AI-Assisted Content Google Takes Seriously: The Checklist I Actually Use

As of September 2026, Google's position on AI-assisted content is the one it published on February 8, 2023: "Our focus on the quality of content, rather than how content is produced, is a useful guide." The pages that get ignored or removed are not the ones a model touched. They are the ones nobody checked, nobody signed, and nobody could have written from experience. The 12 checks below are what I run on every piece on this site and on the 92-page publication WBX Creative built for a dealership client, and they are the difference between AI-assisted and AI-generated.

Every Google quote here comes from a Google page with the date it was updated, because the guidance changes, and the blog posts about it change faster.

How to think about it

Google's spam policies, updated August 28, 2026, define scaled content abuse as "many pages" generated "for the primary purpose of manipulating search rankings and not helping users," and add "no matter how it's created." The generative-AI guidance added May 21, 2025 says the tools "can be particularly useful when researching a topic, and to add structure to original content." Read those together and the rule is plain. AI as a research and structuring tool on content with an original core is fine. AI as the source of the content is the thing being pruned.

The Search Quality Rater Guidelines, current as of September 11, 2025, draw the same line from the other side. Section 4.6.6 rates as Lowest any main content that is "copied, paraphrased, embedded, auto or AI generated, or reposted" with "little to no effort, little to no originality, and little to no added value," and in the same section says "the use of Generative AI tools alone does not determine the level of effort or Page Quality rating." Effort and originality are being judged. The tool is not.

So the checklist is not about hiding the tool. It is about putting effort, originality, and a person behind every page, visibly.

Using AI doesn't give content any special gains. It's just content.

1. Every fact traces to a source you can name

Before a word of prose exists, a research pack exists: every number, price, version, policy, and quote with the URL it came from and the date it was read. Anything the researcher could not confirm on a primary page is marked UNVERIFIED, and on the dealership site the writer's local checker fails a draft that still contains that word. The sources render at the bottom of every piece here, and prices and versions carry "as of September 2026" because they will not stay true.

A model will produce a confident number on request. The fix is refusing to type one you cannot point to.

2. Put something on the page only you could know

Experience is the first E in E-E-A-T, and it is the one a model cannot supply. On this site that means the 216 MB of images that became 42 MB on my own portfolio, and the 59 mislabeled titles a caption pass found in it the next day. On the dealership publication it meant what is on the showroom floor this week, by count, from a dated inventory snapshot.

Google's AI optimization guide, last updated July 10, 2026, puts it bluntly: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model." The February 2026 Discover update said the same from the ranking side, favoring "in-depth, original, and timely content from websites with expertise in a given area," judged topic by topic. If the page would read the same with the byline removed, it fails this check.

3. Byline a real person, and mark them up

Google's helpful-content guidance, updated December 10, 2025: "We strongly encourage adding accurate authorship information, such as bylines to content where readers might expect it." The 2023 post adds the corollary: "Giving AI an author byline is probably not the best way to follow our recommendation."

Google's Article structured-data page, updated September 8, 2026, says "In the author.name property, only specify the name of the author," recommends author.url, and adds that "If the URL is an internal profile page, we recommend marking up that author using profile page structured data." The ProfilePage type is "designed for any site where creators (either people or organizations) share first-hand perspectives." Every piece here carries an Article node whose author resolves to my profile page, which carries ProfilePage schema of its own.

4. Should you disclose that AI helped?

Yes, where a reader might wonder. The helpful-content self-assessment asks: "Is the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways?" The 2023 post says disclosures "are useful for content where someone might think 'How was this created?'"

I disclose on the launch note for this section and keep a provenance field in every source file naming the model, the date, and the packs it drew from. That field is for the two reviewers and for me. The sentence is for the reader.

5. Draft by default

Every machine-drafted page starts as status: draft, and a draft cannot publish. On the dealership site, production builds exclude drafts, staging renders them under a banner, and the content schema refuses to mark a page approved without both reviews recorded. Publishing becomes a separate decision from writing, made by someone who is not the writer.

The alternative is what the losses look like. The March 2024 manual-action wave deindexed 837 of 49,345 tracked sites in one analysis. Those sites did not have a draft state. They had a publish button.

6. Two reviews by people who did not write it, recorded

The fact-check pass traces every claim to its source, fetching the page when the pack is not enough. The editorial pass reads for invented experience, filler, duplication, structure, and voice. Each records a qc entry with the stage, the reviewer, the date, a verdict, and notes on what was fixed or cut. The writer never reviews their own page.

What that catches is the subject of the Outpost case study: a spec quoted for a trim the manufacturer never published it for, an arithmetic slip in an engine comparison, a course-provider count wrong by two, a warranty term nobody could confirm. None were visible to a reader. All were visible to a reviewer with the source open.

7. Ban the vocabulary

Every editor knows the words that mark a page as unedited machine output. My content checker greps for them and warns:

delve · tapestry · testament to · seamless · robust · elevate · unleash · game-changer
leverage (as a verb) · navigate the · landscape of · in today's fast-paced · whether you're a
look no further · in conclusion · it's worth noting · let's dive in

It also warns on British spellings and exclamation points in body copy, the second tier of tells. They appear when nobody with taste read the draft, and a reader feels that without naming the word.

8. Answer in the first 100 words

The first paragraph contains the topic, the year, and the answer to the question in the title. The H2s are phrased the way a person types them, two to four of them as questions. Data goes in a table. Anything technical goes in a code block with a language tag.

That structure is for people, and the answer engines reward it too. Bing's AI Performance guidance, published February 10, 2026, advises headings, tables, FAQs, and to "support claims with evidence." Google's AI features documentation, updated December 10, 2025, says "There are no additional requirements to appear in AI Overviews or AI Mode." A clear page is the optimization.

9. Is FAQ schema still worth adding?

Only for the readers. The FAQ at the bottom of every piece here is three to six real searcher questions with a complete answer in the first sentence. It is still marked up as FAQPage. It will not earn a search feature. Google's Search Central changelog records "Added a deprecation notice to the FAQ rich result documentation" on May 8, 2026, and on June 15, 2026 the documentation was removed with the note that "The FAQ rich result feature is no longer shown in Google Search results." Keep the questions because a reader and an answer engine can quote them. Drop the expectation.

10. Schema that matches the visible page

Google's May 21, 2025 post on AI search has two instructions that belong together: "Focus on making unique, non-commodity content" and "Make sure structured data matches the visible content." Every page here emits one JSON-LD graph: the Organization, the Person, the WebSite, then a WebPage or ProfilePage, a BreadcrumbList, FAQPage where there are questions, and Article on posts, with the author resolving to a real profile. Nothing in the graph claims something the page does not show.

The dealership site adds a rule worth copying: no Product or Offer markup on pages that link out to buy, because Google disallows it there and the vehicle rich results were retired in 2025. Schema describes the page. It is not a wish list.

11. Gate the build

A script runs after every build, parses the finished HTML of every page, and fails the deploy on thin pages, placeholder text, orphans, broken links, missing titles or canonicals or Open Graph images, JSON-LD that does not parse, a noindexed page in the sitemap, or a draft that leaked into production. There is no bypass flag, on purpose. The Next.js checklist lists what mine checks on this site; the Outpost's runs ten checks on Astro output. A reviewer reads for what is wrong. The gate reads for what is missing, on every build, whether or not anyone remembers to look.

12. How many pages should you publish?

Fewer than you could. The helpful-content self-assessment asks, in so many words, "Are you using extensive automation to produce content on many topics?" The site reputation post from November 19, 2024 adds that Google does not "simply take a site's claims about how the content was produced at face value." The dealership site could have had a page for every unit in inventory. It has none, by rule, and a hub page needs at least three units in stock to exist at all.

Every page you do not publish is one you do not have to check. Scale is not the goal. Being right at whatever scale you can review is.

Mistakes I see

• The fake staff writer. An AI byline with a stock photo. Sports Illustrated's "Drew Ortiz" in November 2023 is the cautionary tale, and Google's guidance names the practice.
• Chasing detectors. Rewriting a true, sourced paragraph so a detector scores it lower. The detectors are unreliable in both directions.
• Files for robots. An llms.txt, a Markdown mirror, a "chunked" copy of the page. Google's 2026 AI optimization guide says "You don't need to create new machine readable files, AI text files, markup, or Markdown," and warns against "Rewriting content just for AI systems."
• Topics you do not work in. A design studio publishing on tax law because a model can. Expertise is judged topic by topic, by Google and by readers.
• FAQ stuffing. Twelve questions for a rich result that no longer exists.
• Disclosure as the fix. A sentence admitting AI helped does not make a wrong number right. It is check four of 12, not the whole list.

What it costs / what to expect

Expect the review to be the larger cost. Drafting got cheap. On the dealership site the fact-check pass fetched OEM spec pages, statutes, agency pages, and dealer sheets for each guide, and the editorial pass rewrote passages by hand. That is the work now.

Expect nothing fast from Google. The pages need to be crawled, indexed, and judged, and Search Console is the only honest scoreboard. As of August 31, 2026 its generative-AI report shows impressions from AI Overviews and AI Mode worldwide, without clicks. SE Ranking's 2025 analysis found pages updated in the previous three months were twice as likely to be cited by ChatGPT, which argues for keeping pages current, not for gaming a date.

If you want this run on your own pages, creative services is where the writing and editing lives, and consulting is the right door if you have a team producing content and need the process around it.

Where to go next

The Outpost case study shows the whole pipeline on a 92-page site, with the ledger of what the reviews caught. The launch note for this section is the one-page version of how pieces here are checked. The Next.js checklist has checks 10 and 11 in code.