AI will not fix bad SEO. It will scale it.
- Flex Your Idea
- 33 minutes ago
- 13 min read
The problem is not that AI does search marketing badly. It is that AI now makes bad search marketing look exactly like good search marketing, and that is a buyer's problem before it is a technology one.
Short version: AI multiplies whoever is holding it, and it does not supply the judgement. In search that judgement enters at two points, choosing what to work on and deciding whether the result is any good, and a model can do neither. Point it at someone who cannot tell the difference and you get work that is confident, well presented and wrong, produced faster than anyone can check it.
I get a version of this question most weeks now, and it is almost always from a business owner who has just been quoted a suspiciously low monthly number by somebody promising AI-powered SEO. Can I just use AI for this? Do I still need a person?
One thing you should know before I answer, because it is the reason my answer is worth anything. I use AI heavily, every working day, across research, drafting, technical analysis, code and data work. In over a decade in search, nothing else has changed my weekly output like it has. I am not telling you where these tools fail from the outside.
The answer is still no. Not because AI is bad at search marketing, because it is genuinely good at large parts of it, but because the parts it is good at were never the parts that decided whether you ranked.
Where the judgement actually enters
Any piece of SEO work is a loop with three seats. Someone decides what to work on. Something produces the work. Someone decides whether what came back is right.
AI sits in the middle seat and only the middle seat. It executes a brief it did not choose, and hands back an answer it has no way of grading. It cannot see your live search results, your margin, or which of your customers is worth having. From inside the model, a brief worth following and a brief that will quietly cost you the year arrive as the same thing: a reasonable-sounding instruction.
The other two seats are human. So the only thing in the loop capable of noticing that the work is wrong is the person operating it. Take a competent person out of those seats and you have not made the work cheaper. You have removed the error correction.
2 points in the process where human judgement decides the outcome
0 of them are visible in the deliverable you get sent
Bad SEO used to advertise itself. You could open a deliverable and see it: thin pages, keywords jammed into sentences no human would write, a strategy document that was really a keyword list with headings. Producing something that looked professional took at least partial competence, so a polished document was weak but real evidence of skill.
Fluency is free now. The audit, the content calendar, the technical recommendations, all of it presents at a professional standard whether the person behind it has years of judgement or a few weeks of practice with a chatbot. So the detection method every buyer has always used, look at the work and see whether it seems good, has stopped working.
A wrong brief executed beautifully is still wrong. It just no longer looks wrong.
Three places weak judgement shows up
1. They keep the whole keyword list
Ask any tool for keywords and it returns a long, tidy list sorted by search volume and difficulty, because those are the two numbers that exist in the data. What does not exist in the data is whether the person typing that phrase has a budget, a deadline, or any intention of buying anything.
Volume counts curiosity. Revenue needs intent and the ability to pay, and neither is visible to a keyword tool. There is a second problem underneath that one, and it is specific to working in a small market. Keyword tools bucket and round low numbers, so a term that genuinely gets searched here can report as zero, or as a rounded band that looks like nothing. It then gets filtered out before a human ever reads it. That is a measurement artefact, not an absence of demand, and the only way to know the difference is to have watched terms like that close business.
What that costs you: rankings climb, traffic climbs, the monthly report fills with green arrows, and nothing sells. A year in, the conversation becomes an argument about why the sales team has not noticed. I have written separately about choosing revenue over visibility in Google search, and the same trap sits underneath both channels.
2. They give a "do" query a "read" page
Searches fall into classes and the shape of the page has to match the class. Informational, where someone wants to understand something. Transactional, where they want to act. Navigational, where they are looking for a specific business by name. And local, where they want to visit somewhere physical.
Ask a model to make a page for a term and you will get an article, because an article is the default shape of generated text. If the searcher was trying to book or buy, an article loses to a page with a form on it. If the searcher wanted somewhere nearby, you are in a different competition altogether: the local pack, which Google ranks on relevance, distance and prominence, and which does not care about your blog post at all.
Someone who does not know the local pack is a separate surface will scope you half a job and report it as a whole one. AI will not raise it, because you did not ask.
3. Nothing on the page exists anywhere else
This is the structural ceiling, and the one nobody selling cheap AI SEO wants to discuss.
Google's quality raters are asked, in effect, whether a page contains original information, reporting, research or analysis: your own data, your own photographs, a named practitioner making a call and putting their name on it. Everything a language model produces is a rearrangement of material that has already been published, so a site built purely from model output has, by construction, nothing on it that was not already on the web. It can be accurate, readable, better organised than its sources, and still have no particular reason to outrank them.
No amount of prompting lifts that ceiling, because the missing ingredient is not in the model. It is in your business.
Here is the part practitioners get wrong in both directions. Google's spam policies define scaled content abuse as producing many pages mainly to manipulate rankings rather than to help people, and the wording is explicit that this applies "no matter how it's created". Their own examples include using generative AI tools to make many pages "without adding value for users". AI-written is not the offence. AI-written with nothing added is.
|
Someone who has not understood that distinction goes wrong one of two ways. They refuse to touch AI out of superstition and work far slower than they need to, or they publish at volume straight into the policy. Both will tell you they understand Google's position on AI content.
Why the cheap quote is cheap
Search work splits into two piles. Production, which is writing, formatting, building pages, generating schema and assembling reports. And decision, which is choosing what to fight for, judging whether the output is any good, and changing direction when it is not. Production used to absorb most of the budget. It now costs almost nothing.
So when a fee comes in at a fraction of what the work costs, the production has not been discounted. It has been automated, which is legitimate. What has usually gone with it is the decision time, because that is the only remaining place the hours can come from. You are not buying cheaper SEO. You are buying SEO with the judgement taken out, and judgement is the part you were never able to inspect in the first place.
I will not print a price range here, because a number without your market, your margin and your competitors attached to it is worse than no number. What I will give you is the shape of it. On a small account, the thinking that changes the outcome is measured in hours a month rather than minutes, and it is always the first thing a cheap quote removes. So the question to put to any quote is not "what do I get". It is "how many hours a month will a person spend thinking about my account, and what are they deciding in that time".
The Operator Test
If the deliverable no longer tells you whether someone is good, the conversation has to. I am not the first person to say that, and I should not pretend otherwise: the trade press has been arriving at the same conclusion through 2026, and anyone hiring in this field has felt it. What has been missing is the practical half. Knowing to interview rather than read the report does not help if you cannot tell a good answer from a smooth one.
A model can produce the document. It cannot sit in the room. Ask a question the document does not answer and whatever judgement genuinely exists is what replies.
Five questions, none of them tricks, and a good operator answers all five without going away to prepare. Each comes with two rubrics: what a real answer sounds like, and what a fluent wrong answer sounds like. The second matters more. Anyone who says something obviously stupid was never getting your money. The one to catch is the one who sounds excellent and is hollow underneath.
QUESTION ONE
"Show me a keyword you decided not to go after, and tell me why."
A list of what was deliberately rejected, with a commercial reason attached to each item, only exists if a person made decisions. Nothing in a keyword tool knows what is not worth having, so exclusions never appear unless somebody put them there.
Real answer: names specific terms, and the reason is about money. This one attracts people who never buy. This one is held by a competitor we cannot outspend. This one converts, but the leads are poor quality. Sounds right, still hollow: rejects terms on difficulty score or volume alone. That is the tool's opinion rather than theirs, and a difficulty score is not evidence about your business.
QUESTION TWO
"What does the top result have that we do not, and can we realistically get it?"
This forces two admissions at once: that they have looked at the live search result rather than a tool's export, and that they have an honest view of the gap. Open the search on your phone while they answer. If the page they are describing is not the page actually ranking, you have your answer in ten seconds.
Real answer: describes what a visitor gets from that page, separates the gaps you can close from the ones you cannot, and is willing to say that a term is not worth chasing. Sounds right, still hollow: a fluent inventory of word count, headings and keyword usage, with nothing on what a customer actually gets from that page or whether the gap is closable at all.
QUESTION THREE
"Which of my existing pages should not exist?"
Anyone who has never removed anything has never really audited anything. In the SME accounts I have worked on, consolidating weak overlapping pages has moved more than adding new ones, because it concentrates signals you already have instead of spreading them thinner. It also tests the instinct that matters most right now, which is restraint.
Real answer: names overlapping or thin pages, says where each one's traffic goes afterwards, and raises redirects without being prompted. Sounds right, still hollow: a decisive list of pages to delete with no plan for the traffic, the links or the old URLs. Confident pruning without a redirect plan does more damage than leaving it alone.
QUESTION FOUR
"What will be on this page that exists nowhere else on the internet?"
The originality question in plain language, and the one a model cannot answer on anybody's behalf, because the answer has to come out of your business. It is also the easiest of the five to grade, because you are not judging the answer. You are noticing whether they have asked you for anything.
Real answer: they ask you for material and know exactly which material. Your pricing logic, your process, your photographs, the objection every customer raises, the thing you refuse to do that your competitors do. Sounds right, still hollow: a polished promise of in-depth, expert, well-researched content. If nobody has asked you for anything, nothing original is going onto your site.
QUESTION FIVE
"If we rank and nothing sells, what would you change first?"
The most revealing of the five. It shows whether they think in rankings or in revenue, and whether they have a route to follow when the obvious thing fails. Anyone who has run accounts has been here more than once, so the answer arrives as a memory rather than a theory.
Real answer: fast and ordered. Check whether the page matches what the searcher wanted, then whether the page can convert at all, then whether that term was ever commercial to begin with. Sounds right, still hollow: an immediate move to conversion rate work as a separate paid project, without first asking whether the term was worth ranking for.
Then ask one follow-up to whichever answer impressed you most. That is where it holds or falls apart, because a rehearsed answer has a floor underneath it and real experience does not.
One caveat, since it follows from my own argument. This article is public, so the people it describes can read it and rehearse. Your protection is the follow-up question, and the fact that four of the five ask about your business specifically, which nobody can prepare for in advance.
What good actually looks like
Take a brief I see constantly. A Singapore business wants to rank for the main term in its category: biggest volume in the market, obvious commercial meaning, the phrase the managing director types into Google on a Sunday night.
Run through a model, that brief produces exactly what was asked for. A well-structured page targeting the term, supporting articles around it, internal links, schema, a content calendar. Every item correct. Every item on brief.
An operator does something else before any of that gets made. They run the search and read what is actually holding the first page: if it is directories and an incumbent brand, no better page displaces them, and that is a two-year fight rather than this quarter's work. They read the results themselves for what the searcher wanted, because a page of comparison articles and "best of" lists tells you the intent is comparing, not buying. Then they go to the client's closed deals and work backwards to the phrases those buyers actually used, which is the slowest part of the job and the part that decides everything. The brief changes. The head term becomes a long-term brand play and the effort moves to the terms attached to money.
Same tools, same models, and a completely different year, decided before a single word was generated.
That is the boundary. AI is excellent at work with a checkable answer and dangerous at work with a judgement answer, and search looks like the first kind while being mostly the second. Pulling entities out of a competitor set, writing schema, finding a pattern in a large export: all checkable, all faster with a model. Deciding what to fight for and what to abandon is not checkable, and it never will be.
Frequently asked questions
Can AI do SEO on its own?
It can produce almost every SEO deliverable on its own, and it cannot decide whether any of them were the right thing to make. Those are different jobs. The results of an AI-run search programme track the judgement of whoever is directing it, which is why identical tools produce excellent outcomes in one pair of hands and expensive nothing in another.
Will Google penalise my site for AI-generated content?
Not for the tool itself. Google's spam policies target scaled content abuse, meaning many pages made mainly to manipulate rankings with little value to a reader, and the policy states this applies no matter how the content was created. Publishing genuinely useful pages with a model's help is not the violation. Publishing volume with nothing original added is, whoever or whatever typed it.
Does any of this change now that people search inside AI assistants?
The surface is different, the logic is not. An assistant still needs a reason to name you rather than a competitor, and that reason is still original substance attached to a business someone can identify and verify. If anything the ceiling described above gets lower, because content that only rearranges what is already published is precisely what an assistant can produce for itself.
About the authorI am Christopher Han, and I run Flex Your Idea in Singapore.
I have spent over a decade in search marketing. As Search Lead at Omnicom Media Group Singapore, the group's regional headquarters for South East Asia, I led digital visibility strategy for Singapore Airlines and KrisFlyer, Changi Airport Group, IKEA, Chanel, Audi, Porsche, Resorts World Cruises, HSBC and MyCareersFuture. Those are brands I led search for in that role, not a client roster of my own company.
Enterprise budgets buy scale. They do not buy better thinking, and the discipline underneath is identical at every size. That is the work I do now with small and growing businesses here: the same rigour, without the enterprise price tag attached to it.
If you want the practical groundwork first, start with SEO optimisation for SMEs in Singapore, or the counterpart to this piece on the parts of AI that genuinely do earn their place, boosting your online presence with AI SEO strategies. If you would rather it were built in properly from the start, that is what search and AI optimisation means here. And if you simply want somebody to sanity-check a proposal sitting on your desk, get in touch. I will tell you straight, including when the honest answer is that you do not need me.
|
Find out where you actually stand first
Before you hire anyone, know your baseline. The Search Visibility Score checks how visible your business is in Google and across the AI assistants people now ask instead, then tells you plainly which parts are working and which are not. Singapore market, free, no call required.
Sources
Google Search Central, "Spam policies for Google web search". Read directly from the source page at time of publication.
Google Search Quality Rater Guidelines, on original information, reporting, research and analysis.
Google Business Profile Help, "Improve your local ranking on Google", on relevance, distance and prominence.
A note on limits, because it matters. The quality rater guidelines describe what Google's systems aim to reward. They are not the ranking algorithm, and nobody, including me, can promise you a position.
Comments