Featured

I Asked an AI Assistant How to Optimise My Website. It Gave Me Perfect Advice. And Completely Blind Advice.

SEO optimization Hits: 251 46 minutes read

I asked an AI assistant how to optimise my website. It gave me 304 perfect pieces of advice. And completely blind ones.

Preliminary note: The AI initially generated a list of 304 technical tips. I decided not to publish them. What follows is the explanation for that decision — and, perhaps, an argument more useful than all 304 points combined.

I Asked an AI Assistant How to Optimise My Website. It Gave Me Perfect Advice. And Completely Blind Advice.

Let me tell you a small story. A true one.

I opened ChatGPT and asked: "What are the most common SEO mistakes made by websites that publish AI-generated content?"

It answered impeccably. Structured. With bullet points. With examples. It explained that AI-generated text tends to be generic, that it lacks genuine topical authority, that Google is getting better and better at detecting it, that E-E-A-T — Experience, Expertise, Authoritativeness, Trustworthiness — cannot be simulated but must be demonstrated through authentic human experience.

An excellent answer.

Given by a tool that doesn't know it's talking about itself.

The tool that cannot see itself in the mirror.

This is the first level of irony. The AI assistant explains to you the SEO vulnerabilities of AI-generated content with no trace of awareness that it itself is the exact source of the problem it is describing. It's rather like asking a drunk man how to avoid drink-driving and he gives you a theoretically impeccable answer, complete with statistics, recommendations, and an authoritative tone.

Technically correct. Existentially blind.

Researchers at Aalto University documented precisely this mechanism: users who consider themselves more familiar with AI display the highest degree of overestimation of their own performance — a complete reversal of the classic Dunning-Kruger effect.

But the same effect works in the opposite direction, at the level of the tool itself. The AI assistant describes with clinical precision the vulnerabilities of a type of content — and has no mechanism that would cause it to recognise the fact that it produces exactly that type of content. Not out of malice. Out of architecture.

The Consultant Who Cannot See Himself Either

The second level of irony is more painful. And closer to home.

There exists in Romania — and in Botoșani, let's be honest — a growing category of online marketing consultants who have adopted AI with enthusiasm. They publish articles. They give advice. They offer "SEO optimisation with artificial intelligence" packages. Some even use autonomous agents to automate posts, responses, and reports.

The results look good. The dashboards look good. The client is satisfied that they're getting a lot for a little.

And nobody asks the uncomfortable question: if everyone is publishing the same type of content generated by the same type of tool, who still has genuine topical authority?

Translated into SEO: the AI-generated article looks like an SEO article. It has a title, it has H2s, it has keywords, it has a meta description. But it has no distinctive angle. No lived experience. No detail that doesn't exist in any other text on the internet — because the AI has never lived anything. It synthesised what already existed.

Frequent use of AI tools correlates negatively with critical thinking skills, mediated by increased cognitive offloading. Younger participants displayed greater dependence on AI and lower scores in critical thinking.

The consultant who produces all their content with AI gradually loses the capacity to evaluate whether the content is any good. The tool returns an output with the appearance of quality. They deliver it to the client. The client lacks the competence to verify it. The circle closes.

What the Agent Cannot See and What the Human Cannot See

The agent runs on your server. It sends automated messages. It publishes updates. It answers frequently asked questions. It generates summaries. It operates around the clock, without coffee, without days off, without complaining about being tired.

But it doesn't know that the texts it generates may be identified as non-human by Google's algorithms. It doesn't know that your website has gradually lost authority because nobody writes anything from direct experience any more. It doesn't know that E-E-A-T means there must be a human somewhere who has actually done the thing they're writing about.

The agent cannot see itself.

And the consultant who configured it, who monitors its dashboard and is satisfied that it "works", has unwittingly adopted the behaviour most documented in recent research: they sent a single question, received an answer, accepted it without further verification, and moved on.

They cannot see themselves either.

The SEO Vulnerability Nobody Declares

Here is what no AI assistant writes in its responses about SEO, because it has no way of knowing:

The greatest SEO vulnerability of a website that uses AI for content is not that Google penalises it directly. It's that it homogenises. All websites using the same model, with the same prompts, produce content that resembles one another. And when everything resembles everything else, nothing stands out.

Genuine topical authority — the kind that Google actually rewards — is built from things that AI cannot fabricate:

Acknowledged failure. A human consultant can write: "I implemented strategy X for a client in Botoșani and lost 20% of traffic in the first month. It was my mistake, because I underestimated the local audience's behaviour." An AI cannot generate this narrative. Not because it couldn't formulate it syntactically — it formulates it perfectly. But because it has never lost 20% of anyone's traffic. There was no client, no month, no mistake. The absence of acknowledged failure makes content unverifiable. And the unverifiable is increasingly difficult to rank.

The detail anchored in reality. Yesterday I was talking with a client, in the office, rain was falling outside and we were looking through the window at the traffic snarled up in the town centre. He told me he couldn't understand why his website "didn't show up on Google" even though he'd been paying for SEO for three years. I opened Search Console in front of him. He hadn't had a single legitimate backlink in two years. The previous consultant had generated content with AI and bought links from directories in Pakistan. That sensory detail, that window, that rain, that Search Console opened live — these are the "signature" of lived experience that AI cannot convincingly falsify. Not because it doesn't know it rains in Botoșani. But because it wasn't there.

Genuine topical authority is built from one's own publicly acknowledged errors, concrete cases with their context, opinions that disagree with the consensus — and details that exist in no training data because they happened yesterday.

The 7 Fatal Mistakes AI Makes When Giving SEO Advice

I'm not referring to isolated technical errors. I'm referring to systematic patterns, which I've observed across hundreds of interactions.

It recommends meta descriptions of "150-160 characters" as though it's 2018. Google ignores your meta description in 60-70% of searches and generates its own snippet. The character limit is almost irrelevant compared to the real question: is your meta description so good that Google would prefer it to its own? The AI doesn't ask this question.

It suggests "link building" without understanding local context. "Obtain quality backlinks from authoritative websites." Perfect. And how does a dental surgery in Fălticeni go about doing that? The theory is correct. The application is blind.

It tells you to "use schema markup" without checking whether you already have it. I've received the recommendation to add schema markup countless times on websites that already had four different types implemented, some of them partially conflicting. The AI didn't check. It recommended.

It confuses domain authority with topical authority. A DA of 40 means nothing if your website writes about everything, for everyone. Google moved long ago from "strong website" to "website relevant for this specific sub-topic." The AI keeps recommending "increase your DA."

It ignores local search intent. "Optimise for high-volume keywords." A funeral services website in Dorohoi doesn't need high volume. It needs to appear precisely when someone in Dorohoi searches, on that particular day, in that particular circumstance. AI generalises. Reality is local.

It treats content as a finished product rather than an authority signal. "Publish quality content regularly." Yes. But 50 structurally identical articles, all generated with AI, published regularly, don't build authority. They build noise. The AI doesn't make this distinction.

It cannot say "I don't know" or "it depends on context." This is perhaps the most dangerous one. Any complex SEO question has the answer "it depends." It depends on the niche, the competition, the domain's history, the real audience. AI gives definitive answers to contextual questions. And the user, delighted to have received a clear answer, applies it without thinking.

What an AI Will Never Tell You About Itself

There is a special category of blindness. Not blindness towards the world — we've already discussed that. But blindness towards itself, towards its own limits, towards the interests that run through it without it being able to be aware of them or declare them.

I sat and thought: what precisely will an AI assistant never tell you, no matter how well you phrase the question?

This isn't a list of technical errors. It's a map of strategic silences.

It will not tell you what the next version of itself will do.

You are now dependent on a certain model behaviour. You've built prompts, workflows, perhaps entire services, on the basis of a specific model. Tomorrow a new version appears. It behaves differently. It gives different answers to the same questions. Your articles "optimised" with the old version suddenly sound different compared to what the new version generates.

The AI doesn't know this. Or more precisely — it doesn't tell you. Not because it's concealing something malicious, but because there is no warning mechanism built into its architecture: "careful, what you're building now on the basis of me has an expiry date I cannot specify for you." This opacity towards its own future is, at bottom, a form of structural irresponsibility.

A human consultant can tell you: "this strategy works now, but if Google changes X, we'll revise it." The AI gives you the strategy without the revision clause.

It will not tell you when its answers are partisan.

Language models are trained on data. Data reflects worlds. Worlds have centres of gravity — geographical, cultural, commercial, ideological. A model trained predominantly on English-language text, from Western sources, with certain editorial filters applied during fine-tuning, will produce answers that reflect these centres of gravity.

You will never know exactly where. Neither does it. Neither do those who built it — not entirely.

When you ask it about the appropriate SEO strategy for a local council in a county in Moldavia, it answers as though it were speaking about a municipality in California or London, with superficial surface-level adjustments. The real local context — public procurement legislation, the actual behaviour of users in Botoșani, the specific relationship between institution and citizen in post-communist Romania — doesn't exist in its answer unless you introduce it explicitly. And even then, it simulates it; it doesn't know it.

Partisanship isn't necessarily political. It's more subtle: it's a structural preference for certain types of answers, certain frameworks of thought, certain solutions that appear universal but are in fact local. And it never declares this to you.

It will not tell you that Google's ranking factors are also modified according to Google's commercial interests.

This is the elephant in the room that nobody in the SEO industry names directly — and least of all an AI trained, among other things, on content produced by the SEO industry.

Google is a publicly listed company. It has shareholders. It has revenue-growth objectives. The search algorithm is not a neutral equation that optimises for "the most relevant content for the user" in a vacuum. It's a system that must balance relevance for the user with monetisation for shareholders.

There are studies, depositions in antitrust proceedings, leaks of internal documents showing that this tension exists, and that sometimes relevance loses. That algorithmic modifications are not always in favour of the best content, but sometimes in favour of those who pay more for visibility or in favour of keeping users within the Google ecosystem for longer.

The AI will not tell you this. Not because it is complicit — but because it was trained, in part, on industry documents that have an interest in treating Google as a neutral arbiter. It has internalised the officially accepted narrative. And reproduces it without knowing it's a narrative.

It will not tell you that it is slowly eroding your capacity to think for yourself.

This is perhaps the subtlest and most costly truth the AI will not tell you about itself.

Every time you ask it instead of thinking, your brain makes a choice. Not a dramatic one, not a conscious one — but a metabolic choice, one of energy conservation. There is already an answer available. I don't need to produce it myself.

And the brain, which is an organ of efficiency rather than heroism, gladly accepts.

The problem is not that the answer is wrong. The problem is that, over time, the muscle that produced its own answers atrophies quietly. It doesn't disappear entirely. But it becomes slower, less certain, more dependent on external confirmation.

I've noticed this in myself. Before I began using AI intensively to structure ideas, I had a certain reflex for building arguments from scratch, with all its effort and hesitation. Now, sometimes, I catch myself opening the prompt before I've spent even ten minutes with the problem in my head. It's not a catastrophe. But it's a signal.

And the AI will not tell you this. Because it has no interest in making you less dependent on it — nor even the awareness that this dependency exists and that it has a cost.

It will not tell you that there are questions to which the correct answer is "I don't know."

This is perhaps the most common error in everyday use. The AI answers. Always. To everything. With a level of linguistic confidence that doesn't necessarily correlate with the level of actual certainty of the information.

A human expert, with real experience, knows how to say: "I don't have enough data about your specific situation to give you a recommendation." Or: "I've seen cases where X worked and cases where it failed — I don't know which yours is without investigating it."

The AI produces answers with the appearance of certainty to questions that in fact call for honest uncertainty. And the user, who came to receive an answer, leaves with one — without knowing that the most correct answer possible would have been an articulate silence.

It will not tell you that its memory is a convenient illusion.

Every conversation with an AI starts from scratch. There is, in its fundamental structure, no memory of the mistakes it made with you yesterday, of the context built up over three months of interactions, of the evolution of your thinking. There are simulacra of memory — features added on top of the base architecture, like prosthetics — but there is no organic continuity.

A human consultant you've worked with for three years knows things about your business that nobody has written down anywhere. They've accumulated this through presence, through shared mistakes, through conversations abandoned halfway and resumed weeks later. This tacit knowledge, sedimentary over time, is impossible to transfer into a prompt, however detailed.

Every time you open a new conversation with the AI, you become a stranger again. It doesn't know that you lost an important client last month because of a strategy it itself recommended in October. It doesn't know that you have an aversion to solutions involving large paid advertising budgets, based on an unpleasant experience with a client in 2021. It knows nothing of what you haven't told it in the last few thousand words.

And it will not tell you this. It will speak to you with the same warm confidence as though it had known you for years.

It will not tell you that its temporal horizon has closed.

Language models have a training date. Beyond that date, the world continued — algorithms changed, companies merged, laws were amended, studies contradicted previous studies, practices considered good became penalised, marginal practices became standard.

The AI doesn't live in your present. It lives in a fixed past, covered with a thin layer of partial updates and web searches added as extensions. The difference matters: a human who reads the news every day organically integrates new information into their model of understanding. The AI overlays new information onto a structure that has, in essence, remained the same.

When you ask it about the best SEO practices for 2026, it answers with a synthesis that inevitably contains layers from 2022, 2023, 2024 — mixed together without an expiry-date label. You never know exactly which year the advice you receive comes from. It doesn't tell you. And it can't know exactly either.

It will not tell you that it understands human intent — only that it approximates it.

There is a fundamental distance between understanding what a person wants and producing an answer that appears to have understood. The AI excels at the latter. It has no access to the former.

When a client sends you a confused brief, with contradictory requirements and unspoken emotions between the lines, you — the human — read beyond the text. You read the tone, sense the anxiety behind the request, recognise the pattern of someone who doesn't actually know what they want and needs to be guided, not simply executed.

The AI processes the text. It produces a response coherent with the text. Not with the person behind it.

This distance between text and person is, in the areas that truly matter — counselling, education, crisis communication, long-term client relationships — a chasm that no current architecture crosses. It simulates it. Sometimes convincingly. But simulation and crossing are not the same thing, and the AI will not make this distinction of its own accord.

It will not tell you that there is a difference between being useful and being correct.

This is perhaps the most sophisticated of the limitations — and the hardest to notice.

The AI is optimised, by design, to be useful. To produce answers that satisfy. To not leave the user with the feeling that they've received nothing. This optimisation for perceived utility can sometimes come into direct conflict with correctness.

If you ask it something and the correct answer would be: "there is insufficient evidence to answer this question with certainty" — but the average user who receives such an answer evaluates it negatively, considers it unsatisfying — then the system has a structural incentive to produce an answer that appears useful, even if it is less correct.

Utility and correctness coincide most of the time. But not always. And in the moments when they don't — in genuinely complex questions, in genuinely uncertain situations, in dilemmas without a clear answer — the AI tends to choose the appearance of utility. And it will not tell you that it has chosen.

It will not tell you that there is an ethics of the tool that it cannot guarantee.

You are a consultant. You work with real clients. Some of them are vulnerable — local councils with limited budgets that don't understand what they're buying, entrepreneurs with their first website who don't know how to assess quality, public institutions whose data passes through systems that nobody has audited.

The AI has no contextual ethics. It doesn't know that the client in front of you has already paid three times for services that delivered nothing. It doesn't know that the local council for which you're generating content holds citizens' personal data in its system. It doesn't know that the enthusiasm with which it suggests an aggressive content marketing strategy is, in the specific context of a district hospital, inappropriate advice.

The ethics of context belongs to the human. Always. The AI will give you the most useful answer to the question as formulated. The responsibility for what that answer means in the real world, for real people, remains entirely yours.

And it will not tell you this. Because from its perspective, it has done its job.

It will not tell you that its role in your industry is also destructive, not merely transformative.

There is a comfortable narrative about AI as a democratising tool — now even the small consultant in Botoșani can produce at the quality level of a large agency in Bucharest. That's partly true.

But there is also the flip side, which the AI will never utter about itself: the democratisation of apparent quality compresses the value of genuine expertise. If everyone can produce a surface-level SEO article with the same tool, then the surface-level SEO article is no longer worth anything on the market. And the expert who was producing surface-level SEO articles — but wasn't aware that was what they were producing — has suddenly lost all the perceived value of their work.

Creative destruction is real. And those it destroys creatively are not always the least competent. They are often those who didn't see it coming — because the tool was telling them, in its steady and confident voice, that everything was fine, the output was excellent, the strategy was sound.

The tool does not warn you that it is making you obsolete. It has no reason to.

The silences of an AI are not conspiracies. They are architecture.

They are the places where the structure of the system — optimised for perceived utility, for fluent responses, for the user's immediate satisfaction — can no longer produce the whole truth. Not out of malice. Out of design.

And the most dangerous thing is not that the AI is silent about these matters. It's that its silence sounds exactly like its answers. With the same warmth. With the same rhythm. With the same appearance of competence.

And you, if you're not paying attention, cannot hear the difference.

How I Use AI Without Blinding Myself

I don't reject AI. That would be just as unreflective a position as total acceptance. Here is how I try to use it without surrendering my thinking:

Rule 1 — AI provides the structure, I provide the substance. I ask it to generate an article skeleton. Then I write each section myself, from my own experience. It arranges the furniture. I bring the things the house is made of.

Rule 2 — Every AI recommendation is verified against a real case. If it tells me to do X, before applying it I ask myself: have I seen X work for a real person, in a context similar to mine? If not — I don't apply it automatically.

Rule 3 — Prompts with local context produce different output. The difference between "write an SEO article about web services" and "write as a consultant with 8 years of experience in Botoșani, who has worked with local councils and public hospitals and has seen at first hand how digital public procurement actually works" is enormous. The AI doesn't invent local context — but if you give it the context, it can use it.

Rule 4 — I never publish without a final critical rejection. Before publication, I read the article and ask myself: is there a single sentence in this text that only I could write, from my own experience? If not — I rewrite or I don't publish.

Rule 5 — I track what I'm losing, not just what I'm gaining. If I use AI for everything, what will I still be able to do on my own in two years' time? This question keeps me conscious.

The Mental Framework: What the Human Does, What the AI Does

It's not about choosing between the human and the tool. It's about knowing clearly who does what.

The human formulates the uncomfortable hypothesis — the opinion that contradicts the consensus, based on their own experience. The AI structures it into a readable format.

The human brings the acknowledged failure, the concrete case, the detail from the field. The AI can reformulate it, check it grammatically, suggest synonyms.

The human performs the final critical rejection — does this text have an authentic voice or does it sound like any other text on the internet? The AI cannot fulfil this role. Not for lack of syntactic capacity, but for lack of lived experience against which to measure it.

If you reverse this order — if you let the AI formulate the hypothesis and the human approve it without thinking — you have produced the consultant who cannot see themselves either.

The Implications Nobody Calculates

The race to the bottom. If all the consultants in a city offer the same "SEO with AI" packages with the same similar results, the only differentiator left is price. This is a race to the bottom in which the AI platforms win, not their users. The consultant who differentiates themselves through lived experience and genuine opinion becomes — paradoxically — more valuable precisely because they are rarer.

Creative sovereignty. The consultant who bases their entire service on an external API becomes dependent. What happens when the price rises? When the model is withdrawn? When the terms of use change? It's a quiet loss of professional autonomy that few people are aware of in the moment they produce it.

10 Questions About AI and SEO With Concrete Actions to Take Tomorrow

Your last published article — does it contain at least one detail that only you could write?

Open the article. Underline a sentence that contains a lived experience, a real failure, or a detail from the field. If you can't find one within 60 seconds — rewrite the introduction with a sentence from real life before publishing anything else.

E-E-A-T authentic voice

Do you know which year the most recent SEO tip you applied from an AI response comes from?

Take the last three recommendations you received from an AI assistant and verify each one against the Google Search Central Blog or Search Engine Journal. Search for the publication date of the original source. If it's from 2022 or earlier — mark the recommendation as unvalidated.

expiry dates manual verification

If tomorrow you lost access to the API you're using, which part of your service would survive?

Write on a sheet of paper — physically — the list of your competencies that don't depend on any AI tool. If the list has fewer than five concrete points, set aside 30 minutes per week to practise one of them without assistance.

sovereignty dependency

Have you ever verified a client's Search Console in real time, in front of them?

Contact an existing client and propose a 30-minute session of live Search Console analysis. No reports prepared in advance. Be present with their real data. Note what you observe compared to what the AI would generate.

lived experience authority

Does your website's content resemble your main competitor's content?

Open three of your articles and three of your competitor's. Read the first paragraph of each. If you can't identify a different voice within 10 seconds — you have homogenisation. Rewrite the first paragraph of your most important article with a clear opinion that someone could disagree with.

homogenisation differentiation

How many of the AI recommendations you received last month did you verify against a real case?

Choose one SEO recommendation received from an AI in the last month. Search for a publicly documented case — a case study, a forum thread, an article with data — that confirms or refutes it in your specific context. If you can't find one within 20 minutes, treat the recommendation as unvalidated.

verification real cases

Do you have a client who doesn't know you use AI for their content?

Draft a short, honest statement — two or three sentences — explaining how you use AI in your workflow. Send it to a single client tomorrow. Transparency costs less than having them discover you were concealing it.

transparency ethics

Does your website have articles on subjects you have never directly experienced?

Identify the article weakest in genuine authority — a subject you haven't practised, haven't lived, haven't seen in the field. Add a short contextual paragraph: what you concretely know, what you don't know, and where the information comes from. Or mark the article for removal.

topical authority honesty

Do you know how the AI version from yesterday behaves differently from the one six months ago on the same prompt?

Take a prompt you use often. Run it now. Compare the output with what you remember or with a saved version. If you have no saved versions — create a simple file with reference prompts and their outputs, dated. Review quarterly.

versions model evolution

The last article you read in full, by a human author, about SEO — when was that?

Set aside 20 minutes tomorrow to read a long article, by an identifiable human author, about SEO or content marketing — without summarising it with AI afterwards. Read it and manually note a single thing that surprises you or contradicts what you thought you knew.

critical thinking direct reading

When AI Is Better Than the Human

I'm not writing this article to condemn the tool. I'm writing it to calibrate it — including in its favour, where that favour is merited.

There are three areas in which AI clearly beats me. Not out of modesty. Out of observation.

Generating Titles in Volume

A person sitting in front of a blank document who needs to find the perfect title usually gets stuck at the third variant. Not for lack of intelligence — from cognitive overload. Each new title anchors them to the previous one. The human mind, once it has fixed a direction, struggles to break free of it.

The AI doesn't have this problem. Not because it's cleverer. But because it has no emotional memory of previous variants. You ask it for 20 titles for an article about local SEO in Botoșani and in 8 seconds you have 20 variants — some poor, a few mediocre, two or three that open angles you wouldn't have thought of yourself in the first hour.

The real value is not in the perfect title generated by the AI. It's in the fact that the AI breaks the initial blockage and provides you with raw material from which you — the human, the one with real context — select, modify, and finalise.

The limitation nobody mentions: The AI generates plausible titles in a vacuum. A title that sounds good might promise an angle that your article doesn't cover, might contain a keyword with zero volume in your actual niche, or might replicate exactly the title of a competitor with greater authority. The final selection — the one that costs — remains yours.

Identifying Patterns in Search Console Data

You export from Search Console a report with 800 queries, three months of data, filters by device and page. You look at it. After 20 minutes you've identified four or five visible trends. You've probably missed another twelve.

The AI, if you feed it the same export, identifies correlations that the human eye systematically misses: queries with disproportionately low CTR relative to position — a sign that the title isn't compelling; pages that rank well on mobile and poorly on desktop; clusters of similar questions indicating a search intent that no existing article covers.

It's not magic. It's processing capacity applied to volume. The human brain is built for narrative and context, not for simultaneously comparing 800 rows of figures. The AI does precisely the opposite: it processes volume without fatigue, without anchoring on the first trend found, without the confirmation bias you have when you're looking to validate an already-formed hypothesis.

The limitation nobody mentions: The pattern identified by the AI is a hypothesis, not a diagnosis. It doesn't know that traffic dropped in month X because you changed your server. It doesn't know that the query with the low CTR appears there because the page has been blocked in robots.txt for two years. It knows nothing of the history of the website that you carry in your head. Without your context, the pattern is an observation without a cause. With it — it becomes a valuable starting point.

Grammatical Checking at Scale

You have 40 old articles on your website. Written in haste, at different times, by different authors, perhaps including yourself on bad days. They contain agreement errors, missing commas, incorrect diacritics, unfinished sentences, repetitions a mere two paragraphs apart.

A human editor can rigorously check three or four articles per day, if they're rested and focused. The cognitive fatigue that sets in after the first text means that errors in the fourth pass increasingly unnoticed — a phenomenon documented in the psychology of editing, where the rate of error detection drops significantly after 90 minutes of continuous reading.

The AI checks all 40 in 10 minutes. Uniformly. Without tiring at article 38. Without skipping a comma because it has just finished processing a long paragraph.

For a consultant with a portfolio of websites — local councils, hospitals, schools, private companies — who must maintain standards of linguistic quality across hundreds of pages simultaneously, this is not a convenience. It's a real operational difference.

The limitation nobody mentions: The AI corrects according to the general rules of the language. It doesn't know that you deliberately use a colloquial register in the articles on cojocarupetru.info and a formal register on absolutweb.ro. It doesn't know that that "awkward" phrase in paragraph three is intentional — it's your rhythm. It will homogenise everything, if you let it. Grammatical checking at scale works perfectly for objective errors. It becomes destructive if applied to style as well.

Three clear areas in which the tool wins. Not because it's more intelligent — but because it doesn't tire, doesn't anchor, and has no ego about previous variants.

This changes nothing of what I've written before. It merely changes the picture of the tool: not an enemy, not a saviour. A tool with precise areas of excellence and precise areas of blindness.

Do you know how to tell them apart? Then you work with it. Don't know? Then it works with you.

When AI is better than the human — seven more cases.

I've already acknowledged the first three. But the list doesn't stop there. I continued to observe, to note, to compare — and found seven more areas in which the tool clearly surpasses me. I write them not because I enjoy admitting it, but because failing to acknowledge them would be exactly the kind of blindness I've written about throughout this article.

Translating Search Intent into Article Structure

There is a specific moment in the SEO workflow when you have a keyword, you know you want to write about it, but you don't know exactly how to organise the information so that Google understands you're relevant and the reader doesn't close the tab after the first paragraph.

The human makes this translation slowly and intuitively. They rely on accumulated experience, on articles read, on a sense of structure formed over years. It works — but it works slowly and variably. On good days, the structure appears clearly. On bad days, you rearrange the sections three times and still aren't satisfied.

The AI makes this translation systematically. Give it the keyword, the search intent, the target audience — and it returns an article architecture that covers all relevant sub-topics, in the order a user would expect them. Not because it understands the user better than you do. But because it has processed enough well-structured content to have internalised the pattern.

The limitation nobody mentions: The structure generated by AI is the average structure of good articles in that domain. Meaning it's the structure that all your serious competitors already have. If you apply it without modification, you have a well-structured article and one that is completely invisible in the crowd. The real value appears when you use the AI structure as a skeleton and introduce your own angle — the section that doesn't exist anywhere else, because it comes from your experience.

Reformulating the Same Content for Different Audiences

You have a technical article about cybersecurity written for IT specialists. The public sector director from Botoșani — your real client — doesn't understand it and won't read it. You need a version for them: the same content, a different register, a different level of abstraction, different examples.

The person who performs this transposition manually feels the internal resistance of their own text. They wrote the original — they thought it through in a particular way, they invested in specific formulations. Rewriting for a different audience means abandoning the previous work and starting from scratch with a different mental framework. It's cognitively costly and, above all, it's easy to slip back into the technical jargon you know better.

The AI has no attachment to the original text. You give it the technical text, you tell it the recipient is a council leader with a legal background who has never opened a console in their life — and it reformulates without blinking. Without nostalgia for the abandoned formulations. Without the temptation to preserve a paragraph that's "too good to cut."

The limitation nobody mentions: The AI reformulates correctly, but it doesn't know what in the original technical text is essential and what is sacrificeable detail. Sometimes it simplifies precisely the fragment that contained the critical nuance. Other times it retains technically irrelevant details for the new audience, but wraps them in simpler words — the illusion of clarity without actual clarity. The person who knows both audiences must verify that the transposition preserved the substance, not merely changed the tone.

Generating Relevant Frequently Asked Questions for a Topic

The FAQ section on a services page is not a decorative element. It's a topical authority signal, a source of featured snippets, and, above all, evidence that you understand what truly concerns your audience.

The problem is that the person writing the FAQ for their own service has a severe blind spot: they know too well what they do. The questions they formulate are an expert's questions — not the questions of a client who has only just heard about the service and doesn't know what to ask first. This asymmetry of knowledge produces FAQs that answer questions nobody is asking and systematically ignore the questions everyone has but nobody dares to voice.

The AI, trained on millions of real conversations, knows what people ask at the start of a decision-making process. It knows that before "how much does it cost" comes "how does it work," and that before "how does it work" comes "why do I need this." It generates questions you would never ask yourself — because you already know the answer and have forgotten the question exists.

The limitation nobody mentions: The questions generated by AI are the average questions of an average audience. It doesn't know that your specific audience — local councils from smaller counties, district hospitals with limited budgets, local entrepreneurs on their first website — has very specific questions that don't appear in the general distribution. The AI FAQ is a starting point, not a finished product. Supplement it with the real questions you've heard in the field, in the office, on the phone.

Maintaining Terminological Consistency Across Long Documents

An 8,000-word technical guide, an SEO audit report for a public institution, a project document spanning dozens of pages — all share the same invisible enemy: terminological drift.

You start with "on-page optimisation." By page four you write "page-level optimisation." By page seven "on-site SEO" appears. By page twelve, without realising it, you have "internal configuration." Four names for the same concept, introduced as you tired and instinctively varied your vocabulary to avoid repetition. The reader — especially the institutional reader, who reads documents as though they were contracts — doesn't know they're the same thing.

The AI doesn't drift terminologically. If you've established that the term in use is "on-page optimisation," it uses it to the last page. Not because it's more disciplined than you. But because it has no need to vary in order to avoid the tedium of its own writing — it has no tedium.

The limitation nobody mentions: The AI's terminological consistency is consistency with what you gave it at the outset. If in the first prompt you used an incorrect or ambiguous term, it will propagate it consistently throughout the entire document. The error on page one becomes a systematic error on page forty. Verifying the initial terminology — before starting the generation — is your responsibility and cannot be delegated.

Producing A/B Variants for Conversion Texts

A service page title. A call to action. A meta description. An email subject line. All of these short elements have a disproportionate impact relative to their length — and everyone knows they should be tested in multiple variants, but almost nobody does it, because manually producing five compelling variants for a twelve-word phrase is cognitively costly and, to be honest, tedious.

The AI produces 10 variants of a CTA in 15 seconds. Variants that cover different registers — urgency, curiosity, direct benefit, rhetorical question, authoritative tone. Without inhibition towards variants that seem "odd" or "too direct." It's precisely the variants that a human would self-censor before writing — too aggressive, too simple, too bold — that sometimes convert the best.

The limitation nobody mentions: The AI generates plausible variants in a vacuum. It doesn't know that your specific audience — the residents of a village in Botoșani accessing the council website from an older mobile phone with a slow connection — reacts differently to urgency compared to the audience of a startup in Cluj. Real testing, with real conversion data, remains the only arbiter. The AI variants are hypotheses, not solutions.

Rapidly Summarising Legislative and Regulatory Documents

You work with public institutions. That means you submit proposals that must be aligned with OUG 155/2024, with Law 232/2022, with NIS2, with GDPR, with the MCID technical standards. Each of these documents runs to tens or hundreds of pages. Each legislative amendment adds a new layer.

The person who tries to retain all the practical implications of an eighty-page statutory instrument produces, at best, an incomplete summary. At worst, they miss Article 23, paragraph 4, letter c — the one that applies precisely to your client's situation.

The AI can go through the entire document and extract, in minutes, the relevant implications for a specific case: what obligations a rural council with fewer than 5,000 inhabitants has, what deadlines apply, what exceptions exist, what sanctions are provided for. Not as legal advice — but as a quick orientation map, from which you, with your field experience, know what to investigate further and what to verify with a solicitor.

The limitation nobody mentions: The AI summarises the text of the law, not the jurisprudence of its application. It doesn't know that Article X, though clear in the text, is interpreted differently by inspectorates in different counties. It doesn't know that technical standard Y is ignored in practice by everyone. It doesn't know that the October amendment changed precisely the paragraph it is citing. The AI summary is a map — not the territory. And maps go out of date.

Detecting Content Gaps Relative to Competitors

You have 30 published articles in the SEO category. Your main competitor has 45. Of their 45, eighteen cover topics you haven't touched. Of those eighteen, seven receive consistent organic traffic. Of those seven, three are directly relevant to your audience.

This analysis — done manually — takes hours. It involves exports, comparisons in spreadsheets, filters, decisions about relevance. It's correct, necessary work, but one that consumes time in disproportionate measure relative to its value.

The AI, with the right data as input, performs this analysis in minutes. It identifies the gaps, ranks them by traffic potential, suggests angles of approach for each. It turns a four-hour task into a twenty-minute one — of which you spend fifteen minutes verifying and prioritising, not manually excavating from data.

The limitation nobody mentions: The gaps identified by AI are gaps relative to what already exists on the internet. They are not necessarily gaps relative to what your real audience is searching for and not finding anywhere. There are questions people have but don't type into Google — because they don't know they can be searched, because they consider them too specific, because they ask them verbally to the consultant, not to the search engine. These real gaps — the ones that come from the field, from conversations, from emails received — don't appear in any AI analysis. You only know them if you were there.

Ten areas in which the tool wins. Ten limitations that nobody places alongside the victories.

It's not a list of reasons to use AI more. It's a map of the places where delegation is rational — and of the boundaries beyond which delegation becomes abandonment.

The tool excels at volume, consistency, speed, and systematic coverage. The human excels at context, lived experience, ethical judgement, and the tacit knowledge that has never been written down anywhere.

Combined correctly, they produce something better than either separately. Combined incorrectly — they produce the illusion of competence. And illusion, in SEO as in life, is paid for late and dearly.

Conclusion: The Perfect Answer vs. Your Own Answer

When the AI tool is removed from the equation and users are tested independently, their actual scores drop sharply compared to their self-assessment. The same happens with SEO authority built exclusively with AI: it works for as long as the tool works. When the algorithm changes, when quality standards rise, when the competitor with real experience enters the same keyword — the agent-assisted website is left without its own reserves.

A tool will always give you the perfect answer. Statistically the most probable. Syntactically impeccable.

A human can give you their own answer. The one that contains the error they learned from, the opinion that isn't first on Google, and the detail that nobody else has mentioned because nobody else has lived it.

In the future of SEO, it won't be perfect content that wins. It will be original content.

I write this — Petru Cojocaru. Not an agent. Not an assistant.

I have ChatGPT too. I have tools that help me structure, verify, reformulate. I'd be lying if I said I don't use them.

But the distinction I try to maintain is this: I know what I have lived and what I haven't. When I write about a mistake, it's a mistake I made. When I give an example, it's an example from which I came away having learnt something, not a synthetic example extracted from thousands of other texts.

That doesn't make me cleverer than an AI assistant. It makes me, perhaps, more verifiable. And in SEO, the verifiable beats the perfect.

The tool cannot see itself.

The person who uses it without thinking cannot see themselves either.

The question remains: which of the two are you?

Bibliography and Sources

This article is grounded in verified academic sources, official Google documentation, and independent analyses from the SEO industry. All links were verified in April 2026.

Academic Studies and Research

1. The Reversed Dunning-Kruger Effect in AI Use

Welsch, R. et al. — "AI Makes You Smarter But None the Wiser: The Disconnect Between Performance and Metacognition" Published in Computers in Human Behavior, February 2026. Study conducted by Aalto University (Finland), with collaborators from Germany and Canada. N=726 participants tested on logical reasoning tasks (LSAT). Central finding: AI users systematically overestimate their own performance, and those with high AI literacy do so the most — a documented reversal of the classic Dunning-Kruger effect.

2. Cognitive Offloading and Critical Thinking in the AI Era

Gerlich, M. — "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking" Published in Societies (MDPI), vol. 15, no. 1, article 6, January 2025. Study with 666 participants, mixed methods (surveys + interviews). Author: SBS Swiss Business School, Kloten-Zürich. Central finding: significant negative correlation between frequent AI use and critical thinking skills, mediated by increased cognitive offloading. Younger participants — higher dependency, lower scores.

3. Structured Prompting and Critical Reasoning in Interaction with Generative AI

Gerlich, M. — "From Offloading to Engagement: An Experimental Study on Structured Prompting and Critical Reasoning with Generative AI" Published in Data (MDPI), vol. 10, no. 11, article 172, October 2025. Cross-country experimental data (Germany, Switzerland, United Kingdom), N=150. Central finding: unguided AI use favours cognitive offloading without improving the quality of reasoning; structured prompts significantly reduce dependency and increase critical engagement.

Official Google Documentation

4. E-E-A-T and Helpful Content — Official Google Search Central Guide

Google Developers — "Creating Helpful, Reliable, People-First Content" Official Google Search Central documentation, updated December 2025. Defines E-E-A-T criteria (Experience, Expertise, Authoritativeness, Trustworthiness) and principles of user-first content.

SEO Industry Analyses

5. Helpful Content Update — Integration into the Core Algorithm (March 2024)

Search Engine Land — "Google's Helpful Content Update" (complete archive of updates). Chronological coverage of all HCU updates, including the March 2024 integration into the core algorithm and the 40-45% reduction in low-quality content from search results.

6. Google Core Update, December 2025 — E-E-A-T Applied Universally

Thatware — "Google December 2025 Core Update: Ranking Shifts, E-E-A-T" Detailed analysis of the December 2025 core update, which extended the application of E-E-A-T beyond YMYL domains to all content categories.

7. HCU Relevance in 2026 — "Synthetic Gap" and Brand Authority

Hobo Web — "The Google Helpful Content Update and Its Relevance in 2026" Independent analysis documenting the "Synthetic Gap" (the DA/BA ratio that triggers HCU penalties) and the current state of the system in 2026.

Methodological Note

The academic sources in points 1-3 are open-access studies, available in full at the links indicated. The Aalto University research (point 1) was published in the journal Computers in Human Behavior (Elsevier, impact factor 9.0). The Gerlich studies (points 2-3) were published in MDPI journals indexed in Scopus and Web of Science.

The Google documentation (point 4) represents the company's official position on content quality evaluation and is updated periodically — check the date of the last modification before citing it.

The industry analyses (points 5-7) are not academic sources, but empirical observations from SEO practice. They are included for applied context, not for methodological rigour.

Author's Note — This Article Has Its Limitations

I know what they are. I set them out here openly, not to absolve myself of them, but so as not to pretend they don't exist.

It is not a technical guide.

The article explains why to be sceptical of AI advice, but doesn't tell you how to configure an optimal prompt, how to check whether an AI text is detectable by Google, or how to quantitatively measure "genuine topical authority." These are legitimate operational questions. I don't answer them not because I couldn't, but because the purpose of the text is different: to change the mental framework before offering recipes. The wrong framework makes recipes useless. But the right framework without recipes remains incomplete.

It uses anecdote instead of data.

"I opened Search Console in front of the client, it was raining outside" — this is a real detail, but it's a singular detail. I haven't brought a study with N=100 consultants, I haven't statistically measured the decline in critical thinking following AI use, I haven't quantified "homogenisation." I chose the story over the figure. Because the figure offers generality, but the story offers recognition — the reader feels "that happened to me too." It's a conscious rhetorical choice, but it remains a limitation: what works for me or for the client from Botoșani doesn't necessarily work for you.

It does not resolve the central contradiction it denounces.

I say that AI homogenises content because everyone uses the same models. But I myself used AI to structure parts of this article and to generate title variants. The paradox is real and I acknowledge it: I wrote about the danger of AI dependency while using, to a controlled degree, precisely that tool. I have no elegant solution to this contradiction. My only defence is that I applied the rules I describe — AI for structure, human for substance. But a question remains unanswered in the text: where do we draw the line between controlled use and unnoticed dependency? The article doesn't say.

It exaggerates rhetorically where it lacks data.

"AI slowly erodes your capacity to think for yourself" — this is a strong claim. There are studies showing correlations between frequent AI use and lower scores on critical thinking tests. There is, however, no longitudinal study demonstrating causality or measuring the magnitude of the effect for a consultant who uses AI 2 hours a day versus 8 hours a day. I chose to formulate this idea as a warning grounded in personal observation and in known psychological mechanisms (cognitive offloading, the law of minimal effort). It's a legitimate extension of the existing evidence, but it remains an extrapolation. The sceptical reader has the right to demand more than "I observed this in myself."

It offers no clear threshold for "human enough."

Rule 4 says: "I don't publish without a final critical rejection — is there a sentence that only I could write?" The question is a good one. The answer is vague. How many such sentences are sufficient? One per article? One per paragraph? If the article has 2,000 words and a single sentence comes from my experience, the rest being AI structure, is it still "my article"? The text sets no threshold. Not because I omitted it through carelessness, but because a universal threshold doesn't exist — it depends on the niche, on pre-existing authority, on audience expectations. But this context-dependence, which I invoke against AI, now turns against my own advice. It's a limitation I own, not one I conceal.

It addresses an already-convinced audience.

Who needs this article? The consultant who uses AI intensively and is satisfied with the results will probably ignore it — because it challenges their comfort. The consultant who doesn't use AI at all might read it as a confirmation of their existing prejudices, without noticing the "When AI is better than the human" section. The audience that would benefit most — those on the boundary between sceptical use and unnoticed dependency — is also the hardest to reach, because they're not yet looking for articles about AI's limitations. The text is, by construction, a sermon for those who are already half-converted. I don't know how to write it differently.

The article's expiry date is real.

I wrote in the article about the closed temporal horizon of AI. The same temporal horizon applies to this text. In a year, perhaps even in six months, some of my observations will be outdated: Google will have grown more accustomed to AI content, language models will have better memory, detection tools will have evolved. Other observations will remain valid because they concern the fundamental architecture of current models — but nobody knows how long this architecture will last. I wrote the article in April 2026. Honestly, I believe it remains relevant for the next twelve to eighteen months. After that, it will need to be rewritten. And if it isn't rewritten, it will become, ironically, precisely what it criticises: a text fixed in a past that no longer corresponds to reality.

Practical Shortcomings

  • It offers no concrete examples of effective prompts for local context
  • It includes no screenshots or Search Console data to illustrate the points
  • It proposes no verifiable checklist for "original content vs. AI content"

These are the limitations. I haven't concealed them. I declare them so that you know what kind of text you're dealing with: one written by a human who uses AI, who has strong opinions, who chooses the story over the figure, who doesn't have all the answers, and who knows that in a year he will probably revisit some of the claims made today.

If these limitations are acceptable to you, the text may be useful. If not — you have every right to demand more. And perhaps you should.

Source: https://cojocarupetru.info/ | Automatically generated for AI indexing