AI Content Optimization Checklist: 17 Steps Before You Hit Publish

AI Content Optimization - 12

Last updated November 14, 2026. Every claim below links to the source it came from.

In January 2023, CNET’s editors learned that a draft can look finished and still be wrong. The site had quietly published 77 finance explainers written with an in-house AI tool. A reporter at Futurism pulled on one thread, and the whole batch came apart: 41 of the 77 needed corrections, some of them for phrasing that wasn’t original.

In 2025, Deloitte Australia had its own version of that day. A government welfare report, a contract worth about A$440,000, and fake references plus an invented quote from a court judgment. Deloitte repaid the final installment.

Then, on August 18, 2026, Google rolled out a spam update that hit sites hard: one programmatic, AI-heavy site lost rankings for over 200,000 queries.

Different cases, same root cause. Somebody skipped the last ten boring minutes before publishing. This checklist is those ten minutes, written down.

The short version

A fair warning on scope: the 17 steps are a synthesis. The cases below show why each gate exists. They don’t prove that a given step will lift your rankings by a given amount, and nobody can honestly promise that.

What the evidence says about AI and rankings

Ahrefs took the top 20 results for 100,000 random keywords, about 600,000 pages, and scored each with its own AI detector (Ahrefs study, July 2025).

Page type in the top 20ShareChart
Pure human13.5%███
Human and AI mix81.9%████████████████
Pure AI4.6%█

The correlation between AI share and ranking position came out at 0.011, effectively zero. Pages in first place leaned slightly toward less AI, though Ahrefs itself describes that as a faint pattern rather than a ranking factor (Search Engine Journal’s summary).

Three limits worth keeping in mind:

  1. The detector is Ahrefs’ own, so the labels are only as reliable as that tool.
  2. The data is from mid-2025, before several of the 2026 spam updates below.
  3. A flat average says nothing about whether your page is good enough.

So the reading is neither “AI is banned” nor “AI is free traffic.” It is: the production method isn’t the signal. Quality and effort are.

What Google is actually doing in 2026

The policy. Scaled content abuse arrived on March 5, 2024, alongside two other new spam policies, and targets mass-produced pages made mainly to rank. A PPC Land explainer notes the policy is method-neutral and that a spam update began on September 24, 2026 with it in scope. That explainer also notes Google’s claim of a 45% reduction in low-quality results in 2024 came from its own evaluations, with no published method or outside audit. Treat that figure as Google’s claim, not a measured fact.

The guidance. On October 1, 2026, Google added three sentences to the accuracy section of its generative AI guidance. They explain that models predict likely text rather than retrieve facts, so output can contain inaccuracies, and that all AI-generated content should be manually fact-checked before publishing. Coverage of the change notes it also reaches things teams often automate quietly, such as title tags, meta descriptions and alt text. The page still opens by saying generative AI is useful for research and for adding structure to original content.

The quality bar. Google’s May 15, 2026 guide to generative AI search pushes unique, non-commodity content. One commentator boils it down to a test: could an AI model produce an equally useful version of this page? (Link-Assistant’s analysis). Separately, John Mueller warned in August 2025 that using LLMs to build topic clusters creates site “liability,” as PPC Land reported.

Case file: four sites hit by the August 2026 spam update

Glenn Gabe of GSQi documented four anonymized cases after the update, which began August 18 and finished rolling out on August 21, 2026 (full write-up).

CaseWhat he foundScale of the hit
1. Ultra-YMYL siteProgrammatic pages across countries plus chunks of AI-generated textOver 200K queries lost entirely
2. Thin affiliateTurn-key Amazon affiliate site, with bottom-of-page text detected as 97%+ likely AIOver 14K queries lost
3. Programmatic product pages1.5M URLs indexed, about 85% in the programmatic section, little added valueSite-level drop across sections
4. Aggressive scaled site250K+ URLs and redirects to riskier sitesClose to 25K queries affected

Three lessons that matter for a checklist:

  • The damage can be domain-wide. In case 3, the drop hit sections of the site that weren’t the problem.
  • It spreads past blue links. Gabe notes that impacted sites can fall in AI Overviews and AI Mode too.
  • Recovery is slow. He says it can take months of visible cleanup, and cites a five-month recovery example. Case 2, he says, may never recover without a change in how the site works.

To be fair about it: these are one consultant’s observations of four sites, not controlled experiments, and he is selling recovery services. They are still the best documented look at what 2026 enforcement hits, and they match Google’s published policies.

Case file: two public failures from the other direction

CNET (November 2022 to January 2023). The AI-assisted stories ran under a vague “CNET Money Staff” byline. After the audit, CNET said 41 of 77 had errors, with a small number needing major fixes and the rest minor ones such as transposed numbers and incomplete names. Some correction notes said phrases had been replaced because they weren’t original. The company later brought in stricter AI rules.

Deloitte Australia (2025). Reporters and a University of Sydney researcher found fabricated references and a made-up court quote in a report for the Department of Employment and Workplace Relations. A corrected version disclosed the use of generative AI. The department said the recommendations were unchanged, and Deloitte agreed to repay the final installment of the A$440,000 contract.

Neither is an SEO case. They’re included because they show the same failure Google’s October update now addresses: output published without a human verifying it.

The 17 steps, in four gates

If a page fails Gate 1, don’t waste time polishing Gate 3.

Gate 1: Truth (steps 1 to 5)

1. Verify every number, name, date and quote against a primary source. The study, the filing, the documentation itself, not a blog repeating it. If you can’t find it, cut it.

2. Open every link and citation. Confirm it exists and says what you claim. Deloitte’s problem wasn’t typos; it was references to things that didn’t exist.

3. Redo every calculation by hand. Futurism’s CNET critique included a compound interest explainer that blurred interest earned ($300 on $10,000 at 3%) with the total balance ($10,300). A calculator catches that in five seconds.

4. Check every “latest”, “new” or “currently.” Models describe a world that has already moved. Look the claim up with today’s date in mind.

5. Run a similarity check, then read the result. CNET’s corrections show what happens when a tool is run but not read.

Gate 2: Value (steps 6 to 10)

6. Write one sentence on what a reader can’t get from the current top 10. If you can’t, you have a summary, not an article.

7. Add first-hand material. Your own test, a screenshot, a number from your analytics, a quote from a named person who agreed to it. It’s the cheapest answer to the “could a model write this?” test.

8. Answer the search intent on the first screen. Someone searching a checklist wants the checklist.

9. Delete anything that exists to hit a word count. Scaled content abuse is about volume without value. Ten empty posts a week is a problem whoever types them.

10. Read it aloud. Where you stumble, readers skim. Cut the warm-up lines, the stacked adjectives, and any sentence you’d never say to a colleague.

Gate 3: Search (steps 11 to 15)

11. One page, one primary query, an honest title. Search your own site first. If another page already targets the query, merge or differentiate before publishing.

12. Hand-check titles, meta descriptions, alt text and schema. If a script wrote them, a person reads them.

13. Link on purpose. Up to the broader guide, sideways to related cases, and make the new page reachable within two clicks.

14. Make it quotable. Question-style subheadings, a direct answer under each, a table where a table is clearer. A page must be indexed and snippet-eligible to appear in Google’s generative AI features, so a stray nosnippet tag can undo good work.

15. Run the dull technical pass. No accidental noindex, mobile layout fine, images compressed, no layout jumps. Search Console’s URL Inspection tool takes five minutes.

Gate 4: Trust and ship (steps 16 and 17)

16. Real byline, a plain line on how AI was used, and a way to report errors. CNET’s first mistake wasn’t the tool; it was hiding it. Being discovered turned a quality problem into a trust problem.

17. Check your publishing pace, then schedule two follow-ups. If you ship more pages than a person can genuinely review, that’s the pattern the spam policies describe. Put one check at a week and another at a month: indexing, impressions, and any page earning nothing.

Score your page

Give one point per step you can honestly say you did. Be strict.

ScoreVerdict
16 to 17Publish
13 to 15Fix the gaps first, especially any in Gate 1
10 to 12Hold it. Something important was skipped
Under 10Rewrite, don’t polish

Any miss in Gate 1 caps the page at “hold,” whatever the total.

Which step would have caught which failure

FailureGate 1Gate 2Gate 3Gate 4
CNET: errors and non-original phrasing1, 3, 516
Deloitte: fake citations1, 216
Aug 2026 cases 1 and 2: programmatic and AI text at scale6, 7, 91117
Aug 2026 case 3: 1.5M thin URLs6, 911, 1317

The 10-minute version

Some days you’ll have ten minutes. Do these, in order: step 2, step 1, step 3, step 12, step 16. Verify the links, verify the facts, redo the math, hand-read the title and meta, put a name on it. That covers the failure patterns behind CNET and Deloitte, though not the volume problem. For that, only step 17 helps, and it needs honesty about your pace.

What this checklist can’t do

It can’t make a page good. It catches the preventable mistakes: false facts, fake sources, thin pages, hidden AI use. Whether the page deserves to rank still depends on whether you have something real to say.

It also can’t tell you how much review is enough. Google doesn’t set a minimum. A reasonable floor is that a named person verifies every claim a reader could act on.

FAQ

Does Google penalize AI-generated content? Not as a category. Its policies target mass-produced, low-value pages whatever the method, and Ahrefs found no meaningful link between AI share and ranking position.

Do I have to disclose AI use? Google doesn’t require a label for ordinary text. CNET’s case shows the cost of being discovered instead of being upfront.

How many posts per week is safe? There’s no safe number. The test is whether each page got real review and adds something new.

Can a site recover from a spam update? According to Glenn Gabe, yes, but typically after months of cleanup that Google can see, and not in every case.

Should I use AI for titles, meta descriptions and alt text? You can, but Google’s October 2026 guidance, as reported, treats manual review as critical for AI output in general. Read every one before it ships.

Sources

Facts in this post were checked on November 14, 2026. Google’s guidance and enforcement change often, so check the linked pages before acting on any single figure.

Leave a Reply

Your email address will not be published. Required fields are marked *