AI SEO Services: Six Flex Points and Who Decides
AI SEO services run on judgement, not output volume. See the points in a real workflow where a human decision still changes the result.
By David Jubé · Jul 20, 2026 · 14 min read
AI SEO services are sold by the output they produce, and the output is the cheap half.
Run the entire workflow with AI and you will still make every decision no model makes for you. Which question the page answers. Which claim can carry a number. Which fact about your business is worth declaring in markup. Which draft is accurate and still not worth quoting.
Generation moves faster than any writer I have worked with, and it moves fastest in whatever direction it was pointed. That is the shape of the job now: the production half got cheap, the pointing did not, and what you buy from anyone running an assisted workflow is a set of six decisions rather than a volume of pages.
Key takeaways
- An AI-assisted workflow has six points where the tooling hands a decision back. Those six are the service; the rest is production.
- The tooling does genuine heavy lifting at all six, drafting, structuring and validating faster than a person working alone.
- What it will not do is choose the direction, and it moves fastest in whatever direction it was given. A wrong call produces a fluent page aimed at nothing.
- Three of the four conditions a page must meet are largely automatable. The fourth, another domain stating the same facts about you, is not reachable from inside your workflow.
- Vetting an AI-assisted provider comes down to who reads the output before it ships and on what cadence.
An assisted workflow has six flex points, and judgement lives at each one
The label itself is used loosely, and Graphite’s research on what AI SEO actually names as a body of work shows how far the competing terms overlap, describing one production line seen from different ends. Underneath the naming, the work is a sequence, and each step in it has a version the tooling runs and a version it hands back to a person.
Each of those six calls has a failure mode of its own, and the first one sets the cost of the other five, because everything downstream inherits the aim.
Intent comes first, because nothing decides for you who the page is for
Ask for a page about a subject and you get the subject, covered evenly. That is the tool working correctly, since even coverage is the safest reading of a vague brief. A query is narrower than a subject, and the page either meets it in the first two sentences or it does not.
Get the call wrong and the failure is invisible from the inside. The page is accurate, well organised and pointed at nobody, so nothing looks broken and no review that grades writing quality will catch it.
Get it right and everything downstream gets cheaper. The draft comes in shorter, the evidence it needs becomes obvious, and the review has a standard to measure against instead of a general impression of quality.
Evidence has to be supplied, since generation has no access to your numbers
A generated pass asserts what is generally true. It has no way to assert what is specifically true about you, because nobody told it, and the generally true version is what every competitor in your category has already published.
So the second decision is which claims carry something only you hold: a measured result, a dated observation, a constraint you hit, a figure out of your own records. Two or three per page is the working number. A page with no specifics repeats the general statement, and a page saturated with numbers reads like a spec sheet and buries the one figure worth quoting.
My rule is one evidence-carrying claim per section, with the rest of the section carrying the reasoning that makes the claim matter. That is what moves a passage from correct to citable, the mechanism How to Get Cited by ChatGPT, Perplexity and AI works through in detail.
Drafting is where the tooling earns its place, and the range genuinely expands
Everything above is a constraint on the tool. Drafting is where it changes what one person can do, and the change is not marginal. A single practitioner with a clear brief now covers ground that used to need a team: a full page set drafted, restructured, and rewritten in three registers to see which one lands.
The range is real and I would not go back. Pew’s survey of how teens use and view AI documents how ordinary these tools have already become in the cohort that took them up fastest.
The decision at this step is the brief itself. Aim it at a topic and it returns a long, fluent, unplaceable article. Aim it at a query, a reader and a stated position, and the first pass arrives close to shippable, which is the same speed working for you rather than against you.
Journey position 10, Website Content Writer: The Point You Hand It Over, covers this decision inside the content lane, where it lands as a question of when a second reader sees the draft.
The direction is the deliverable
Speed multiplies whatever direction you set, in both signs. A brief naming the reader and the question turns volume into coverage, and a brief naming a topic turns the same volume into pages nobody was looking for.
Structure declares what is true about you specifically, not about your category
Every build path emits structured data now, and the default block is valid, complete, and about your category rather than about you. A block saying you are a business in a sector is present without being informative, so validity is not the test that decides anything here.
The decision is which fields describe this business: the service names you sell, the area you cover, the identifiers connecting this site to the places you already appear elsewhere.
Every field you declare is also a promise to keep it true. A value that drifts out of step with the visible page becomes a contradiction, and one contradicted field costs more than an absent one, because it hands a reason to discount everything sitting beside it. Declare the fields you will maintain, and leave the rest of the vocabulary alone.
Review catches the accurate sentence that no engine has a reason to quote
A drafted page passes every check that grades writing. It is clear, ordered, free of errors, and it can still contain nothing worth lifting. That is the failure a package description never mentions, and the one that changes the finished page most.
The review question differs from the editing question. Not “is this sentence correct” but “would anything quote this sentence, and if not, what is it doing on the page”.
Nielsen Norman Group’s work on evaluating AI output makes the point about assessment itself: one output read once tells you almost nothing, because judging a single sample is not judging the process that made it. Reading your own draft once and finding it fine is the expected result rather than evidence about the draft.
So the review runs adversarially by design. Mark every sentence that could appear unchanged on a competitor’s site, then cut it or make it specific. What comes out is a shorter page with a reason to exist, and the sentences that survive the pass are the ones something can lift.
Publishing without observation is where unsupervised scale starts
Publishing does not finish the workflow, because the only evidence about whether any of it worked arrives afterwards.
Some of that evidence is precise. Server logs record which AI crawlers fetched which pages and how often, and Oncrawl’s log-file evidence of what AI crawlers do on a site shows how much sits in that record already. Which agents to expect is documented too, in OpenAI’s crawler documentation, so a log line can be attributed rather than guessed at.
The traffic side is coarser. Assistant referrals show up in analytics, Cloudflare Radar on AI crawler and referral trends tracks how those flows move in aggregate, and the referral market keeps splitting across more assistants, which StatCounter’s data on the AI chatbot referral market fragmenting records directly. A single source of assistant traffic is a thinner read every quarter.
Then there is the part that leaves no trace at all. SparkToro’s account of what AI visibility tracking can and cannot see is honest about the ceiling, since an answer a reader accepted without clicking is invisible to you and to any tool sold to measure it. How to Measure AI Traffic With No Referrer covers the workable proxies and labels the gap as a gap.
Timing changes what any of this means. Orbit Media on how long organic search work takes before it shows results is the corrective for anyone checking a week after publishing and concluding nothing happened.
The decision is what gets read back into the next round. Answer it and each round corrects the last. Skip it and you are producing at volume with nothing coming back, which is what unsupervised scale means and what eventually gets a site flagged. The tooling does not cause that outcome. Publishing without ever reading the result does.
Deciding which flex points you keep, and which you would rather hand over
Six decisions, and you can make all of them yourself. Capability is not the constraint here, and confidence is: which of the six do you call quickly, and which do you sit on for a week and then guess at anyway.
One thing no version of this workflow produces, whoever runs it. Corroboration lives on somebody else’s domain, and no pass over your own site creates a source outside it. Three of the four conditions a page has to meet are largely automatable now, and the fourth is a different kind of work, done by asking people rather than by producing pages.
If you bring somebody in, the vetting question is narrow. Moz on finding the right SEO provider sets out the general questions worth asking, and for an assisted workflow the sharpest one is who reads the output before it ships and on what cadence. A provider who names the review step and shows what it changed last month is running the workflow above. One who answers with page counts is selling the production half.
Journey position 18, Thirteen Questions to Ask a Web Vendor Before Signing, carries that line of questioning further, and it hands you things you can check rather than a portfolio you cannot.
The AEO category collects the reasoning behind these decisions, and the case notes show them applied, by vertical rather than by name. What I do in my services is those six decisions and the review cadence around them.
Book a free diagnosis
Want to know which of the four conditions your site is failing, which flex point is producing the gap, and which of those calls you should be making yourself? I read the site, the pages, and what your logs and analytics show, then tell you where the judgement is missing rather than where the output is thin.
Ending the human step reads well in a pitch and badly on inspection
A workflow with nobody in it can be advertised, and every step above has a version that runs unattended.
Each one then defaults to the safe general answer. The page covers the subject, the claims stay unspecific, the schema describes the category, the review passes, and nothing gets read back into the next round. Every individual output is defensible, and the site is invisible.
Nothing failed in that sequence. The tooling did exactly what it was pointed at, six times over, with nobody doing the pointing. That is the whole distance between a workflow that extends a practitioner and one that replaces the practitioner with an average.
Design and communication are addressed to people, so a workflow proud of having none has quietly answered a question nobody asked. Use the tools hard, point them well, and keep the six calls.
Frequently Asked Questions
What are AI SEO services?
AI SEO services are the named work items that get a site found and cited: intent research, evidence gathering, drafting, structured data, review, and observation after publishing. An AI-assisted provider runs each step with tooling doing the production. What you are buying is the decision inside each step, not the volume of pages that arrive.
Will AI replace SEO?
No, and the honest version is more interesting. Tooling has extended how much ground one practitioner can cover, which is a real change and a welcome one. The decisions it accelerates still need somebody who knows the business: which question the page answers, which claim carries evidence, which facts are true about you.
How do you know if your SEO agency is bad?
Ask who reads the output before it ships and how often, then ask to see what that review changed last month. A provider running an assisted workflow should name the review step and show you an edit it produced. Deliverable counts, page volumes and traffic promises answer a different question.
Can SEO be done by AI?
Largely, and quickly, for three of the four conditions a page has to meet. An assisted workflow produces crawlable pages, distinctive passages and valid structured data faster than a person working alone. It cannot produce corroboration, because corroboration means somebody else’s domain saying the same facts about you.
Can I do SEO myself?
Yes, and the tooling has made that far more realistic than it was. The real question is not whether you can do the tasks, it is which of the six decisions you are confident calling: the page’s question, the evidence, the direction of the draft, the structured fields, the cuts, the read-back.
How do you measure AI visibility?
Measure what leaves a trace: assistant referrals in analytics, AI crawler hits in server logs, and direct or branded traffic arriving after a citation. What stays unmeasurable is the answer a reader accepted without clicking. Treat that gap as a gap rather than assuming a tracker closes it.
Continue Reading:
Previously in this series
- AI Overviews and the Six Schema Types a Build Emits, the schema types a build emits by default, which is one of the six points where judgement comes back to you.
Next in this series
- Thirteen Questions to Ask a Web Vendor Before Signing, the questions that test the same judgement in a vendor, before you sign anything.
From the library
- Use AI to Write Without Sounding Like AI, spotting the tells that surface once the review flex point gets skipped.
- How to Structure a Page to Rank and Get Cited, what the structure flex point looks like when a person makes the call.
- Schema Markup for AI: Answer-First Writing Guide, the schema choices a vendor should explain, not leave to a default.
- How to Produce Content That Ranks and Gets Cited, seeing these six flex points as one brief, draft and proof system end to end.
Explore TDM Insights Categories