TL;DR: This site went public this month with essentially zero search presence. The first working week was about plumbing, not vanity: submit the sitemaps once, seed IndexNow with all 370 URLs, turn on CDN logging to see AI crawlers, run a full seven-layer audit, and ship the first search-driven enrichment wave. Logging began Sep 11, and the first two-day read already showed 322 AI-engine requests from every major answer engine, with machine traffic nearly matching human traffic. The first weekly Search Console export read 6 clicks, 389 impressions, 127 pages. Tiny, but the machinery to grow it is now in place.
This site is two surfaces. One is a marketing site for humans. The other is a machine-first registry of open-source GTM tools, hundreds of entries, built with structured data, an llms.txt, and markdown twins so it can be read by AI engines as much as by people. It is agent-operated, human-gated, and updated daily. It launched publicly in September with essentially zero search presence. The practice behind it has been delivering growth systems under contract since August 2024; the web property is the new thing. So this is a from-zero visibility program with the work and the numbers in the open.
The first signal
On Sep 8 the site produced its first Search Console signal ever: 18 impressions and 1 click. The same day brought the first generative-AI-feature impression on Google's AI surfaces: exactly 1. One impression is not a trend. It is proof the pipes are connected, which is the only thing you can ask for on day one.
Connecting the consoles, once
On Sep 9 both sitemaps went to Bing Webmaster Tools and the Google domain property was verified via DNS. On Sep 10 the one-time indexing setup finished and IndexNow was seeded with all 370 URLs.
That produced a standing decision worth stating plainly: sitemaps are set-and-forget. The consoles re-fetch them on their own schedule, so re-submitting by hand is motion without movement. The only recurring action is an IndexNow ping per publish. Automate the recurring thing, do the one-time thing once, and stop touching it.
Measuring the machine plane
On Sep 11 CDN access logging went on. This is the only way to see AI crawlers and agents at all. They run no JavaScript, so they are invisible to normal analytics. If you only watch a standard analytics dashboard, an entire class of visitor, the one this registry is built for, simply does not appear.
The decision that came out of it: measure the machine plane separately from the human plane, because the human-plane tools are structurally blind to agents.
The first partial baseline: who the machines are
Here is the honest constraint on this section. Logging only started on Sep 11, so the first snapshot I could pull covered just the opening two days, and it arrived as a single combined count rather than a clean day-by-day split. I cannot report what I did not measure, so treat this as a first partial baseline, not a trend.
Even so, the read was striking: across those first two days, non-human requests (around 1,740) roughly matched human ones (around 1,700). Nearly half of everything hitting the site was machine traffic that a JavaScript analytics dashboard would never have shown me.
Inside that machine traffic, the AI answer engines were already present. The named crawlers in the log, by request count over the two-day window, were Anthropic's ClaudeBot (64) and Claude-User (18), OpenAI's SearchBot (57), GPTBot (20) and ChatGPT-User (19), Applebot (34), Cohere (20), Perplexity's User (19) and Bot (17), CommonCrawl's CCBot (15), ByteDance's Bytespider (14), Google-Extended (13), and Meta (12). That is 322 AI-engine requests in total, from every major answer engine, within days of logging turning on.
The point of the section is not the magnitudes, which are two days old and directional at best. It is that the machine plane is real, it is large, and now it is measured. This is the floor I will grow from, reported openly.
The heavy day: a seven-layer audit
Sep 12 was the audit. A full seven-layer pass across the site found zero P0 or P1 issues, three P2s, and all three were fixed the same day. Three catches are worth telling because each carries a lesson.
First, a stale deploy exclude had once wiped a live verification file that indexing depends on. The audit caught the pattern and the rule was hardened. The lesson is not about one file; it is that you audit your deploy pipeline, not just your pages, because the pipeline can quietly delete the thing your indexing rests on.
Second, the llms.txt link formatting was tripping real AI crawlers. Amazonbot and ClaudeBot were visible in the CDN logs, catching on line-end punctuation and 404ing. Machine readers are more literal than human ones; a human forgives a trailing character, a crawler does not.
Third, product analytics turned out to be dark in production from missing environment keys. It was found by actually checking rather than assuming, and then lit up. The recorded decision: standard analytics for the human plane plus CDN log analysis for the machine plane, because JavaScript analytics cannot see agents.
The first enrichment wave
On Sep 13 the first search-data-driven enrichment wave went out. Eleven registry pages that Google already ranks in the top 10 but that were getting no clickthrough got hand-written summaries and re-verified descriptions. One tool's license was corrected after re-verification, because accuracy is the product and a wrong license is a defect.
Then the first weekly Search Console export landed, covering Sep 4 to 10: 6 clicks, 389 impressions, 127 pages with impressions.
What the numbers say, and what is next
Six clicks is a rounding error to anyone measuring at scale. It is also real, and more to the point the machinery to grow it now exists: the consoles are wired, IndexNow fires on publish, both planes are measured, the audit surface is clean, and the enrichment loop has run once end to end.
The open question for next week is Bing. It has processed the sitemaps, but the pages are not being crawled. That is the diagnosis waiting to be done, and it is the more interesting problem, because ChatGPT reads Bing's index, which means Bing is not optional.