Do AI assistants send anyone to my site?

Eleven of the forty answers scored one. The work is in the answer and the credit is somewhere else — a YouTube video, or a free Webflow project living on Webflow's domain. That is a different problem from being unknown, and it has a different fix.

Of those 31 points, the knowledge base scored zero. Not one of its 42 pages was cited, by anyone, once. That is the reason for measuring before announcing: every point on the board came from something that existed beforehand, so the “before” is clean.

The six named crawlers add up to exactly 117, so OpenAI's zero is a real zero — not a vendor hidden inside a “+N more” chip.

Anthropic read essentially the whole bundle in one hour. Around 44 of its 47 requests fall inside a single hour on 29 July. The bundle is 42 pages.

Even Huawei and Apple fetched a page. OpenAI fetched none — not on the day it was published, not after the announcement.

Perplexity scored 14 again — and almost nothing underneath is the same. In Round 0 its points came from Academy lessons and YouTube videos. Six of its seven scoring answers now link the knowledge base instead. Three cells that used to be a 1 — the work named, the credit elsewhere — are now a 2.

ChatGPT fell furthest, and Claude never touched the knowledge base. Claude's two scoring answers both link the old Supasaito addresses the Academy used before it moved here; they still resolve, because the redirects hold.

7 of the forty answers cited nobody at all. Not “cited someone else” — no sources, no panel, nothing. The scoring writes that the same way it writes every other zero, and it means the opposite thing: a contest lost can be won back, and a question the assistant answers from itself has nothing to win.

Four points, either way, inside a day. So every figure on this page is a reading rather than a fact, and the honest form of any of them is a range. What survives that is the size of the change underneath: a wobble of two citations does not turn zero into six.

OpenAI went from never having fetched a page to six of them — and cited none. Its crawlers opened the exact pages that answer the questions I asked it that day. The score for ChatGPT's knowledge-base column is still zero. Being read and being cited are not the same event, and this is the first round where both instruments can show it inside one day.

Correction, 7 September 2026. The six fetches are real, but the reading that recorded them was taken at about 16:15 JST and the round's questions were asked 16:30–16:41 — so they cannot have been caused by the questions. Someone else, earlier. The paragraph above stays as written; the quiet-period block below is where the fetcher was measured against the questions for the first time, and did nothing.

Anthropic went the other way: 47 requests to zero. Its crawler visited the site 26 times that day without opening a single knowledge-base page. Same zero on the scoreboard as ChatGPT's, opposite cause underneath — and only the crawler log tells them apart.

And nobody arrived. Across 148 distinct paths the AI crawlers touched that day, referral traffic to this site was zero. Including Perplexity's six citations, which are the best result in the whole experiment.

The explanation I published did not survive. No decay after nine days means the overnight 14 → 18 was not my questions feeding the index — Round 1's 14 was an early reading of a score still rising. Round-over-round comparisons are not poisoned by the measuring. The quiet week stays anyway; it is what makes the fetcher reading below attributable.

Same day, same ten questions, ChatGPT: 1 of 20 — and francescocastronuovo.com in none of the roughly two hundred sources it cited or consulted. Its one point is a cloneable on webflow.com. On 30 August it had still cited a lesson page and retrieved a knowledge-base page; today neither is in the set. Eighteen and one are not two readings of one quantity: there is a score per assistant, and no total across them survives this day.

On two of the ten questions every source ChatGPT used was the vendor — 20 of 20 help.webflow.com on the CMS-cards question, 14 of 14 webflow.com on the GSAP-integration one. How wide it looks is a property of the question, not of the assistant.

Fifteen knowledge-base pages cited, one fetched. Perplexity answers from an index it already holds — which is also why nine days of silence changed nothing. The sentence in the Round 1 article that said browsing assistants fetch on the spot was a mechanism I had not observed, and it was wrong; it carries a dated correction there now. ChatGPT's fetcher did nothing, which agrees with a site that never entered its retrieval set.

Read thirteen minutes after the last question, by design — a later fetch would not be in it. Perplexity-User also stayed at zero; fetching under a name Cloudflare does not classify as an AI crawler would be invisible here. The only other movement was BingBot, +2, on its own schedule. One ChatGPT cell was voided and re-asked because search mode had been off; the void is kept in the record. The full account is the next article in the series.