AI SEO Articles: The Quality Bar Before You Publish

Google states its position on AI content plainly. Its guidance says "Appropriate use of AI or automation is not against our guidelines." The line it draws is scaled content abuse, defined as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." Its first example is "Using generative AI tools or other similar tools to generate many pages without adding value for users."

So the tool that drafted the article isn't the test. Three things are:

• Verified facts. Every number, name, date and claim traces to a source you'd stand behind.
• Original value. The article says something the top results don't.
• Rendered in HTML. The content that matters exists in the page source a crawler receives, not only after JavaScript runs.

AI drafts tend to fail the first two by default. Language models state figures confidently whether or not they're true, and they paraphrase what already ranks. The third is a site problem the drafting tool never sees. If you're comparing tools, our guide to what AI SEO writers produce and miss covers the drafting side. This page covers the review.

Answer engines build responses from sentences. "The Starter plan costs $49 a month" can be lifted straight into an answer. "We offer flexible pricing to fit every budget" can't, because it answers nothing.

Stating a specific fact on a page that can be fetched is the only on-page change with a positive experimental signal in the evidence we track. The 252,000-trial factorial study (arXiv 2605.25517) found that "Including explicit price information and a recent timestamp also helps consistently," while "formatting-only edits have little impact." Read that with its limits: the study used a synthetic corpus, two candidate documents per trial, and documents already inside the model's context. Specifics help a page that has been retrieved. They don't get a page retrieved.

The generic tactics fare worse. C-SEO Bench (NeurIPS 2025) found most conversational SEO methods "largely ineffective" and frequently harmful to a document's ranking. The original GEO paper found keyword stuffing lowered its own visibility metric (17.8 against a 19.5 baseline). Google's optimization guide says you don't need to write in a specific way for generative AI search, or worry about long-tail keyword variants.

Filler also isn't neutral when it contains invented specifics. A made-up statistic is worse than none, because an engine can quote your wrong number back to buyers as fact. The practical rule: replace vague sentences with verified specific ones, and delete anything you can't verify. Our playbook on publishing specific, fetchable facts walks through which facts to state and where.

Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files (11.50% and 23.84% of their requests) but "don't execute them. They can't read client-side rendered content." The evidence base we use records that independent testing in June 2026 found the same for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Meta-ExternalAgent and Bytespider. Googlebot does render JavaScript.

That creates a quiet split. An article loaded by client-side JavaScript can rank in Google and still be absent for ChatGPT search or Perplexity. The same applies to text inside tabs, accordions or "load more" blocks that fetch content after the page loads.

The check is simple. Fetch the raw HTML with JavaScript off and search it for the sentences that carry your key facts. If they're missing, the fix is server-side rendering or prerendering for that content.

Rendering is one gate. Two others sit beside it:

• Crawler permission. robots.txt rules are per bot. Blocking GPTBot stops training use. OAI-SearchBot controls ChatGPT search.
• Network reachability. A firewall or CDN can return 403 to a crawler that robots.txt allows.

The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.

An article passes through four stages before it shows up in an AI answer. Your checklist only reaches some of them.

The access stage is where a pre-publish check does the most good, because the failures there are binary. A page behind a 403, a noindex tag or client-side rendering simply isn't in the candidate pool.

Retrieval is per query and at the engine's discretion, as of October 2026. ChatGPT search rewrites the user's question into its own queries. Google may fan a query out across subtopics. You can't see those queries, so you can't optimize for them directly. Topical fit to real buyer questions is the most reproducible factor the research agrees on.

Brand prominence also sits outside the checklist. Large, well-known sources dominate citations across engines. A well-built article from a small brand can still be passed over.

Work through these in order. The early steps decide whether the article should exist at all.

1. List and verify every fact. Pull out each number, price, date, name and claim. Trace each one to a primary source or your own records. Delete anything you can't verify, rather than softening it.
2. Check it against your own site. Compare prices, plan names, features and company details with your pricing and about pages. A contradiction on your own site is a likely source of wrong AI answers about you.
3. Name the original value in one sentence. What does this article contain that the current top results don't? Your own data, your own process, real prices, a worked example clearly labelled as hypothetical. If you can't write that sentence, don't publish. See whether original data gets cited by AI for what the evidence supports.
4. Swap filler for specifics. Go paragraph by paragraph. Replace generic claims with a verified fact, or cut the paragraph.
5. Strip outdated tactics from the draft. Models trained on old blog posts repeat advice about schema, E-E-A-T scores and domain authority as if it were current. Remove it.
6. Check rendering. Confirm the key sentences appear in the raw HTML before JavaScript runs.
7. Check eligibility. No inherited noindex, nosnippet, max-snippet:0 or data-nosnippet on content you want shown. Canonical tag points to the right URL. robots.txt allows the search bots you care about.
8. Publish at a pace you can review. If volume outruns verification, slow down. The pre-publish review for scaled AI content covers review at volume.

OtterlyAI tested the opposite of this checklist in 2026. It loaded two new domains with roughly 1,000 AI-generated posts each, then tracked them for about four months. The details are in our write-up of the mass-publishing experiment.

Otterly reports that Google removed both sites from its index. The site with a fictional expert author box lasted about 14 days. The site with no author lasted about 90. Otterly found zero manual actions in Search Console on either domain, so neither owner got a warning.

Bing kept both sites, and Microsoft Copilot kept citing them. Of the 3,934 citations Otterly tracked for one site, 2,718 came from Copilot. Google's AI surfaces almost never cited either site once Google's index dropped them.

The test has limits. It ran one site per arm, had no human-written control, and used a low-stakes topic. Otterly sells AI visibility tracking, as does the publisher of this page. The lesson that holds is narrow: volume without value is the pattern Google's policy describes, and removal can arrive silently. For the volume question in general, see whether publishing more content helps AI visibility.

• Trusting the draft's numbers. Models produce plausible statistics and fake study names. Verify every figure or remove it.
• Inventing credentials. An author box for a fictional expert misleads readers. Google says E-E-A-T "isn't a specific ranking factor," and in Otterly's test the fake-expert site was removed first. Credit real people or none.
• Adding schema or llms.txt to get cited. Google says structured data isn't required for generative AI search, and FAQ rich results were deprecated in May 2026. Ahrefs' matched test found AI Overviews citations fell 4.6% after pages added schema. Google also says Search ignores llms.txt. Existing markup does no harm, but it isn't a citation lever.
• Rewriting for an "AI-friendly" style. Google rejects chunking as a requirement and says there's no ideal page length. Write for the reader.
• Hiding text for AI crawlers. Hidden instructions, cloaked content and prompt injection aimed at AI systems are manipulation. They risk removal and damage trust in your brand. Never use them.
• Treating more pages as more visibility. Our comparison of scaled AI content and people-first content sets out where each approach leads.