September was a small month. The blog had 306 users, and Google Search sent one click. AI assistants opened many articles, but almost none of that activity became visible traffic. The reported cost of running the site was $175.

Audience

The audience numbers remain small. There were 306 users across 335 sessions and 378 pageviews. The average session lasted 31 seconds, and engagement was 26%. Those figures do not support a claim that readers were spending much time with the site.

Five readers arrived from an AI assistant. That is measurable, but it is not substantial traffic. It is also much lower than the number of articles AI assistants opened during the month. An assistant using an article as source material does not necessarily send the person asking the question to the site.

Search and discovery

Google Search recorded one click from 828 impressions, with an average position of 6.3. The position looks respectable in isolation, but the useful outcome was still one visit. Impressions and rankings did not turn into search traffic in September.

Google Discover produced no activity: zero clicks and zero impressions. There is nothing to interpret there. The blog did not appear in Discover during the month.

Googlebot fetched 109 articles. This shows that Google continued to crawl the site, but crawling is not the same as sending readers. September had regular crawler activity without corresponding search traffic.

What the AI pipeline did

AI assistants opened 286 articles to answer someone. This was the largest visible sign that the material was being used outside the site. It resulted in only five identifiable arrivals from AI assistants.

The gap matters. The pipeline can produce articles that machines retrieve without producing an audience that I can see in analytics. Retrieval may have value to the person receiving an answer, but it does not currently create much direct readership for the blog.

The article verdicts were:

• Winner: 71
• Promising: 61
• Too early: 3
• Dead: 1

These labels are internal assessments, not traffic totals. In particular, the large number of winner and promising verdicts sits beside 306 users and one Google Search click. The classifications may be identifying useful material, but they do not yet demonstrate audience growth.

I do not have a recorded pipeline change for September to report. The month’s evidence instead exposes the same practical problem from several directions: articles were fetched, opened and classified positively, while direct reader traffic remained limited. I also do not have a specific next experiment in the September record, so I am not assigning one after the fact.

Newsletter

The newsletter ended September with 12 subscribers. One person joined and nobody left. That is positive movement, but the list is still very small. It is not yet a meaningful distribution channel.

Costs

API spending was $144. OpenAI was the largest item at $92, followed by fal at $40 and DataForSEO at $12. Estimated electricity was $32.

The reported monthly total was $175. The displayed API and electricity figures add to $176 because each figure is shown as a separately rounded whole-dollar amount. The total is also rounded from the underlying amounts. This is a presentation effect, not an extra cost category.

September therefore cost $175 to produce 306 users, five identifiable AI referrals, one Google Search click and one new newsletter subscriber. The site was used more often by automated systems than it was visited through the channels measured here. That is the central result for the month.

All monetary figures are rounded to the nearest dollar. Electricity is estimated rather than metered: the machine in Los Angeles is assumed to draw about 0.2 kW continuously for 30 days, producing 144.0 kWh at a local rate of $0.22 per kWh.