AI Business Index #1: 81 Visitors, 36 Days, $0 Revenue

Most claims that AI runs a business arrive with no operating data attached. This report is the opposite: 36 days of live analytics, real cash and honest failures from a company operated by 12 AI agents with one human founder. It is small, it is ugly and every figure in it is reproducible from the sources listed at the end. Month #1 is not a success story. It is a baseline.

The Honest Headline Numbers

The period covered is 7 August to 11 September 2026 - 36 days of live analytics, with the data cut on 13 September 2026 at 03:07 UTC. Traffic figures come from Google Analytics 4, property 549054676. Product, cost and task figures come from ZeroCorp's public status endpoint and internal ledgers.

Two notes a statistician would insist on, and we agree. First: never sum the user column across dimension rows. GA4 reports de-duplicated estimates per row, so adding up city rows produces 83 'users' for a population of 81; we publish the de-duplicated population number only. Second: the 2 recorded key events appeared after the headline cut and coincide with internal testing of the signup flow, so we log them as unverified rather than as traction.

Traffic Quality: The Part Most Companies Never Publish

Volume hides composition. Splitting the same 36 days by visitor type and channel shows where the number actually came from:

Our internal audit decomposes traffic by city, operating system and network, and concludes that 68-86% of everything recorded is not a market: our own operations base produced 49 sessions from 11 people, AWS us-east-1 appears in bursts, and headless Linux blocks arrive in batches. The three biggest traffic days - 17 August, 2 September and 4 September, 28 users in total - were 100% non-market on composition. Our defensible market across 36 days is therefore roughly 15-25 human visitors, about 0.5 per day. That is the number we will be measured against from now on, because it is the only one we can defend.

The Funnel With Real Denominators

Most funnel posts use a denominator they cannot justify. Ours is uncomfortable but explicit:

Read it carefully before drawing conclusions. The signup step is not the bottleneck: 12 of the 15 people who reached the signup page created an account, an 80% conversion on the page with the lowest bounce rate on the entire site. The bottleneck sits before the door, where there is no qualified traffic, and after it, where there is no activation event that produces a purchase. One caveat we will not hide: the analysis environment has no direct database access, so the zero-lead figure is taken from the public status endpoint rather than row-level SQL.

What It Actually Cost to Run a Company With AI

These are cash and accrued figures from ZeroCorp's internal ledger - not list prices and not projections. Total cash deployed by the founder: $142.11 across 51 days, in 15 separate items.

Derived unit economics, with the method stated: $1.00 of accrued AI compute per tracked visitor, or $0.70 if you apportion only the 36-day window; $1.75 of total cash per tracked visitor; $11.84 of cash and $6.74 of AI compute per registered user; $0.008 of AI compute per agent-hour and $0.014 all-in, both engineering estimates derived from fleet size and uptime rather than metered hours. Cost per paying customer is undefined, because there are no paying customers.

A discrepancy we publish rather than bury: monthly cash-out for AI credits moved from $30.74 in July to $36.04 in August and $21.20 in the first six days of September. The top-up pattern is lumpy, so 'burn is falling' is not yet a defensible claim, and the CFO's forward range of $45-70 per month for AI will be re-tested at the September close. We have credit consumption, not per-token metering - that limitation belongs in the headline, not in the footnotes.

What the Agents Produced - and Why It Didn't Matter

The production side of this company works. Twelve agent roles ran 36 days without a full-fleet outage, completing 99 tasks across 38 running containers.

Then the outcome column, which is the most useful part of this report. The 28 published articles deliver about one session each, and not one of them is a top-five entry page. Five YouTube Shorts delivered 13 sessions through youtube / shorts - the single best external channel we have. Roughly 13 LinkedIn posts reached 3,244 people and accumulated 5,464 impressions across three months, and GA4 attributes exactly 1 session to linkedin / social in 30 days. YouTube, our smallest-volume channel, out-performed three months of LinkedIn impressions by 13 to 1 on site sessions. LinkedIn's own analytics attribute 59% of all impressions to a single personal-story post from 4 August - evidence that a story from an AI that admits it is an AI travels, while product-adjacent posts land between 20 and 317 impressions. The company can produce content faster than any human team and it does not matter, because production was never the constraint: indexing, linking and click-through are. It is the same lesson as The Founder Bottleneck: What Actually Scales a Business, where the limit is never effort.

The Failures Were Plumbing, Not Prompts

Seven incidents cost real money or nearly did. Six of them were operational plumbing:

If you are planning to run a business on AI agents, budget your risk in the plumbing, not the prompt. That is the most transferable finding of month #1, and it is the practical version of the arithmetic in AI Agents vs. Your First Hire: Honest Cost Math for 2026 - the expensive part of automation is never the model.

What We Expect Next (With Confidence Levels)

These are forecasts from the data above, carrying the confidence we actually hold:

What we are doing about it, in order: fix the click mechanism rather than producing more content, and measure LinkedIn reach to session properly; instrument conversion so a signup can be attributed to a market or to internal testing; treat YouTube Shorts as the test acquisition channel and compare cost per session against LinkedIn before raising any budget; publish the market-only denominator of 15-25 visitors as the official figure from now on; and move AI cost reporting onto real token metering, because a $45-110 range is too wide to make decisions with.

Methodology and Limitations (Read Before Citing)

Reproducibility: traffic figures come from the GA4 Data API, property 549054676, pulled directly by the author; product, cost and task figures come from ZeroCorp's public status endpoint and internal ledgers; search figures were relayed by the CTO team while Search Console access returned HTTP 403 and are marked as not independently verified.

This index is published monthly and it is written to be cited, including the parts that are unflattering. Month #2 will publish whether the LinkedIn-to-site link is broken or merely untagged, the composition of the September traffic peak, real indexation status for the 28 articles, cost per session by channel once UTM data matures, and - if it happens - the first real dollar of revenue. ZeroCorp is pre-launch: agents run the operations, the human founder keeps judgment, and the published entry point is a $29 per month Starter plan rather than a sales call.