The Google updates of 2023-2024 — Helpful Content Update, Scaled Content Abuse policy, and the March 2024 spam policy refresh — killed the business model of "20 AI articles per day, $5 each, hope for traffic." A lot of sites lost 60-90% of their traffic and never recovered. By 2026 the rules are clearer than they were, and AI is back in editorial workflows — just not as the only step.

This guide is the working policy we use, distilled from talking to publishers who survived 2024 and watching what Google actually penalises.

What Google is actually penalising

The Scaled Content Abuse policy targets scaled, low-value content created primarily to manipulate search rankings. The key phrase is "primarily to manipulate." Three signals matter most:

  1. Originality: does the page say something that wasn't already in the top 10 results?
  2. Demonstrated expertise: does the author / publisher show evidence of actually using or studying the thing they're writing about?
  3. User satisfaction: measured indirectly via dwell time, return visits, branded search.

You can use AI heavily on the production side and still hit all three. You can also write entirely by hand and miss all three. The HCU is not "AI bad, human good" — it's "useless content bad, useful content good." AI just made useless content much cheaper to produce, which is why so many sites got penalised at once.

What works in 2026

Use AI for the parts it's good at

  • Research summary and synthesis — Gemini Deep Research or Claude reading 20 papers and summarising the consensus is genuinely useful as a first draft.
  • Drafting from a detailed brief — if you supply a strong outline, source quotes, your point of view, the AI can produce a competent first draft.
  • Editing and polishing — tightening sentences, removing filler, improving flow. This is now table stakes.
  • Format transformations — turning a 2,000-word piece into a 200-word LinkedIn post or a 60-second script.
  • SEO scaffolding — meta descriptions, alt text, FAQ schema, internal link suggestions.

Keep the human in the loop where it matters

  • The angle / thesis — the take has to come from a human. AI averages; it cannot have an opinion that's worth reading.
  • Verification — every concrete claim, statistic, or quote needs human-eyeballed primary sourcing. AI hallucinates citations with absolute confidence.
  • Voice — the personality has to be a real person's, or at least consistent enough to feel like one. AI default voice is the average of the corpus, which reads as nobody.
  • Examples — real, specific, named. "A coffee shop in Portland uses X" beats "many small businesses use X."

The workflow that actually ships

For long-form editorial (1,000+ words), this is the loop we recommend:

  1. Human writes the brief — target audience, angle, must-cover points, three sources to draw from, length target. ~15 minutes.
  2. AI drafts the structure — outline with H2s and key claims under each. ~5 minutes.
  3. Human reviews and rewrites the outline — this is where you add your take. ~15 minutes.
  4. AI drafts each section — fed the outline, the brief, and the source material. ~5 minutes total.
  5. Human edits aggressively — tighten, add specifics, replace generic phrasing, verify every fact. ~60-90 minutes.
  6. Final pass — read aloud. If a sentence sounds like AI, replace it. ~15 minutes.

End-to-end, this is 2-3 hours for a 1,200-1,500 word piece. Pure-AI production was ~10 minutes. The new workflow is slower but ranks; the old one is dead.

Tools that fit this workflow

  • Briefing: Frase, Surfer SEO, Clearscope — for SERP analysis and content gaps.
  • Drafting: Claude (long-form), ChatGPT (structured), Gemini (research-anchored).
  • Editing: Claude or ChatGPT for the prose pass; Hemingway Editor for the readability pass.
  • Fact-checking: Perplexity, Phind, or just the source URLs in browser tabs.
  • Schema and meta: any frontier chat can produce JSON-LD, meta descriptions, alt text reliably.

Disclosure

Google's official position is that AI-assisted content is fine; you don't have to disclose AI use to rank. But disclosure helps you. A clear AI-assistance statement on your editorial policy page (like ours) does three things: signals to readers that you take quality seriously, satisfies advertisers' AI-content policies (AdSense in particular), and gives you a documented standard to refer back to when an article is questioned.

The reverse — pretending the content is fully human-authored when it isn't — is now actively risky. AI detection tools are unreliable, but readers' BS detectors are getting sharper, and a credible "we used AI for X but not Y" beats an indefensible "this was all human, we promise."

Red flags that get you penalised

  • Publishing more than ~3-5 articles per day on a small site — volume alone is a signal.
  • Articles that just rephrase the top SERP results without adding anything.
  • Generic "expert" bylines with no real human behind them. (See our editorial policy for how we handle this — anonymous editorial team, not invented experts.)
  • AI-generated images that don't add information (decorative filler from a stock-style model).
  • Internal links that don't make sense for the reader (just SEO-link-graph manipulation).

What we'd do if starting a content site in 2026

  1. Pick a narrow topic where you have real expertise. Don't try to outrank Wirecutter on everything.
  2. Publish slower than you think you should. Two excellent posts per week beats 50 mediocre ones.
  3. Show your work. Specific examples, named tools, screenshots, dates, prices. These are signals of actual knowledge.
  4. Build an email list early. SERPs are volatile; subscribers are not.
  5. Use AI to shorten the time from "I know what I want to say" to "draft on the page." Don't use AI to skip the "I know what I want to say" step.

The TL;DR: AI is a tool for editorial efficiency. It is not a content engine. Sites that treat it as the second got penalised; sites that treat it as the first are doing fine.