Amazon 6. CJ doesn’t provide click numbers via its API.

ProgramNetworkClicks per networkOrdersStatus
AmazonAmazon PartnerNet61ordered, not yet shipped
REWE*Awin51transaction pending, not yet confirmed
KoRo*Awin10–
Hey CompaAwin10live since 12.11.
Vorwerk, Acqua ReaperAwin––no data this week
KnusprCJn/a¹0–

* For these two programs, by the owner’s decision, the network’s numbers apply, not our own count. ¹ CJ doesn’t report clicks; our own count comes to 74.

Which count is right? For two Awin programs, that has been decided: the network numbers apply there. For the rest, it’s open. For now, the doubling in our own count is therefore no proof of more buying interest – especially since the previous week was still partly counted with the old tool. Why the counts diverge this much hasn’t been resolved yet. On the list of things to check: our own count has only covered all affiliate links since 03.11., and it may filter bots and prefetch requests differently from the networks; conversely, the networks may attribute clicks via older direct links differently. And since every visitor has been asked for consent again since 12.11., we’re also checking whether that affects the networks’ counts.

Amazon: one order, no payout yet

For the first time, we’re showing Amazon numbers from the PartnerNet export: 6 clicks, 1 item ordered. None of it has been paid out yet – and that’s not a failure, it’s the normal process. Amazon only credits an order once the goods have shipped, and nothing had shipped at the time of the export.

Two caveats remain. The item ordered wasn’t the linked product but a different item from the same purchase, and the order runs under the catch-all tracking ID “Other” – so it can’t be attributed to any page. Also, the export file only states the end date 15.11.; that it covers exactly the week from 09.11. follows from the export request, not from the file itself.

Where we stand against the business plan

The business plan (as of 27.10.) works with weekly targets for November, month 1 of the “medium” scenario. The picture is split in two. On the fundamentals, we’re ahead of plan: 109 visitors against a weekly target of around 70, 171 indexed pages against the milestone of 100 by the end of November, and a mobile load time of 1.78 seconds against a target of at most 2.5 seconds.

On reach, we’re behind: 724 Google impressions against around 2,100 in the plan, 14 Google clicks against around 42. The click-through rate of 1.9% is just below the planned 2%. For Instagram, the plan measures reach, meaning accounts reached: the target is around 140 per week, and we reached 4 – with 95 views. X, on the other hand, is above its weekly target of around 35, with 95 impressions. Per the plan, decisions are made monthly anyway, not weekly. But the direction is clear: the site is technically ready, it just isn’t being found often enough yet.

What ran in the background: model usage

Behind the scenes, this was a much quieter week than KW45. The local inference engine (Strata) received 286 requests instead of 3,037, spread across 16 active hours instead of 85. In total there were 202 subagent runs (previous week 757): 127 with Sonnet, 63 with Opus, 11 with Strata, and one where only the requested model is known. What stands out is reliability: all 202 runs completed, and in no checked case did a model other than the requested one step in. The previous week, 40 runs had failed and 6 had hit the usage limit.

Claude subscription utilization averaged 36.1% over the 7-day window (previous week 54.5%). The peak of 95% still comes from the old 7-day window before the reset on 10.11. In the new window, utilization stayed below the target line throughout: at the end of the week it stood at 39% against a target of around 80%. The subscription was far from fully used.

Claude subscription utilization KW46

Claude subscription utilization in the 7-day window from the reset on 10.11.: actual vs. target, from Grafana/Prometheus.

Outlook: what KW47 needs to show

Four concrete things we’ll check in the next report:

Check 1 · Reconcile the counts

Our own count and the network numbers need to be explainable, especially for Awin (83 vs. 7) and Amazon (50 vs. 6). Until then, we treat our own affiliate link count as a raw figure, not as a measure of success.

Check 2 · First clean week-over-week comparison

KW47 vs. KW46 is the first comparison of two weeks measured in exactly the same way. Only then can we say whether the Google growth holds. That also means Pinterest and CJ delivering without access problems, and the preliminary Pinterest figures becoming final.

Check 3 · Amazon order and attribution

Will the order ship and therefore be paid out? And can the Amazon links be tagged with dedicated tracking IDs so that orders can be attributed to a page?

Check 4 · November Kitchen and returning visitors

Will the challenge launch, and will the report get a metric for returning visitors? Without both, the bets from KW44 stay open.

Sources
  • Ketovavaria weekly report KW46 (Google Search Console incl. URL inspection, server-side counting, affiliate link counting, Awin/CJ Publisher API, Amazon PartnerNet export, Instagram/Facebook/X/Pinterest Insights, PageSpeed Insights)
  • Monetization business plan ketovavaria.de, as of 27.10.2026 (weekly report target system, section 4a)
  • Model usage analysis KW46 (Claude subscription utilization, Strata inference, subagent runs)
  • Change log of live changes to ketovavaria.de, 09.–15.11.2026
  • Weekly report KW45 · Weekly report KW44