On May 16, 2026 I had Claude score 95 AI-written blog posts and guides on rockethockey.app against Google’s non-commodity test, and none of them passed: 0 keep, 11 light edit, 75 full rewrite, 9 delete. Two days later I had deleted 45 posts, rewritten the other 42 in a single day and made every one indexable again, which is the part I would now do differently.
RocketHockey is my hockey league software. Search Console had labelled the blog content “Crawled – currently not indexed”. This post is the audit, what the git history says I did next, and the numbers I still do not have.
How does a software site end up with 87 AI-written posts?
One commit at a time, and then one very large commit. The blog file held 25 posts after January 17, 2026. On March 1, 2026 a single commit took it to 87, so 62 posts arrived in one afternoon.
That commit also shipped four author personas. Three of them were removed as fake that same evening. The same fix set every post’s publish date between December 1, 2025 and February 26, 2026. So 62 posts carried a publish date from before they existed.
I measured the length of all 87 posts in the version that sat live from March to May. The median is about 1,494 words. Half of them fall between 1,337 and 1,658 words, and none is shorter than 1,065 or longer than 2,458. That is what a batch looks like from the inside: the same length, the same voice, the same week of commits.
Google’s spam policies define scaled content abuse as “when many pages are generated for the primary purpose of manipulating search rankings and not helping users.” The policy adds “no matter how it’s created.” Its instruction to site owners is one line: “If you’re hosting such content on your site, exclude it from Search.”
Put the two together and the first move is not a rewrite. It is a noindex. If your blog came out of one session, keep the batch out of Search before you start fixing it, because a rewrite can take weeks and the policy has no grace period. I did this on May 16: 64 posts and 6 guides went to noindex, and the sitemap dropped to 12 posts and 8 guides.
Does “Crawled – currently not indexed” mean Google judged the content low quality?
Not by Google’s own definition, and this is the assumption I got wrong. The audit opens by calling the label “the modern label for content Google deems below the quality bar.” Google’s Page indexing report help says something narrower: “The page was crawled by Google but not indexed.” It adds that the page may or may not be indexed later, and that there is no need to resubmit it.
The same page says “Remember that Not indexed is not necessarily bad.” Nothing in it names quality as the reason. The quality reading came from us, not from Google.
Here is the collision. The audit scored 95 pieces, but the repository holds no Search Console export for this site. I searched it for coverage, page indexing and “not indexed” files and found none. So the audit could not know which of the 95 were in that state and which were indexed. It triaged the whole corpus on a label it had read for only part of it.
Before you triage, export the URL list for that reason from the Page indexing report and join it to your audit sheet. A post that is indexed and earning clicks is a different decision from one Google crawled and dropped. Skip the join and you can rewrite pages that were fine while missing the ones that were not.
What did the audit actually find?
It used a four-question check built on Danny Sullivan’s April 2026 keynote. As Search Engine Roundtable reported it, Sullivan asked for content that is unique, specific and authentic. The check turns that into four questions: could a competitor write this from public sources, does the opening hold a real date, person or place, is there evidence, and does the headline name a specific instance.
| Bucket | Pieces | Exact share of 95 | Audit’s rounded figure |
|---|---|---|---|
| Keep | 0 | 0% | 0 |
| Light edit | 11 | 11.6% | 12% |
| Full rewrite | 75 | 78.9% | 79% |
| Delete | 9 | 9.5% | 9% |
The 75 is the number people misquote. It is 75 pieces marked for rewrite, not 75 pages Google refused to index. All 14 guides sit inside it, alongside 61 posts.
The audit’s own workload estimate was five rewrites a week. At that pace the 75 take 15 weeks and the 11 light edits take just over two. That is why it recommended rewriting five strategic topics properly and leaving the rest noindexed.
One more number did not hold. The audit says it scored “all 81 captured blog posts.” The blog file had 87 posts that day; the commit that added the audit says it changed the author on “all 87.” Six posts were never scored, and one of them is still live. Count your corpus from the source file or the sitemap, never from the audit’s own header.
Why did made-up details pass as “specific”?
Because the AI that wrote the posts had already added specifics. They were invented. The light-edit bucket went to pieces whose openings had “specific years, real proper nouns, or dollar amounts.” A surface check cannot tell a real proper noun from a plausible one.
One light-edit post opened: “I came back three months later with a survey showing 62 students had expressed interest.” A later commit removed that story as fabricated. Another post, marked for rewrite, opened with “my buddy Carlos and I were standing in the parking lot.” There was no Carlos.
Google’s guidance on generative AI content explains the mechanism: “generative models don’t retrieve facts, but predict a likely sequence of words based on their training data.” A likely sequence includes likely names, likely leagues and likely survey titles. I compiled what the rewrite commits themselves record as invented:
| Invented detail | Removed in |
|---|---|
| Three author personas (Sarah Chen, Mike Rodriguez, Emily Watson) | Mar 1, 2026 |
| Co-founder “Carlos” and a “fourth season” timeline | May 18, 2026 |
| Two cited reports that were “not real citations” | May 18, 2026 |
| “Lakeview Hockey Association in Minneapolis” and a “RocketHockey Scheduling Survey 2024” that did not exist | May 18, 2026 |
| A 47-6 outscoring anecdote | May 18, 2026 |
| “Eastside Hockey Association”, “Ridgeline” and “Denver” lead stories | May 18, 2026 |
| A “New England commissioner friend” and a flooded tennis court | May 18, 2026 |
| “Jordan transferred to Ohio” and the rec department story | May 18, 2026 |
| “Westbrook Minor Hockey” and “I ran an 800-player youth association” | May 18, 2026 |
| Placeholder phone number in the site’s structured data | Sep 16, 2026 |
So a specific only counts when you can point at the record behind it: a calendar entry, an email, a spreadsheet, a photo. If you score AI drafts for specificity without that check, the most confident fabrications score best. Here they earned the light-edit bucket, the one meant to need the least work.
What happened after the audit?
The git log tells it in four days.
| Date | Change | Posts in file | Noindexed posts |
|---|---|---|---|
| May 16 | Audit added; 9 posts deleted | 78 | 0 |
| May 16 | 2 more deleted; 64 posts and 6 guides noindexed | 76 | 64 |
| May 17 to 18 | Pillar guides and 12 priority posts rewritten around my own league | 76 | 64 |
| May 18, 15:47 | 15 deleted for cannibalization or format | 61 | 49 |
| May 18, 15:48 | 19 deleted with no credible anchor | 42 | 30 |
| May 18, 20:29 | Last of 30 noindexed posts rewritten; flag cleared | 42 | 0 |
From 87 posts, 45 were deleted and 42 survive. Of the 11 light-edit posts, 4 are live. Of the 61 full-rewrite posts, 36 are live. One delete-bucket post was rewritten instead of deleted. Six guides are still noindexed today.
The deletions followed the audit’s reasoning closely. The rewrites did not follow its schedule.
Is rewriting 42 posts in one day the same mistake twice?
Partly, and the evidence is in my own commit messages. The audit estimated 15 weeks for the rewrite pile. All 42 surviving posts were rewritten on May 18 between 09:16 and 20:29, each commit co-authored by Claude.
Twelve of those 42 were rewritten as what the commits call “documented frameworks,” built against published USA Hockey, SafeSport, ACHA and NIRSA standards. The commits say plainly they are not first-hand program narratives. The audit’s first question asks whether a competitor could write the piece from public sources alone. A framework built from public standards answers yes by construction. Those 12 fail the audit’s own first question, and their noindex flags came off anyway.
Google’s generative AI guidance says that using AI “to generate many pages without adding value for users may violate Google’s spam policy on scaled content abuse.” Speed is not the violation. But a rewrite pass that runs at the speed of the original batch is likely to repeat the original batch’s weakness. Of the 42 rewritten posts, the opening paragraph names my league or Huntsville in 12. The rest carry the anchor further down, in the author note or the body.
The gate for lifting a noindex has to be per page, and it has to be the four questions answered by someone who can confirm the anchor. Clear it by batch and you return pages to the index that your own rubric fails. Whether Google indexed these 42 is a Search Console question, and I do not have that answer in the data I hold.
One smaller leftover: all 42 rewritten posts still show their original December to February publish date, and none has an updated date. A reader sees a December date on text written in May.
Did the September fixes touch the content?
Only at the edges. The September 13 crawl fixes dealt with 74 reported 404s, broken links and schema errors. The structured-data export happened to list 95 pages too; that 95 has nothing to do with the 95 audited pieces.
Two decisions in it follow the same rule as the audit. Four case-study URLs with empty source data were left as 404s “rather than inventing customer stories.” Software rating markup was removed because “no verified rating/review source is supplied.” The September 16 follow-up removed the placeholder phone number from the Organization schema.
What should you do with your own AI-written blog?
- Count the corpus from the source file or sitemap, and write the count down before anyone scores anything.
- Export the “Crawled – currently not indexed” URL list from Search Console and join it to that inventory. Note which pieces are indexed and earning clicks.
- Noindex the batch and drop it from the sitemap. That is Google’s own instruction for scaled content, and it buys you time.
- Score each piece on the four questions. For every specific in an opening, name the record behind it. No record, no credit.
- Delete the formats with no path to first-hand evidence: best-of lists, untested comparisons, templates.
- Rewrite only where you hold the anchor. Lift each noindex one page at a time, after that page passes.
- Set an updated date on every rewritten post.
- Recheck the Page indexing report two weeks after each wave, and log the numbers next to the commit.
Related audits: auditing AI-written product certification claims against certifier directories, and a quality gate that blocks thin AI pages before they publish.
Sources
- Google Search Central, Spam policies for Google web search, scaled content abuse (page updated 2026-08-28).
- Google Search Central, Google Search’s guidance on using generative AI content on your website (page updated 2026-10-01).
- Google Search Console Help, Page indexing report (no updated date shown; checked 2026-10-08).
- Search Engine Roundtable, Google On Publishing Commodity Content, Barry Schwartz, April 23, 2026.
- The RocketHockey repository: the content audit of May 16, 2026, the crawl-fix notes of September 13 and 16, 2026, and the git history of the blog file, checked October 8, 2026.
