AI-Assisted SEO: A Human-in-the-Loop Workflow

A language model is fast at the parts of SEO that are about shape: grouping keywords, proposing an outline, writing a first draft, tightening sentences. It is unreliable at the parts that are about truth: your current price, the year you launched, what a competitor offers, where a statistic came from.

Split the work on that line.

If you are choosing software for the drafting half, our guide to AI SEO writing tools covers how to set them up safely.

Google's spam policies define scaled content abuse as being "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The first example it gives is "Using generative AI tools or other similar tools to generate many pages without adding value for users."

Google's guidance on AI content is just as direct about the tool: "Appropriate use of AI or automation is not against our guidelines." The violation is volume without value, not the model that wrote the words.

What the policy looks like in practice

OtterlyAI tested the shortcut in 2026. Two new domains each published roughly 1,000 AI-generated posts. Google removed both from its index, one after about 14 days and the other after about 90, and Otterly reports no manual action in Search Console on either. The removal was silent.

The same test showed the indexes diverge. Bing kept both sites, and Microsoft Copilot, which grounds on Bing, kept citing them. Google's AI surfaces almost never did once the pages were gone. That fits Google's own rule that a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode.

Treat this as one illustration, not proof. It was one pair of sites on a low-stakes topic, with no human-written control, run by a vendor that sells AI visibility tracking.

What it means for you

The review step is the value Google is asking for. A page a human has checked, corrected and decided is worth publishing is a different object from a page generated in bulk.

AI-drafted pages fail in predictable places. Models fill gaps with plausible numbers, repeat last year's pricing, merge your product with a competitor's, and invent studies that sound real.

Those errors cost more now than they used to. AI platforms already get specific facts wrong about specific brands. A wrong fact on your own site gives them a source for the error.

The reverse also holds. Stating a specific fact (a price, a number, a one-line definition) on a page an engine can fetch gives it a sentence it can quote. That is the only on-page content pattern with a positive experimental signal in our evidence base, and it rests entirely on the fact being right.

Two checks follow from this:

• Consistency across your site. The same fact stated differently on your pricing page, about page and footer is a likely cause of an AI answer getting it wrong. Managing this is part of treating your brand as an entity.
• Visibility of the fact itself. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so they can't read client-side rendered content. A correct price that only appears after JavaScript runs may be invisible to them. Googlebot renders JavaScript, so this gap is specific to most non-Google AI crawlers.

Be plain about the split. It stops you spending review time on things that do not move.

In your control

• What you publish, and how fast.
• Whether every fact on the page is correct and consistent with your other pages.
• Whether the content that matters is in the HTML before JavaScript runs.
• Whether your robots.txt allows each AI search crawler you care about (rules are per bot: GPTBot governs OpenAI training, OAI-SearchBot governs ChatGPT search).
• Whether your CDN or firewall actually lets those crawlers through.
• Whether you use noindex, nosnippet or data-nosnippet by accident.

Not in your control

• Whether Google indexes a given page. Google says indexing and serving isn't guaranteed.
• Whether an engine searches the web for a given question at all. As of October 2026, retrieval is discretionary per query on every major platform.
• Which retrieved pages the model chooses to cite.
• What a model learned before its knowledge cutoff.

Rewriting a draft to make it more citable does not cross that line. The C-SEO Bench study found most conversational SEO methods "largely ineffective" and frequently harmful to a document's ranking. Strong classic SEO also does not guarantee AI presence, which we cover in whether strong SEO means you are visible in AI search.

Run this on every page before it goes live. It takes longer than generating the draft, which is the point.

1. Break the draft into single claims. Highlight every sentence that states something checkable: a number, a date, a name, a comparison, a cause.
2. Check prices, numbers and dates against your source of truth. Your billing system, your contracts, your own records. Not another web page, and not the model.
3. Trace every statistic to its primary source. If you cannot find the study, delete the number. Never round it, estimate it or keep it with a vague attribution.
4. Strip unsupported claims about competitors and people. No invented quotes, testimonials or customer names.
5. Confirm the byline is real. Do not publish a fictional expert. In the Otterly test, the site with an invented wildlife biologist was removed first.
6. Ask what the page adds. If it repeats what your other pages or the top results already say, merge it or cut it. This is the value test the spam policy turns on.
7. Check consistency. Every fact on the new page should match the same fact everywhere else on your site.
8. View the raw HTML. Confirm the key facts appear without JavaScript.
9. A named person approves publication. Someone owns the decision and the corrections.

For a lighter version suited to free tools, see the free SEO writing fact-check workflow.

Do the work in this order. Each step removes a block before you add more pages.

1. Fix access first. Check that robots.txt allows the search crawlers you want, then check that your firewall does not turn them away. A robots.txt that allows a crawler while your CDN returns 403 removes you quietly. The AI Crawler Checker requests a page as each AI crawler and shows which ones get blocked.
2. Correct the facts on pages you already have. Pricing, product, about and contact pages first. These are the pages engines are most likely to quote about you.
3. Add new AI-assisted pages at the pace review allows. If your reviewer can check three pages a week, publish three pages a week.
4. Check index status engine by engine. Look at Google Search Console and Bing Webmaster Tools separately. One healthy index tells you nothing about the other.
5. Look at what AI engines actually say. The free checker on this site shows whether engines name you for up to three buyer questions. It is a snapshot, so repeat it after changes.

Teams running autonomous agents for parts of this should read an AI agent SEO workflow with human checkpoints. For adapting pages you already rank with, see how to adapt your existing SEO for AI answers.

• Publishing faster than anyone can review. Volume without value is the exact pattern Google's policy describes.
• Inventing authors or credentials. Google says E-E-A-T isn't a specific ranking factor, and a fake persona adds risk with no evidence of benefit.
• Keeping a statistic because it sounds right. If the source cannot be found, the number goes.
• Adding schema or an llms.txt as an AI fix. Google says structured data isn't required for generative AI search, and Google Search ignores llms.txt. Existing markup does no harm, but neither is a lever for AI answers.
• Rewriting in an "AI-friendly" style or chunking pages. Google's AI optimization guide says you don't need to write in a specific way for generative AI search. Chunking is done by the retrieval system, as the document chunking study shows.
• Padding with long-tail keyword variants. Google says you don't need to capture every variation, and the original GEO paper found keyword stuffing lowered its own visibility metric.
• Allowing GPTBot and assuming ChatGPT search is covered. GPTBot governs training. ChatGPT search uses OAI-SearchBot.
• Hiding content from users but showing it to AI crawlers. Cloaking and hidden instructions are manipulation. They are a risk to your index status and your reputation, and they are never a legitimate tactic.