The Best Use of AI in SEO Has Nothing to Do With Writing
Most SEO and AEO teams point AI at the draft, the one step where it reliably produces commodity content. The real leverage is everything around it: business context, ICP psychographics, journey stages tailored to the business, briefs that carry that context, cluster-wide internal linking and CTAs, create/optimize/consolidate/prune decisions, and a feedback loop from Search Console that keeps the strategy alive. Notes by Mihir Naik.
Most SEO and AEO teams point AI at the draft, the one step where it reliably produces commodity content. The leverage is in the ten steps around it: business context, ICP psychographics, journey stages, briefs, link maps, portfolio decisions, and a feedback loop that keeps the whole thing honest.
Everyone pointed AI at the wrong 10%
I spend a lot of time on LinkedIn and in Slack channels with people doing this work, and I keep seeing the same setup. A team adopts AI, and within a month the entire investment has collapsed into one activity: generating articles. The calendar fills. Output triples. Six months later, the pages sit there getting no traffic, no citations in ChatGPT or Perplexity, and no pipeline, and someone concludes that AI content doesn't work.
I don't think that's the lesson. Writing the words was never the expensive part of a content program. Knowing what to write, for whom, at which moment, in what format, and why anyone should believe you over the incumbent, that was always the expensive part. Teams automated the cheap step and left the expensive ones exactly where they were.
Why AI content reads like everyone else's
Here's the mechanism, and it isn't a flaw in the model. A language model trained on the open web has one default behaviour when you give it a thin prompt: it returns the consensus. Ask for "a blog post about expense management software" and you get the average of ten thousand posts about expense management software. That average is, by definition, what every competitor is also going to get.
The output is generic because the input was generic. And notice what that implies: the fix isn't a better model or a cleverer prompt library. The fix is feeding it the things the internet's average cannot contain, your business, your customers, your data, your point of view. Which is exactly the work most teams skipped in order to get to the draft faster.
Commodity in, commodity out. The model isn't producing average content because it's a model. It's producing average content because average context is all you gave it.
So the question worth asking isn't "should we use AI to write?" It's "where in this workflow does AI compound, and where does it flatten?" My answer, after running this on real programs, looks roughly like this:
Where teams point AIWhere the leverage actually is
Drafting the articleUnderstanding the business and the market it competes in
Rewriting the introSculpting the ICP until it describes one real person
Spinning up meta descriptionsBuilding journey stages that fit this business
Expanding an outline into proseDeciding what format each stage actually needs
Filling a content calendarWriting briefs that carry the whole context
Producing more pagesMapping links and CTAs across a full cluster
Publishing fasterDeciding what to refresh, consolidate, and kill
Everything in that right-hand column is thinking work: slow, context-heavy, hard to scale with headcount, and almost never done thoroughly because there was never enough time. That's precisely the profile of work AI is good at accelerating. Let me walk the chain, because the steps compound in order.
Start with the business, not the keyword
The first artifact isn't a keyword list. It's a written understanding of the business: what it actually sells, how it makes money, which segment is profitable versus which is merely loud, what the sales team hears in objection calls, where the product genuinely beats the alternatives, and where it honestly doesn't.
Then the competitive landscape, and I mean the real one. Not "who ranks for our head term" but who the buyer considers, including the spreadsheet, the incumbent vendor, the agency, and doing nothing. AI is very good here, because this is a synthesis problem across messy inputs: call transcripts, review sites, community threads, competitor positioning, pricing pages, analyst notes. A human doing this properly takes two weeks. With AI reading and structuring the raw material, it takes an afternoon and covers more ground.
This document becomes the substrate for everything downstream. Skip it and every later step inherits the internet's average instead of this company's reality.
Sculpting an ICP, not filling in a template
Most ICP documents are firmographics with a stock photo: company size, industry, job title, maybe a "pain point" bullet list. That describes a market segment, not a person, and it produces content aimed at nobody.
The version that changes your content is psychographic. What are they actually trying to make progress on? What are they afraid of? What does a win look like to their boss, and what does failure cost them personally? Whose opinion do they need on their side internally? What words do they use for the problem, before they know your category exists?
I say sculpting deliberately, because this is iterative. You draft it, then push on it: argue with the model, feed it real customer quotes and lost-deal reasons, ask it where the profile is generic, and cut everything that would be equally true of a competitor's buyer. Take a spend-management SaaS selling to finance teams. "Finance leaders at mid-market companies who want to control spend" is a segment. This is a person:
- A controller at a 200 to 800 person company, two years into the role, one promotion away from VP Finance.
- The functional job: close the books faster and stop chasing receipts across four systems.
- The emotional job: never be the person who approved a tool that failed the next audit.
- Buys under a CFO who will ask, "why this one, and what does it replace?" in a single sentence.
- Calls the problem "expense chaos," never "spend management," until a vendor teaches them the term.
That last line alone changes your keyword research. The line above it changes your entire consideration-stage content, because it tells you the audience for that content isn't only the controller, it's the CFO they have to convince.
Your journey stages aren't TOFU, MOFU, BOFU
Awareness, consideration, decision is a diagram of a funnel, not a description of how anyone buys anything. Every business gets the same three buckets, which is another way of saying the framework carries no information about your business at all. Content mapped to generic stages produces generic content, for the same reason a thin prompt does.
Build the stages from the ICP instead. For that controller, the real sequence has a step no funnel diagram contains:
- Living with it. Expense chaos is annoying but survivable. Nobody is searching for a solution yet.
- A trigger makes it urgent. A failed audit, a funding round, a headcount jump, a controller inheriting the mess.
- Learning the category exists. They discover this is a named problem other people solved.
- Building the internal case. They need to defend a purchase to a CFO who wasn't in any of these conversations.
- De-risking the choice. Security review, audit trail, migration cost, what happens if it fails.
- Proving it worked. Ninety days in, they have to show the CFO the decision was right.
Stage 4 is where most B2B content programs have nothing at all, and it's the stage where deals actually stall. You only see it if you built the journey from a real person instead of importing a template. AI is genuinely useful for drafting and stress-testing these stages against the context document, as long as you keep rejecting the generic version it offers first.
Strategy falls out of the journey and the ICP
Once the stages are real, the content plan mostly writes itself, because each stage has an obvious job. Stage 2 content has to be findable at the moment of the trigger. Stage 3 has to name the category in the buyer's own words. Stage 4 has to be forwardable to someone who will read it without context.
And that's where format flexibility matters more than most teams allow. The stage should decide the artifact, not the publishing habit. Our stage 4, building the internal case, does not want a 2,000-word blog post. It wants a one-page business case the controller can paste into an email, a cost-of-current-state calculator with their own numbers in it, and an objection-by-objection FAQ written for the CFO, not for the controller.
Most content calendars are a list of articles because articles are what the team knows how to produce. When the journey picks the format, you stop publishing blog posts at problems that needed a calculator.
Briefs that carry the whole context
This is where all that upstream work pays off, and where AI does its heaviest lifting. A brief built on the context document, the ICP, and the stage is a fundamentally different object from a brief built on a keyword and a competitor outline. It states which person is reading, at which moment, what they believe walking in, what they must believe walking out, what the piece has to prove, what the business can uniquely claim, what format it should take, and what happens next.
The research underneath it automates well, and this is the honest place for AI to do volume work: pulling the statistics, the regulatory specifics, the benchmark data, the format conventions the topic expects, the objections that show up in review sites and community threads. The critical constraint is provenance. Research is only worth automating if it's scoped to sources you trust and grounded in material that belongs to this business:
- Your own product data, support tickets, and sales-call objections
- Named primary sources with dates, not a plausible-sounding number with no citation
- Customer language pulled verbatim, so the piece sounds like the market, not like a model
- The specific claim this business can make that a competitor cannot
A brief like that is also the thing that makes the eventual draft good, whoever or whatever writes it. Which is the quiet point of this whole essay: the quality was decided upstream, long before anyone opened a document.
The connective tissue: clusters, links, and CTAs
Here is a job that is genuinely tedious, genuinely mechanical, and almost never done well: internal linking across a complete cluster. Not "add three related links at the end," but reasoning over the entire planned content set at once, what links to what, with which anchor, in which direction, and how a reader moves from the trigger-stage piece to the category explainer to the internal business case.
A model holding the whole plan in context can propose that map in one pass. A human doing it by hand across sixty pages does it badly, once, and never revisits it. This matters more in AI search than it did in classic SEO, because retrieval works on passages and relationships: a well-connected cluster gives an engine a coherent structure of your expertise rather than sixty disconnected pages that each half-answer a question.
The same pass should assign CTAs, and the journey makes them coherent instead of uniform. Our stage 2 piece, read by someone who just failed an audit, should not say "book a demo." It should offer the audit-readiness checklist. The stage 4 piece offers the business case template. One primary action per page, chosen by where the reader is, not by what the demand-gen team wants this quarter.
Create, optimize, consolidate, prune
Most content strategies only have a create button. That's how sites end up with four thin posts about the same question, a 2021 guide that still ranks and still says something the product no longer does, and two hundred pages nobody has opened in a year.
A whole-portfolio view asks four questions of every existing URL, against the journey map rather than against a traffic column:
- Create where a stage has no coverage at all. Stage 4 is usually empty.
- Optimize where a page is aimed at the right moment but says the wrong thing, or is now factually stale.
- Consolidate where several thin pages compete for one intent, and one strong page would serve it and be far more retrievable.
- Prune where a page serves nobody in the journey and dilutes what your site appears to be about.
This is a classification problem over a large inventory with a defined rubric, which is to say it's the kind of work AI is excellent at and humans avoid. The judgment stays yours; the pass over four hundred URLs doesn't have to be.
A strategy that breathes
The last piece is the one that turns all of this from a document into a system. Feed the results back in. Google Search Console and Bing Webmaster Tools tell you which queries you actually surface for and which ones you thought you would. Analytics tells you which pieces move people forward and which ones end the session. AI visibility monitoring tells you whether engines cite you when someone asks your category's question out loud.
Read against the journey map, that data stops being a performance report and becomes a correction. When queries arrive that your stage definitions never anticipated, the stages were wrong, not the queries. When a stage 4 page gets impressions but no engagement, the format is wrong. When a competitor starts getting cited for your category explainer, your context document is out of date. Every one of those is a strategy edit, not a content edit.
A content strategy written once a year is a document. A content strategy that re-reads its own results every month is a system. AI is what makes the second one affordable.
Not one line of that was about generating content
Read back through the chain. Business context, ICP, journey stages, strategy, formats, briefs, research, link maps, CTAs, portfolio decisions, feedback loop. Nowhere in it did I argue about whether a human or a model should type the sentences. I don't think that's the interesting question, and I'm not here to tell you never to use AI in a draft.
What the chain is about is the only thing that ever made content work: what your users are looking for, what they're afraid of, what would genuinely solve their problem, how this business solves it, why anyone should choose you, and where you should improve based on what's working and what isn't. Do that thinking properly and the draft is close to a formality. Skip it and no amount of generation saves you, because you'll have automated your way to the average of everything anyone has ever said about your category.
If your team is going to invest in AI this year, point it here:
- The context only your business has, before anything gets written
- The judgment work that never had enough hours: journeys, briefs, link maps, portfolio calls
- The feedback loop that keeps the strategy alive instead of annual
The teams getting commodity results from AI aren't using it too much. They're using it in one place, and it happens to be the least valuable one. More field notes on how AI search is changing this work are on the notes index.
What’s next
More long-form methodology lives in guides; notes stay informal.
About the author
Mihir Naik, AI search (AEO) professional and product leader. Senior Product Manager (AI) at seoClarity, building Clarity ArcAI. Based in Toronto; in SEO since 2011. Available for consulting.
Read full bio →