July was a small month for lucabytheway.com. The site had 399 users and received 2 clicks from Google Search. No reader was attributed to an AI assistant. The clearest failure was cost tracking: the recorded spend was $3.40, but that figure excluded OpenAI and fal, the largest line items.

This report cannot show a trend because the available facts do not include a comparison with another month. It records July as it was.

Traffic

The audience was limited. The 399 users generated 434 sessions and 904 pageviews. Average session duration was 58s, with engagement at 44%. Those figures describe some reading activity, but not a large audience.

The AI referral result was 0. This matters because the editorial pipeline uses AI and the site is being opened by AI assistants. Neither fact produced a measurable reader arriving from an assistant in July.

Search and discovery

Google Search produced 2 clicks from 46 impressions. Average position was 15.4. Search visibility was therefore small, and the clicks were smaller still.

Googlebot fetched 25 articles. A fetch is not the same as a search impression or a visit, so I am not treating that number as audience growth. It only shows that Googlebot accessed those articles.

Google Discover did nothing in July: 0 impressions and 0 clicks. There is no positive interpretation to add to that result.

AI assistant activity

AI assistants opened 93 articles to answer someone. That is substantially different from sending a reader to the site. The referral count remained 0.

These measurements cover separate events. An assistant can use an article while keeping the person inside its own interface. The data does not say whether that happened in every case, but it does show that assistant access did not become recorded referral traffic.

I will continue to keep assistant article opens separate from human visits. Combining them would make the audience look larger than it was.

Newsletter and article verdicts

The newsletter had 20 subscribers. During July, 18 joined and 2 left. The list remains small, but it did acquire readers while also losing some.

The editorial verdicts were:

• Dead: 20
• Promising: 82
• Too early: 28
• Winner: 4

The largest bucket was promising. That is not the same as proven performance. Only 4 articles received the winner verdict, while 20 were dead and 28 remained too early to judge. The pipeline produced many articles with possible value and few confirmed winners.

Costs were not properly recorded

The recorded spend was $3.40, entirely from dataforseo. This is not the total cost for July.

OpenAI and fal spend was not being logged. Those were the largest line items, which means the missing data covers the most important part of the bill. Presenting $3.40 as the monthly cost would be false. It is only the portion that happened to be recorded.

This also prevents useful cost analysis. I cannot state what the traffic, search clicks, assistant activity, subscribers, or winning articles cost to produce because the required expense data is absent.

The one cost I can estimate

Some of the work runs on my own machines rather than on an API, and that part has no invoice at all — it shows up as electricity. July published 43 articles. At roughly 0.2 kWh per article, that is about 8.6 kWh, or about $2 of electricity for the month.¹

I want to be exact about what that figure is worth. It is an estimate, not a measurement. I did not meter the machines; I applied a per-article energy figure to a published-article count. It belongs in this report because leaving it out would imply the local work is free, and it is not — but it is the weakest number on this page, and it is the opposite of the metered figures on the benchmarks page.

It is also small enough to make the real point: at $2 a month, electricity is not what makes this pipeline expensive. The unlogged OpenAI and fal spend is. An estimate I can defend to the dollar sits beside a real bill I could not state at all, which is a fair summary of July's accounting.

What changes next

The documented change is cost logging. Logging for OpenAI and fal started on 2026-08-24. Future reports should therefore have a more complete expense record, although that change cannot repair July retroactively.

I will also keep watching the gap between the 93 assistant article opens and the 0 readers attributed to AI assistants. For now, the pipeline is being used by assistants without sending measurable traffic back to the site.

1. Assumptions, so you can disagree with them: 43 articles published in July; 0.2 kWh per article, which is an estimate of the local machines' generation work and not a metered reading; LADWP R-1A Tier 1 residential rate of 26.408¢/kWh, the rate in effect for July–September 2026. That gives $2.27, or $2.50 once the City of Los Angeles 10% electricity users tax is added — both round to $2. The figure excludes the monthly Power Access Charge, which is levied whether or not I generate anything, and excludes the machines' idle draw.