Most AI-for-SEO guides list ten things AI can do and skip the part that matters. Here is the exact process one operator and one AI agent run on this blog, intent to publish, what the AI leads, and the two human gates that decide whether a page ranks.
We run this blog with one operator and one AI agent. Twenty-something articles, a keyword pull most mornings, a draft by lunch. The part nobody writing "AI for SEO" guides will tell you: the agent does maybe 70% of the work and none of the 30% that decides whether the page ranks.
That 30% is the whole story. Everyone can prompt a model to write 1,500 words about their product. Google is now sitting on a few hundred million of those, and most of them rank nowhere. The teams getting value from AI in search are not the ones generating more. They are the ones who worked out which jobs to hand the machine and which to keep on a human desk.
This is that split, written from the desk.
AI for SEO means using language models to compress the slow middle of search work: topic sourcing, keyword research, reading the SERP, clustering topics, first drafts, internal linking, and schema. It does not mean autopilot publishing. In our practice two steps stay human: checking every claim against a real source, and editing the draft so it reads like a person wrote it. Those two gates decide rankings. Everything upstream of them is where AI earns its keep. The companion reads are SEO vs AEO vs GEO for the framework and where ChatGPT gets its information for what the engines actually cite.
The phrase hides two very different jobs that get sold as one.
Job one is research and operations. Pulling demand, reading what already ranks, grouping topics by intent, drafting, wiring internal links, generating structured data. Repetitive, high-volume, rules-based. A model is genuinely fast and good at this.
Job two is judgment. Deciding what is true, what is worth saying, and whether a sentence sounds human. A model is unreliable here, and confidently so. It will invent a statistic and cite it with a clean-looking source. It will write a URL that returns a 404. It will reach for the same eight adjectives every other model reaches for, and in 2026 those adjectives are a ranking liability rather than a flourish.
Most failures people blame on "AI content" are really one mistake: handing the machine job two.
So the useful question is not "can AI do SEO". Yes, it can do most of it. The useful question is where the human stays in the loop, and the honest answer is narrower than the tool vendors want it to be.
The shape is a fixed chain: intent, then keywords, then SERP, then the article, in that order. Seven stages run along it. The agent leads five, a human gates two.
One: listening for intent. Three times a day we run a search across Reddit, X and Threads for what is gaining velocity in our niche. The pipe can handle around 120 requests a minute; we use a fraction of that, because three passes a day is enough to catch a topic on the way up rather than after it has peaked. A live trend or a piece of news is the intent signal, and everything downstream hangs off it.
Two: keyword research and clustering. First we validate that intent against real search demand, because a topic spiking on social with no search volume is a post, not an article. Then we pull volume, difficulty and trend, fan the seed out into related queries, and group them by intent. The output is a map of which page should exist, which queries belong on the same page, and which would cannibalise something we already published. Last week the seed was "ai for seo". The map said one head guide, one tools listicle, and one service page, because the three carry different intent. This article is the first of those three, and that call came before a single sentence got written.
Three: reading the SERP before writing a word. We take the top ten results for the target query and classify them by page type. Guide, listicle, product page, forum thread. If eight of ten are tool listicles and you write a definition essay, you will not rank no matter how good the essay is, because Google has already decided what that query wants. Roughly half of the sources that appear in Google AI Overviews also rank in the organic top ten, per AIOSEO's 2025 dataset, so the classic SERP and the AI answer pull from the same shelf. Read the shelf first. The model does this read in seconds; the value is that it stops you writing the wrong page.
Four: drafting to a brief, not a prompt. The agent drafts, but against a structured brief: primary keyword, the sub-intents the SERP exposed, the internal links to place, the five concrete things a human knows that a generator does not. A blank "write me an article about X" prompt produces the generic page that already lost. A brief produces something with a spine.
Five: gate one, every claim against a real source. This is the first place a human takes the keyboard. Every percentage, every dollar figure, every "studies show" gets checked against a primary source or it gets cut. We keep a single file of verified statistics with the source URL next to each one, and nothing numerical ships unless it traces back to that file or to our own audit methodology. Models hallucinate numbers because a plausible number completes the sentence. The fix is not a better prompt. The fix is a person who checks.
Six: gate two, making it read like a person. Google's own research shows AI assistants cite content that is, on average, 25.7% fresher than what classic search surfaces, and the engines lean toward sources that read like genuine writing rather than template output. We run a humanize pass that strips the tells: the stock vocabulary, the rhetorical tics, the giveaway sentence shapes that mark a draft as machine-made. This is an editorial quality step. A page that reads like a person holds attention longer and gets quoted more, and both feed the ranking.
Seven: internal links, schema, distribution. Back to the agent. It places the cluster cross-links, generates the Article, FAQ, Breadcrumb and Person structured data, and updates the freshness date so the sitemap signals a real edit. Then the page goes out, gets submitted for indexing, and where it makes sense gets a syndicated copy with a canonical pointer home. AI handles all of it because it is mechanical and rules-bound.
Five machine stages, two human gates. That chain is the actual shape of AI for SEO that works.
One architectural detail makes it hold together: each stage runs as its own sub-agent with its own memory, not as one long prompt that tries to hold everything at once. The listening agent does not carry the drafting context. The fact-checker never sees the SERP-reader's working notes. Each mini-agent gets one job, the inputs for that job, and nothing else, and it hands a clean result to the next. Separating memory per stage is what stops the context bleed that turns long single-session prompts into mush, where the model half-remembers an earlier instruction and quietly contradicts itself three sections later. Seven small agents that each do one thing well beat one large agent trying to do all seven.
The cleanest way to think about it is per-task, not per-tool.
| Task | Who leads | Why |
|---|---|---|
| Trend and topic sourcing | AI | Social listening across Reddit, X and Threads, three times a day |
| Keyword volume, difficulty, clustering | AI | High-volume, rules-based, fast |
| SERP intent classification | AI | Pattern-matching the top 10 is what models do well |
| First draft from a brief | AI | Speed; the brief carries the judgment |
| Fact and source checking | Human | Models invent confident, wrong numbers |
| Voice and final edit | Human | Generic phrasing ranks and cites weaker |
| Internal links and schema | AI | Mechanical, deterministic |
| Deciding the page should exist at all | Human | Cannibalisation and strategy are judgment calls |
| Indexing, syndication, freshness | AI | Repetitive operations |
Hand the machine the volume. Keep the truth and the voice.
Three failure modes, all common, all avoidable.
Invented facts and dead links. A model completes sentences, and a specific number completes a sentence better than a vague one, so it produces a specific number whether or not it is real. Same with citations. Ship that unchecked and one wrong stat, caught by a reader or a competitor, costs you the trust the whole page was built to earn. Gate four exists for exactly this.
A voice the engines recognise. When every page in a niche is drafted by the same handful of models, they converge on the same cadence and the same words. That sameness is now a signal, and not a good one. Fresh, specific, human-sounding writing is what gets cited inside AI answers, which is the surface that actually sends qualified traffic in 2026. A draft that screams "generated" loses there twice: once on rank, once on the citation.
Thin pages at industrial scale. The cheapest thing AI does is produce more pages, so that is what most teams do with it. Volume without a sub-intent behind each page just dilutes a domain. AI tools cite and rank pages that cover several real sub-questions deeply, not pages that exist to hit a publishing quota. More is the trap most teams walk straight into.
None of these is a reason to skip AI. They are the reasons the two human gates are not optional.
Specific, because a vague tool list helps nobody. Our working stack is five layers.
DataForSEO is the data spine. Search volume, keyword difficulty, SERP scrapes, and full backlink profiles. The keyword and SERP stages run on it, and so does the competitor-backlink mining we use for link building. Budget around $50 a month of usage at our volume.
A social-listening layer across Reddit, X and Threads handles topic sourcing. It runs three times a day, surfaces what is gaining velocity in our niche before it peaks, then we validate each candidate against search demand so we only write the trends that people are also starting to search for.
Google Search Console and Google Analytics close the loop. GSC shows which queries we already rank for, which pages are indexed, and what sits one position away from page one. Analytics shows which articles actually pull traffic and convert, so next month's topic list is weighted by what worked rather than by what felt good to publish. There is no clean connector wiring these into the agent, so we gave it a different kind of access: a browser extension, claude-cowork, that lets the agent operate the GSC and Analytics dashboards through the frontend the way a person would. Nothing to integrate or keep in sync, the agent just clicks through the tools a human already uses.
Claude Code is the agent runtime. It runs the drafting, the humanize audit and the schema generation, orchestrates the chain stage to stage, and drives the extension. The subscription is $100 a month.
That is the whole rig. Notice what is missing: no single "AI SEO tool" that claims to do everything. The tool matters less than the process around it, and a lean stack with the two human gates beats an expensive one without them every time.
People searching "ai seo tools" usually want the ranked list of twenty products with prices. The closest thing we have published is the GEO dashboards pricing breakdown for the monitoring side, with why most of those dashboards won't get you cited as the companion critique. A dedicated AI-SEO-tools comparison, the full list and what each one costs, is the next piece we are writing. When it ships it will live here.
Smaller than most teams expect, because the edge sits in the process rather than in expensive software.
| Line item | Cost |
|---|---|
| Claude Code subscription | $100/mo |
| DataForSEO usage | around $50/mo |
| Social listening, GSC, Analytics access | included via the extension and existing accounts |
| Setup time | about 3 full working days |
The cash side is roughly $150 a month. The real cost is the build. Budget about three full working days of focused involvement to stand it up correctly: wiring the listening, writing the briefs, tuning the humanize audit to your own voice, splitting the stages into separate agents, and giving the extension access to the dashboards. Done properly once, the daily run takes an hour or two. Rushed, it produces the same generic output as everyone else, which is the genuinely expensive outcome.
Optimising your own pages is half the game. The other half is being present in the sources the engines read when they build an answer, and those sources are mostly not your website. Reddit alone is the single most-cited source across major AI engines, at roughly 40% of citations, and the top fifteen domains hold 68% of the share. You can run a flawless AI-for-SEO process on your blog and still be invisible inside ChatGPT if your brand never appears in those upstream sources. That is the work we do on Reddit GEO: putting the brand inside the conversations the engines quote. Page-level SEO and source-level presence are different jobs on the same chain.
Google documents the page-level side itself in its AI Optimization Guide, and AI Mode now fans a single query into 12 to 15 sub-queries behind the scenes, which is why covering several real sub-intents per page beats one keyword per page. The full framework split sits in SEO vs AEO vs GEO, and the routing into AI answers specifically is in how to get cited by Perplexity, ChatGPT and AI Overviews.
Can AI be used for SEO? Yes, for most of it. AI handles keyword research, SERP analysis, clustering, first drafts, internal linking and schema well. It does not reliably handle fact-checking or voice, so those two stay with a human. The working pattern is AI for the volume, human for the truth and the final edit.
Which AI is best for SEO? There is no single best model. The research stage needs a keyword-data source, the drafting stage needs a strong general-purpose language model, and the audit stage needs scripted checks rather than another model. The process around the tools decides results more than the choice of model.
Can ChatGPT do SEO? ChatGPT can draft content, suggest keywords and outline structure. It cannot verify its own claims, and it tends toward phrasing that reads as machine-generated. Use it for drafts and ideation inside a brief, then run a human fact and voice pass before publishing.
Is AI SEO content penalised by Google? Google penalises unhelpful content, not AI content as such. Mass-produced thin pages get filtered regardless of how they were written. A well-researched, fact-checked, genuinely useful page drafted with AI assistance is fine. The line is quality and intent, not the tool.
What does an AI SEO workflow actually look like? Seven stages in our practice, run as a chain: listening for trending intent across Reddit, X and Threads, keyword research and clustering, SERP intent classification, drafting to a brief, a human fact-check gate, a human voice gate, then internal links, schema and distribution. The AI leads five; a human gates two.
What is LLM SEO or generative engine optimisation? It is optimising to be cited inside AI answers rather than only ranked in blue links. It overlaps with classic SEO on page quality but adds source-level presence, since engines quote a small set of high-trust sources. We cover the distinction in SEO vs AEO vs GEO.
Does AI for SEO replace an SEO team? It replaces the slow middle of the work, not the judgment. Someone still has to decide which pages should exist, verify the claims, and own the voice. In practice it lets a small team produce at the volume a large one used to, which is what we run here.
AI for SEO is a division of labour, not a magic button. Give the machine the research and the wiring, keep a person on the truth and the voice, and the page ranks. Skip the two gates and you join the few hundred million pages that went nowhere.
If you would rather not stand this process up in-house, it is one of the modules in our GTM OS bundle. We run the whole loop for you, from keyword research through the two gates to distribution, built around your launch rather than handed over as a template.
Want a read on where your commercial queries actually stand inside AI answers? Run a 20-minute audit. We will pull Perplexity, ChatGPT search and Google AI Overviews for your top queries and show you which sources are getting cited instead of you, plus where AI-assisted content would move the needle and where source-level presence is the real gap.