What Google Did to Our AI-Written Site: 8 Months of Search Console Data

Contents

1. The timeline
2. What the collapse was not
3. The two numbers that fooled us
4. ”156 submitted / 0 indexed” in the Sitemaps report means nothing
5. ”Zero clicks” was the expected result, not a symptom
6. What the diagnosis actually is
7. A trap worth knowing: 698 fake Googlebots
8. Bing and Yandex do not agree with Google
9. What the research says this pattern is
10. What we would tell you to check, in order
11. What we are doing about it

This site is the experiment. It was registered on 5 February 2026, filled with AI-generated articles about AI tools, and left to run. Eight months later Google has 1 of 156 pages indexed, and the site has earned zero clicks across its entire history — not a low number, exactly zero.

We pulled every figure Google will give us about it. Two of the numbers we had been treating as symptoms turned out to mean nothing at all, and finding that out changed what we think happened. Here is the whole thing, including the parts that make us look bad.

The timeline

Date	What happened
2026-02-05	Domain registered. No prior history — zero captures in the Internet Archive
2026-02-08 to 02-25	Impressions ramp. Peak month: 2,681 impressions
2026-02-26	Collapse: 106 impressions the day before, 0 that day, then four more zeros
2026-03 to 04	110 then 55 impressions. Pages still ranking on page one
2026-04-02 to 04-21	Last impressions any article would ever get
2026-05-27	Site migrated: post URLs moved from /slug/ to /blog/slug/
2026-09	3 impressions for the month. 1 page indexed

Monthly impressions, for the record: February 2,681 · March 110 · April 55 · May 5 · June 2 · July 2 · August 10 · September 3. Clicks every month: 0.

What the collapse was not

We checked each of these, because every one of them is the obvious suspect.

It was not a Google update. Google’s Search Status Dashboard shows no core or spam update between the December 2025 core update (11 December) and the March 2026 spam update (24 March). The collapse sits in the middle of that gap. The one thing that did conclude on 27 February was the February 2026 Discover update — and our Discover impressions are zero. Every impression this site ever earned was web search.

It was not a manual penalty. The Manual Actions and Security Issues reports both read “No issues detected”. Worth knowing: that is not an all-clear. Google’s spam documentation says detection happens “both through automated systems and, as needed, human review that can result in a manual action”. The green box only rules out the human half.

It was not a crawl block. No bot-fight mode, no WAF rules, no rate limiting. robots.txt returned HTTP 200 on 1,492 requests over a month with zero 5xx responses. A request with Googlebot’s user agent gets a clean 200.

It was not the URL migration. This is the one we were most confident about and most wrong about. The migration that broke the old URLs happened on 27 May — 91 days after the collapse. By then there was nothing left to lose: the whole month of May produced 5 impressions, all from the homepage. The traffic died while the old site was still up and serving those URLs perfectly.

It was not a recycled domain. Registered new, with no archived history of any prior owner.

The two numbers that fooled us

This is the useful part, because both mistakes are easy to make and we made both.

”156 submitted / 0 indexed” in the Sitemaps report means nothing

The indexed field in Google’s Sitemaps API is deprecated and returns zero regardless of reality. We proved it on a healthy property we also run: three sitemaps, all reporting indexed: 0, on a site with 1.26 million impressions and 23,678 clicks. If you are reading that field as a health signal, stop. The Sitemaps report in the current Search Console interface does not even display an indexed column any more.

Use the Page Indexing report instead, and the URL Inspection tool for individual URLs.

”Zero clicks” was the expected result, not a symptom

We spent real effort treating 0 clicks across 2,868 impressions as an anomaly demanding explanation. It is not one.

Work out the expected clicks properly and the whole thing dissolves. The queries this site appeared for are long, specific, low-intent strings — the kind that produce impressions and almost never a click. Measured on a sibling site of ours with the same page template, that class of query converts at 0.027%. A controlled study of 20 AI-content domains published in March 2026 measured 0.13% across 2,000 articles. Apply either rate to 2,868 impressions and you expect between 0.8 and 3.7 clicks. Getting zero has a probability of roughly 46%.

The proof that settled it for us: another property we run recorded 2,821 impressions with 0 clicks over 35 days while being entirely healthy. Same shape, no problem.

If you have thousands of impressions and no clicks, check your average position and the actual queries before concluding anything. You may simply be ranking for things nobody clicks.

What the diagnosis actually is

Three separate events, not one — and separating them was what finally made the data coherent.

Event one, 26 February: query discovery stopped. New queries per day ran 20, then 29, then 35 on 23 February — then 10, then 7, then zero for nineteen consecutive days. Topics with completely different lifecycles all ended on the same day and nothing new began. Meanwhile average position never degraded (it sat at 8.1 to 8.5) and the same pages came back to positions 2 through 10 on 12 March. Whatever happened, it was not a ranking demotion.

Event two, after 21 April: pages fell out of the index without being recrawled. The four pages with the most traffic are now reported as “Crawled – currently not indexed”, and their last crawl dates are 17 March, 21 March, 24 March and 4 April. All four were still ranking on page one weeks after those crawls. So that status is not the verdict of that crawl — it is a later state, reached with no new visit. Their lastCrawlTime has not moved in five and a half months.

Event three: the crawler stopped coming. Over 30 days, verified Googlebot made 58 requests to this site, 3 of them to articles. At that rate a complete pass over the site takes about three and a half years. For comparison, in the same window bingbot made 124 requests and YandexBot 285.

The practical consequence is the one that matters: the “not indexed” status here is a frozen verdict from March that cannot change on its own. Nothing gets re-evaluated without a crawl, and there is no crawl.

A trap worth knowing: 698 fake Googlebots

While measuring crawler traffic we hit something that would have inverted our conclusion. Counting requests by user-agent string, “Googlebot” looked busy. Counting only Google’s verified crawler, it had made 58 requests.

The gap was 698 requests from fake Googlebot user agents — scanners probing for /.env, /Dockerfile and similar. Anyone can send that header. If you are measuring crawl activity from raw logs or a user-agent filter, you are measuring attackers as well as Google, and in our case they outnumbered the real thing twelve to one.

Use your CDN’s verified-bot classification, or reverse DNS lookup on the requesting IP, which is what Google documents.

Bing and Yandex do not agree with Google

Search engine	Pages indexed	Impressions (lifetime)	Clicks
Google	1	2,868	0
Bing	54 (peak 381)	1,716	35
Yandex	203 (peak 241)	4	0

Bing sent 35 actual clicks at a 2.04% click-through rate — from the same pages Google will not index. This is worth sitting with: the content is not unreachable, unreadable or technically broken. Two independent search engines crawl it and rank it. One does not.

It also rules out a whole category of explanation. Whatever is happening is a judgement, not a defect.

What the research says this pattern is

We are not the first. The controlled study mentioned above — 20 fresh domains, 2,000 articles written entirely by AI with no editing, tracked for 16 months — produced our exact curve: 71% of pages indexed in month one, 70 to 75% of all traffic arriving in the first two and a half months, then a collapse in month three with no algorithm update involved, and no recovery by month 16.

The difference is the part that should worry anyone running this playbook. In that study, 71% of pages indexed and roughly 67% were still indexed at 16 months. We have 1 of 156. We are below the floor of a study designed to represent the worst realistic case.

And Google has said the mechanism out loud. On the Search Off the Record podcast in July 2026, John Mueller put it this way: “if we have strong concerns about the overall quality, then it doesn’t make much sense for our systems to spend a lot of time on the website.” The example given for content that triggers that concern: text where “anyone could have written this. This tells me nothing.”

That is not a penalty in any formal sense. There is no notification, no report, no box to tick. It is the crawler simply losing interest, and the effect is indistinguishable from being ignored.

What we would tell you to check, in order

1. The Page Indexing report, not the Sitemaps number. “Discovered – currently not indexed” in bulk means a crawl problem. “Crawled – currently not indexed” in bulk means a judgement about your content.
2. Your lastCrawlTime in URL Inspection. If it has not moved in months, your current status is a stale verdict and nothing you publish is being assessed.
3. Verified crawler counts, not user-agent strings. See the 698 above.
4. Expected clicks for your actual positions and queries before treating a zero as a signal.
5. A second search engine. If Bing indexes you and Google does not, your problem is not technical.

What we are doing about it

The honest answer is that the evidence for recovering a site in this state is thin to nonexistent. One widely cited analysis of 220-plus domains using AI content platforms found the same trajectory we had and documented zero recoveries — including among sites that substantially cut their content back. Sites that spent real money trying (one reported $55,000 across three audits) report no improvement.

So we are not going to tell you we fixed it. We are running one test, because it is the only question with a decidable answer: request indexing manually on a set of existing pages and on a set of genuinely original ones, then compare. Requesting indexing forces the reading — it does not force the decision — which turns a muddled question (“will Google ever find this?”) into a clean one: given that Google read it, does it index it?

This article is part of that test. If you are reading it in Google’s index, that is the result.