Using AI in SEO Without Wrecking Your Site: The Jobs It Does Well and the Ones It Cannot Be Trusted With

The riskiest way to use AI for SEO is also the way most people start: type a keyword into a chat window, copy the answer onto a page, hit publish. It feels like production. It is actually the one use case Google has explicitly built machinery to detect and demote. Meanwhile the uses of AI that measurably improve search work, the unglamorous ones like schema generation and corpus-wide QA, barely appear in the tutorials. This page inverts that. We run an agency on orchestrated AI and we will show you exactly where the line sits, because our own domain lives or dies by respecting it.

What AI actually changes about SEO work

AI changes the cost of production, not the standard of what ranks. The bar Google describes in its documentation, original information, demonstrated experience, a named accountable author, has not moved; what moved is that meeting the bar's surface appearance became nearly free, so the surface stopped counting as evidence. A page that merely reads well proves nothing anymore, because reading well is now the default output of a $20 subscription. Search Engine Land's April 2026 roundup of practical AI uses in SEO lists twenty applications, and the pattern across the useful ones is consistent: AI compresses the hours between question and draft, and the value survives only when a human owns the judgment at the end of the line. Treat the tool as a hundred junior analysts, not one senior strategist, and the whole topic gets simpler.

Where does AI genuinely help? Five jobs it does well

The five highest-return AI jobs in SEO share one trait: their output can be verified faster than it could be produced by hand.

• Research synthesis. Reading forty sources, ranked competitors, forum threads, PDF studies, and returning a structured map of what is claimed, what is sourced, and what is missing. The human checks the map against a sample of the sources, not all forty.
• Entity extraction. Pulling every person, product, place, and concept from a page or a competitor set, so you can see which entities your content covers and which it skips. This is mechanical work models do more reliably than tired humans.
• Schema generation. Structured data is tedious to hand-write and trivially machine-checkable, which makes it close to an ideal AI task. More on this below, with markup you can copy.
• Brief building. Turning research into a working outline: the questions to answer, the claims that need sources, the entities to include, the queries each section serves. The brief is scaffolding; a human still decides the argument.
• QA at scale. Sweeping hundreds of pages for broken links, missing meta descriptions, orphaned pages, contradictory claims between articles, and duplicated phrasing. No human team re-reads a 400-page site monthly. A model can, every week.

Notice what all five have in common: none of them is "write my article for me." They sit before and after the writing, where speed compounds and errors are catchable.

Where raw AI output fails, and why Google notices

Unedited model output fails in three specific, repeatable ways. First, unverified facts: models produce plausible statistics and confident attributions that dissolve under a single source check, and on YMYL topics a fabricated number is not a style problem, it is a liability. Second, generic prose: a model prompted with only a keyword returns the statistical centre of everything already published, which is by definition content with zero information gain, the exact quality Google's helpful content signals exist to filter. Third, volume without accountability: the marginal cost of page 500 is the same as page 5, so unsupervised pipelines drift toward scale, and scale without review is the fingerprint of what Google calls scaled content abuse. Our research team documented the wreckage in the AI slop warning: sites that automated publishing and lost most of their visibility in a single update cycle. Whether AI-drafted pages can rank at all is a separate question with a more nuanced answer, and we tested it directly in does AI content rank in Google. Short version: yes, when a human made it worth ranking.

The January 2025 rater criteria: the line your content has to clear

In January 2025, Google updated its Search Quality Rater Guidelines, the manual its human evaluators use to benchmark result quality, and drew the AI line in writing. Raters are instructed to assign the Lowest rating when the main content of a page is auto-generated or AI-generated and no original or unique value has been added. Read that sentence carefully, because both halves matter. AI-generated is not the trigger on its own; the trigger is AI-generated plus nothing added. Raters do not set rankings directly, but their scores train the classifiers that do, so the guideline describes the target Google's algorithms are being tuned toward. The practical test for any page you are about to publish is the rater's test: could an evaluator point to something on this page, a first-hand observation, an original figure, a genuinely useful synthesis, that did not exist before you made it? If the honest answer is no, the page is a demotion candidate regardless of how clean the prose reads.

The human-in-the-loop workflow, step by step

Here is the sequence that survives the criteria above. It is the one we run, scaled down to a form any business can execute.

1. Research before drafting. Live searches on the target query. What ranks, what it covers, what it gets wrong, what it misses. AI can compile this; a human confirms the gaps are real.
2. Ground the draft in evidence. The model writes from a brief loaded with your data, your customer questions, and verified sources, never from a bare keyword. Input quality is the single biggest lever on output quality.
3. Verify every factual claim. Each statistic traced to a named, dated source. Anything unverifiable gets cut, not softened. This step is non-negotiable and it is the one automated pipelines always skip.
4. Add what only you know. A first-hand observation, a pattern from your own client work, a local detail. This is the "original value" the rater guidelines demand, and no model has access to it.
5. Edit for voice and rhythm. Uniform sentence length and hedge-everything conclusions are machine tells. The de-AI-ing pass is a craft of its own; the sibling piece on how to humanize AI content walks through it line by line.
6. Check the corpus, not just the page. Before publishing, compare the new page against your existing pages for repeated phrasing and duplicated structure. Individually fine pages can still form a collectively templated pattern, and pattern is what classifiers see.
7. Publish under a named author who can be held to it. Accountability is a quality signal precisely because slop operations cannot afford it.

Wondering how your current pages would score against those seven steps? That is a twenty-minute check for us, and asking costs nothing.

Schema generation: the clearest AI win in SEO

Structured data is where AI assistance is closest to pure upside, because the output is machine-validated. A model drafts the JSON-LD in seconds, you paste it into Google's Rich Results Test and Schema.org's validator, and either it parses or it does not. There is no judgment call to get wrong, only a checkable artifact. This work maps to Vector 6, Structure, in our twelve-vector methodology: markup that lets search engines and answer engines read your pages with high confidence. Here is the kind of block a model can draft for a local service business in one prompt:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Plumber",
  "name": "Example Plumbing Co",
  "url": "https://exampleplumbing.ca",
  "telephone": "+15195550147",
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Brantford",
    "addressRegion": "ON",
    "addressCountry": "CA"
  },
  "areaServed": "Brant County",
  "sameAs": [
    "https://www.facebook.com/exampleplumbing"
  ]
}
</script>

One caution from Google's own AI-features documentation: structured data helps machines parse your pages, but adding schema "because AI needs it" is a debunked tactic; generative search features do not require it. Do schema for the parsing confidence, not as a magic token.

QA at scale: the use case almost nobody writes about

The least-discussed AI application in SEO is auditing your own corpus, and in our experience it returns more than another new article does once a site passes a hundred pages. Sites accumulate rot: pages contradicting each other after a policy change, internal links pointing at retired URLs, meta descriptions that repeat, five old posts making a claim you have since corrected. Every one of those is a quality and trust signal decaying quietly. A model tasked with reading the full site and flagging contradictions, duplicates, and dead references turns a quarterly consulting project into a weekly automated sweep with a human reviewing the flag list. This is also the discipline that catches the failure mode described earlier, where individually acceptable pages converge into a detectable template. Page-level review cannot see convergence. Corpus-level review is the only thing that can, and it is exactly the kind of exhaustive, boring reading machines are for.

How Formative Digital itself uses AI, stated plainly

We are not going to write a page about AI ethics in SEO and pretend our articles are typed longhand. Formative Digital runs on an orchestrated system of specialized AI agents, we call it the Formative Forces, in which separate agents handle research, citation verification, markup, and QA, passing work between stages under human oversight. That is a different machine from one person pasting chat output onto a page: every stage has an engineering constraint attached, live research before drafting, sources verified at write time, corpus-wide uniqueness checks after, and Matt signs the result under his own name. The distinction we are drawing on this whole page, tools under constraints versus tools instead of constraints, is the distinction our own operation is built on. We publish this openly because an agency selling AI-search visibility while hiding its methods would be running the exact opacity play the industry is notorious for. You can judge the standard by its output; this page is a sample of it. More of the research behind the approach lives in our research library.

What Google's own documentation says, including the debunked tactics

Google published its guide to appearing in generative AI features in 2025, and it is the closest thing to first-party rules this field has. The positives are the familiar fundamentals: unique, non-commodity content, demonstrated first-hand experience, pages built for people. Just as useful is what the guide shoots down, because several of these are still being sold as services in 2026. An llms.txt file: Google states it does not use one. Rewriting existing pages "for AI systems": no effect. Chunking content into fragments "for retrieval": not a thing. Near-duplicate pages targeting query variations: an actual spam policy violation, not a tactic. If a vendor pitches any of those, you have learned something valuable about the vendor. The honest summary of Google's position is that AI-era optimization is quality work under a new distribution model, and there is no separate bag of tricks. That is inconvenient for people selling tricks and convenient for businesses willing to do the work once, properly.

What to do first if you are starting this week

Sequence matters more than tooling. Start with the two AI jobs that cannot embarrass you: run an entity and gap analysis of your top competitors, and sweep your existing site for broken links, duplicate metadata, and contradictions. Both produce verifiable output and touch nothing public. Next, pick one page topic where you hold genuine first-hand knowledge and run the seven-step workflow above, timing yourself; expect the verification and editing steps to take longer than the drafting, which is the correct shape. Only after that page ships should you think about cadence, because cadence built before quality control just accelerates the publishing of mistakes. Results depend on your industry, competition, and existing digital presence, and a realistic window for measurable organic movement is three to six months. Anyone quoting faster than that with AI as the reason is describing the volume trap, not a strategy.