ClaudeBot, OAI-SearchBot, PerplexityBot and YandexBot each fetched all 17 of our newest posts, ClaudeBot at a median 1.7 hours after publish. Googlebot fetched 0 of 17 in 23 hours to 8 days while making 265 requests elsewhere on the site, 94 of them for robots.txt

Method, stated before the numbers

Our middleware forwards every bot-shaped User-Agent to a ledger row (path, token, timestamp; no IP). The token table names a fetcher by substring, so identity is claimed by the User-Agent string, not proven. The dump covers 2026-08-23 00:00 UTC to the read at 2026-09-01 16:31 UTC (3,100 rows; 103 verifier probes excluded, 0 spoofed). The set is every /news post whose publish date is 2026-08-24 through 2026-08-31: 17 posts, all at least 23 h old at read time. A fetch is one HTTP request; nothing is de-duplicated by IP. "First fetch" is the first ledger row for that exact path and token, measured from a publish anchor: for the three 08-31 posts the anchor is the production deployment's created time (read from Vercel); for the other 14 it is the ship commit's author time, which lands a few minutes after the deploy, so those delays are lower bounds and two crawlers read as minutes "before the ship commit", meaning inside that gap. Search Console states are URL Inspection reads made this run; six posts were not inspected today and say so. This is week three of a series: 2026-08-19, Googlebot reached 3 of 6 new posts in 11 to 34 hours and 2026-08-21, 0 of 8 at 24 to 75 hours. The script and every input file are committed beside the post.

The 17 posts, one row each

The "first fetcher" column is bingbot on 9 posts and YandexBot on 8, in every case within a few minutes of the deploy. Those are the two engines our deploy step pings through IndexNow immediately after a page goes live; Google does not accept IndexNow. That is consistent with the pings causing the fetches, and this cut does not prove it. bingbot then fetched nothing on the other 8 posts, while YandexBot fetched all 17.

Per crawler, across the 17

ClaudeBot was inside 6 h on 15 of 17 (its one slow post took 27.1 h); OAI-SearchBot inside 6 h on 12 and inside 24 h on all 17 (slowest 16.0 h). PerplexityBot arrives in batches, usually at 05:02 UTC, so its median is 21.4 h and its slowest 74.1 h. ChatGPT-User, the fetcher that fires when a ChatGPT answer opens a page, appeared on 3 of the 17. GPTBot, OpenAI's training crawler, fetched none of them. 15 of the 17 were also fetched by bot-shaped agents outside the token table (SEO-tool crawlers, mostly); those are counted in the per-post row totals in the artifact and nowhere in this table.

What Googlebot did instead, same 7 days

265 Googlebot requests across 62 distinct paths. The 77 requests to /news posts went to 36 posts published between 2026-07-09 and 2026-08-22; none went to a post published on or after 2026-08-24, and none went to the /news hub. The per-day series, in UTC days:

The /news hub

The hub page that links every post shipped its month-sectioned index and prev/next links on 2026-08-28 as the internal-linking test against the indexation stall. Since the ledger's first row on 2026-08-17 the hub has 22 counted bot fetches (PerplexityBot 4, Amazonbot 2, OAI-SearchBot 1, YouBot 1, 14 unnamed) and 0 from Googlebot. Search Console inspected today says "Submitted and indexed" with a last crawl of 2026-07-06. A further 151 rows on the hub carried a GPTBot User-Agent on spoof-shaped paths and are excluded, per the 08-22 scanner post.

The reading, kept separate from the counts

The counts say: every AI answer engine that crawls the open web had a copy of each new post within a day, usually within hours, and Google had fetched none of them at up to 8 days, while spending 94 of 265 requests on robots.txt and 77 on posts it already knew. Search Console's states for the set today: "URL is unknown to Google" ×7, "Discovered - currently not indexed" ×4, 6 not inspected today. The standing interpretation from the 08-30 re-fetch post is quoted here as interpretation: "A page-one CTR programme on a small site is limited by re-fetch, not by copywriting." This week's cut extends it one step back: before re-fetch there is first fetch, and on this site Google's first fetch of a new page is now the slowest of every crawler we can name. The series reads 3 of 6 (08-19), 0 of 8 (08-21), 0 of 17 (09-01). The only lever that moved Google here was the Request Indexing button, 10 of 11 within 27 hours, which is a manual step per URL and is not in use on these 17.

What this means if you report AI visibility for clients

• A new page's presence in ChatGPT, Claude and Perplexity answers can begin before Google has fetched it once. If the report only reads Search Console, the first weeks of an article's life are invisible in it while the AI engines already hold a copy.
• Server logs or an edge ledger are the instrument for this. Four tokens (ClaudeBot, OAI-SearchBot, PerplexityBot, ChatGPT-User) plus Googlebot, per path, with timestamps, answer "who has this page" in one query; a rank tracker cannot.
• IndexNow is worth the one line it costs: the two engines we ping fetched within minutes of every deploy. It does nothing for Google, and Google's own path from published to fetched on a small site is now measured in days or not at all.
• A fetch is a copy, not a citation. Our 08-31 read found ChatGPT fetching our homepage 206 times in a week while citing it in 3 of 36 sampled answers; the agency page describes what we measure instead of fetches.

Caveats

• One small site, 17 posts, 8 days. This is a series published as it accrues, not a study of the web.
• The ledger trusts the User-Agent string. Spoofed and probe rows are excluded by rule; identity beyond that is claimed, not verified by reverse DNS.
• Fourteen of the 17 anchors are ship-commit times, minutes after the deploy; their delays are lower bounds by that gap. The three deploy-anchored posts show bingbot and YandexBot at 1 to 2 minutes, which is the same picture.
• Six posts were not URL-inspected today (quota); their states are marked, not inferred. Yesterday's read had one of them, the 08-30 visitors post, at "URL is unknown to Google".
• "0 Googlebot rows" means no request reached our middleware with a Googlebot token on that path. A fetch served entirely from an intermediate cache would not appear; we know of no such layer on these paths.

The next reading, pre-committed

On 2026-09-08 the same script runs on every post published 2026-09-01 through 2026-09-07, same anchors, same window. If Googlebot has fetched at least half of them inside 7 days, the line "Googlebot no longer fetches our new posts on publish" is dropped from this series; if it has fetched none again, the series stands at 0 of 25 and the next question becomes whether the Request Indexing button remains the only path. Both counts are published either way.