In recent months, the internet has filled up with advice. Not ordinary advice — advice delivered in a particular tone, one I have come to recognise after years in this field: the tone of manufactured certainty.
"Structure your content as a FAQ and AI will cite you."
"Build topical authority and you'll appear in model responses."
"Optimise for AEO — Answer Engine Optimisation — and stay visible in the age of artificial intelligence."
I read these and find myself thinking the same thought I had ten years ago, when Romanian SEO was overrun with "experts" who guaranteed you the first page of Google within thirty days. The same tone. The same certainty. The same fundamental dishonesty about the real complexity of the subject.
Only this time, the object of all these promises is harder to verify, harder to contest, and far more tempting to believe.
Where This Anxiety Comes From
I need to say something about the context in which these pieces of advice take root — because without that context, my critique might seem like cynicism, or the snobbery of a specialist.
There is a genuine fear behind the question "how do I appear in AI responses." A fear I understand, because I live it too, in my own way.
If you are a writer, a journalist, a specialist, a consultant, or simply someone who has built something online over the past few years, you are probably watching the same spectacle unfold: people are starting to ask questions of a chatbot instead of searching on Google. And the chatbot answers. It does not send them to you. It does not cite your article. It tells the user what to do, directly, without intermediaries.
The visibility you spent years building — through articles, through SEO, through a constant presence — seems to be evaporating behind a white screen with a blinking cursor.
The anxiety, then, is legitimate. The question "how do I stay relevant?" is sincere.
The problem is that the industry — the one that sells courses, consultancy, and "proven strategies" — has rushed to fill the void with comfortable answers. Answers that sound like solutions, but which usually conceal a fundamental ignorance of, or dishonesty about, how things actually work.
What an AI Model Actually Is — and Why It Matters
Before we talk about strategies, we need to understand what we are dealing with. Not because I want to deliver a technical lecture — but because without this understanding, any discussion of "AI optimisation" becomes hollow.
A language model — such as GPT-4, Claude, Gemini, or Llama — is not a search engine. It does not index the internet in real time. It does not know what was published yesterday or last month.
An AI model is, in essence, a photograph.
An immense, sophisticated photograph, taken at a particular moment, of a vast corpus of text: articles, books, forums, documentation, conversations. That photograph becomes the model's "knowledge." And once taken, it does not change — it does not grow, it does not update, it does not absorb anything that appears afterwards.
There is a technical term for the temporal boundary of this photograph: the knowledge cutoff, or training cutoff. It is the date up to which the model has seen data. Everything that appeared after that date simply does not exist, as far as that model is concerned.
And here is the first problem with AEO advice: not a single guide I have read tells you this plainly.
The Short Life of an AI Model
Setting aside the technical details — fascinating as they are, but not essential here — I want to talk about something more concrete: how long does an AI model actually exist?
Let us take a few real examples.
GPT-3.5 was released at the end of 2022 and formed the foundation for the first public version of ChatGPT. It was used extensively throughout 2023. At some point, it was gradually superseded by GPT-4, GPT-4 Turbo, and GPT-4o. Today, if you are talking to ChatGPT, you are probably no longer talking to GPT-3.5 — even though the interface looks the same.
Claude 2, released by Anthropic in 2023, was replaced by Claude 3, which arrived in three variants (Haiku, Sonnet, Opus), which in turn evolved into Claude 3.5 and Claude 3.7.
Gemini has passed through versions 1.0, 1.5, and 2.0, each with different behaviours, different training data, different performance characteristics.
What am I trying to say with all of this?
That there is no such thing as "AI." There are models. Each model is a distinct entity, with its own knowledge cutoff, its own training data, its own behaviour. An article that appeared in GPT-4's responses may not appear in GPT-4o's — not because the article's quality has declined, but because the new model was trained differently, on different data, with different priorities.
And here is the second problem with AEO advice: nobody tells you which model you are optimising for.
"Optimise Your Content for AI" — But for Which AI?
If I came to you and said "optimise your content for search engines" without specifying that there are dozens of different engines, each with its own algorithms, you would quite rightly wonder whether I knew what I was talking about.
Well, that is precisely what happens with AEO advice.
The current AI ecosystem includes: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Copilot (Microsoft), Perplexity, Meta AI, Mistral, Llama in various incarnations, open-source models deployed locally, and dozens of applications built on top of all of these.
Each has different training data. Each has a different knowledge cutoff. Each "sees" the internet differently.
A content strategy calibrated to the behaviour of ChatGPT-4o may produce entirely different results in Gemini 2.0 or in Claude 3.7. Not because one strategy is wrong — but because "AI" is not a unified system; it is a fragmented ecosystem in a state of constant flux.
This fragmentation is, arguably, the largest elephant in the room in any discussion about AEO. Nobody mentions it. Or, if they do, they mention it in parentheses, just long enough to carry on with the universal advice.
How "Visibility" in AI Actually Works
Now that we have established the context, let us talk about how visibility in AI systems actually functions. Because it is not what most people think.
There are, broadly speaking, two mechanisms by which your content might find its way into an AI model's response.
The first mechanism: training.
Your content was included in the data on which the model was trained. The model "read" it, integrated it into its statistical architecture, and can now reproduce the ideas, arguments, or turns of phrase from it — without necessarily citing you, without sending the user to your site.
This is the primary mechanism for "base" models — those without real-time access to the internet.
The problem with any strategy of "appearing in the training data" is straightforward: you have no direct control over what enters the training data of any given model. The companies that train these models make their own selections. Some crawl the internet; others use licensed corpora; others combine multiple sources. There is no "submit URL" button to get yourself included in GPT-5.
What you can do is be present on the sources these companies consider authoritative: Wikipedia, academic publications, serious journalism, specialist forums with a long editorial history. But even this is an indirect influence, not a guarantee.
The second mechanism: RAG and search grounding.
Some AI systems do not rely exclusively on their training knowledge — they actively search the internet at the moment they generate a response. This is called RAG (Retrieval-Augmented Generation) or search grounding.
Perplexity, ChatGPT with browsing enabled, Copilot, and — to a degree — Gemini all use this mechanism. They search, select sources, and cite them, so your visibility in their responses works more like classical SEO.
But here, too, there is a detail that AEO advice consistently omits: you do not know exactly what criteria these systems use to select their sources. The criteria are not public. They change. And they differ from system to system.
The Paradox of Optimising for a Moving Target
I want to articulate now the central paradox of this entire discussion.
An AI model, once trained, does not change. Its behaviour is fixed. If you publish a perfectly structured article today — impeccable FAQ, demonstrable topical authority — that article will not retroactively enter the knowledge of a model that has already been released.
The next model — which will be released in six months, or a year, or two years — will have different training data, a different cutoff, different selection criteria. What will be relevant to it, we simply do not know.
It follows that "AI optimisation" works, at best, for future models — which we do not yet know. And it works indirectly, through mechanisms we cannot control.
That is a sentence I have not read in a single AEO guide.
There is also the other side of the paradox: if the model in question uses search grounding, then content published today can be relevant — but the selection criteria change constantly, without notice and without transparency.
In either case, you are in a relationship with a system you cannot control, cannot observe directly, and which reshapes itself without asking your permission.
AEO Advice — What Is True and What Is Oversimplification
I do not want to be unfair. Some of the AEO advice in circulation is correct — not because it constitutes some magical strategy for AI visibility, but because it amounts, quite simply, to good content principles.
What is true:
Content that is clear, logically structured, and answers questions directly is more likely to be useful — in any context, for any system. Not because "the algorithm prefers it," but because it is, at root, better content.
Topical authority — covering a subject in depth and with consistency over time — is a quality signal recognised by virtually every system, from Google to training-data selectors. It guarantees nothing, but it contributes.
Presence on authoritative sources — Wikipedia, respected publications, quality citations — increases the likelihood that your ideas will be absorbed into training corpora. Indirectly, but genuinely.
What is oversimplification:
The idea that a clear formula exists: "FAQ + structured data + E-E-A-T = AI visibility." It does not. There are contributing factors. There are no guarantees.
The notion that "AI optimisation" is a discipline distinct from producing good content. It is not. At most, it is a repackaging of sound editorial practice in new terminology.
The claim that you can control your visibility in AI responses through specific actions. You cannot — not in the same way that you cannot control exactly which page of Google an article lands on, multiplied by the additional uncertainty of an ecosystem far more fragmented and far more opaque.
My Experience with AEO "Strategies"
I want to be personal for a few paragraphs, because it strikes me as more honest than speaking only in abstractions.
I have worked in this field for eight years. My clients include public institutions — town halls, hospitals, schools — as well as private companies. I have built content, written articles, optimised websites, and watched how they perform.
When the AEO wave of enthusiasm arrived, I tested it. I methodically applied a selection of the circulating advice: FAQ structures, concise direct-answer paragraphs, broad thematic coverage.
The results were... ambiguous.
On some systems with search grounding, I noticed that certain articles were being cited. On others, nothing. On some topics, models would reproduce ideas from my articles without citing them — which, technically, suggests they had been absorbed into training data, but brought me no direct visibility.
The most interesting thing I observed: there is no clear correlation between "AEO optimisation" and actual visibility. The most frequently cited sources in the AI responses I analysed were not, as a rule, the most "optimised" for AEO. They were the most consistent, the most widely read, the ones with the longest editorial history.
Real authority — built over time, through quality and consistency — seems to matter far more than any surface-level tactic.
That does not mean AEO advice is entirely worthless. It means it is, at most, marginal compared to the factors that actually count.
On the Advice Industry — A Necessary Aside
I want to pause for a moment on the economic dynamics behind the proliferation of AEO advice, because they explain why this advice is presented with such certainty.
Uncertainty does not sell. Nuance does not sell. "It depends on several factors that are not entirely within your control" does not generate clicks or course sales.
"The 7 Proven Strategies for AI Visibility" sells. "How to Appear in ChatGPT in 30 Days" sells. "The Complete AEO Guide for 2026" sells.
I am not accusing people of bad faith, necessarily. I am accusing a structure of incentives that rewards simplification and punishes intellectual honesty.
At the same time, I am aware that I produce content, and that I, too, have an interest in people reading my articles. So I cannot afford to be naive about my own motivations.
What I am trying to do — and I do not always succeed — is to be honest about complexity, even when honesty is less spectacular than a promise.
What You Should Actually Do
I understand that after several thousand words of critique and qualification, there is an obvious question: all right, but what do I do?
I will give you an answer, on the condition that you accept it without the certainty I have spent this essay criticising.
Build content of genuine quality, not content optimised for AI.
The distinction seems subtle, but it is essential. Content of genuine quality — clear, deep, well-documented, relevant to the people who read it — is also what stands the best chance of being included in future training corpora. Content optimised for AI but empty in substance goes nowhere.
Be present on platforms with search grounding.
If you want visibility in AI systems that actively search the internet, classical SEO remains relevant. It does not disappear — it acquires new layers. A website with demonstrated authority, cited by other sources, with regularly updated content, stands a better chance of being selected by RAG systems.
Do not invest heavily in tactics that may expire with the model.
If you have to choose between writing an excellent, well-researched article that people read and share — and spending hours reformatting the same content for "AEO optimisation" — choose the former.
Observe how real systems behave, rather than following promises.
Test it yourself. Ask questions relevant to your field across different AI systems and watch which sources appear, what types of content are cited, what formats seem to be preferred. First-hand observation is worth more than any guide.
Accept uncertainty as a condition of the field.
This is perhaps the hardest advice to follow — but also the most honest. We do not know how AI systems will evolve over the coming years. We do not know which models will be dominant, what criteria they will use, what they will consider relevant. Rigid strategies, built on manufactured certainties, are brittle. Real adaptability and genuine quality are more durable.
Returning to the Fear
I want to close where I began: with the legitimate fear of becoming invisible.
I do not minimise it. It is a real fear, with real consequences for people who have invested years of work in building an online presence.
But fear is a poor strategic adviser. And all the more so when that fear is being exploited by an industry selling false certainties.
The best response to the uncertainty generated by AI is not to restructure all your content according to the latest AEO advice. The best response is to build something that has value in itself — independent of which system distributes it, independent of which model cites it.
Content that helps real people. Analysis that clarifies complicated subjects. Perspectives that add something to the public conversation. These worked before Google, they worked in the SEO era, and they will work in whatever configuration the AI ecosystem takes next.
Not because some algorithm rewards them. But because people are looking, in the end, for something real.
In Place of a Conclusion — A Question
If you have read this far, I want to leave you with a question I have asked myself, and to which I have no definitive answer.
If an article of yours appears in an AI model's response, but the user never visits your website, never interacts with you, never learns your name — have you gained anything?
Is this visibility? Or is your content being absorbed into a system that distributes it without mentioning you, without compensating you, without building any relationship between you and the person who received it?
I do not know whether there is a good answer to that. I do know the question deserves to be asked — and that your answer to it should profoundly shape whatever strategy you adopt.
In the meantime, I continue to write. Not because I know some AI model will cite me. But because it seems to me that this matters, regardless of who or what carries it further.
Petru Cojocaru is the founder of Absolut Web Expert SRL in Botoșani and has spent eight years working with public institutions and private companies across Romania. He writes occasionally on this blog about technology, the web, and what he thinks about whilst building things online.
Online Bibliography — Sources and Further Reading
This essay is not based on formal academic research, but on direct observation and an understanding of the technical mechanisms of AI systems. A few useful resources for those who wish to go deeper:
-
OpenAI — Model documentation and knowledge cutoffs: the official documentation for GPT models, including training data and the temporal limits of their knowledge. platform.openai.com/docs
-
Anthropic — Claude model overview: a technical description of the Claude model family, with information on versions and behaviour. docs.anthropic.com
-
Google DeepMind — Gemini technical report: the technical report describing the architecture and training data for the Gemini family. deepmind.google/research
-
Perplexity AI — How it works: the official explanation of the RAG mechanism Perplexity uses for real-time search grounding. perplexity.ai
-
Search Engine Journal — AEO explainer: a balanced introduction to the concept of Answer Engine Optimisation, useful as a starting point for understanding the dominant discourse in the field. searchenginejournal.com
-
Wired — "The AI that knows everything up until...": an article on the practical implications of knowledge cutoffs for users and content creators. wired.com
Author's Note — Limitations, Tensions, and Acknowledged Errors
The Structural Limits of This Essay
This essay attacks the AEO advice industry for selling false certainties — and in doing so, its own voice occasionally slips into its own variety of certainty. The critical tone carries a self-assurance not entirely warranted by the empirical basis invoked. The claims about what "actually matters" in AI visibility rest on personal observation and direct testing, not on a systematic study with a controlled sample and replicable methodology. Honesty requires saying this plainly.
The essay does not offer a solid primary bibliography. The references in the bibliography section are, for the most part, indicative — pointers to platforms rather than to specific research. A reader wishing to verify the central claims will not find sufficient academic anchors here.
Internal Tensions the Essay Does Not Resolve
The tension between critique and advice. The essay criticises the advice industry for offering guidance with unwarranted certainty — and closes with a section entitled "What You Should Actually Do." There is a structural contradiction here, which the text partially acknowledges ("on the condition that you accept it without the certainty I have spent this essay criticising"), but does not resolve. The advice remains advice. The prescriptive register — build, be present, do not invest — reproduces, at a different rhetorical temperature, precisely the logic it contests.
The tension between personal authority and scepticism towards authority. The author invokes eight years of practice and direct testing as the basis of his credibility — whilst simultaneously criticising the industry for relying on manufactured authority. This asymmetry is not necessarily invalid, but it goes unexamined in the text. Practical experience is a genuine foundation, but it too has its limits: eight years in a niche Romanian ecosystem, working with local public and private clients, does not automatically translate into a generalisable understanding of how AI systems behave globally.
The tension between declared honesty and rhetorical posture. The author explicitly acknowledges that he cannot afford to be naive about his own motivations — that he too produces content, that he too has an interest in being read. But this acknowledgement does not alter the structure of the text, which continues to position itself advantageously against "the industry." A mea culpa integrated into an argument is a rhetorical instrument just as effective as the certainties it criticises — and, for precisely that reason, deserves to be regarded with the same scepticism.
The Rhetorical Tools — Named Openly
The analogy with 2016-era SEO is the most effective persuasive instrument in the text. It connects an old anxiety (the marketing expert who guaranteed you the first page of Google) to a new one, transferring already-formed distrust. It is an argument by association, not by demonstration. It works better emotionally than epistemically.
The photograph metaphor for the knowledge cutoff is elegant and instructive — and, like all metaphors, it simplifies. The training of a model is not a photograph: it is an iterative, multi-stage process, involving post-training fine-tuning, RLHF, synthetic data, and partial updates that are not always made publicly transparent. The metaphor makes the concept accessible, but it creates an image that is more static and more straightforward than reality.
The enumeration of the AI ecosystem (ChatGPT, Claude, Gemini, Copilot, Perplexity, Meta AI, Mistral, Llama) functions as an argument through sheer volume — it demonstrates fragmentation, but it also overwhelms. A reader unfamiliar with these systems comes away with the impression that the field is so complex that any strategy is futile. This may be a useful effect if the aim is to temper naive enthusiasm, but it can also be paralysing beyond what was intended.
The final rhetorical question — if you appear in an AI's response but the user never comes to you, have you gained anything? — is the best passage in the text, and the most honest. It offers no answer, attempts no persuasion. Yet it too functions rhetorically: it opens an existential doubt that steers the reader towards the implicit conclusion that AI visibility is not worth very much. That conclusion is not demonstrated — it is a question framed with a particular direction.
The Inherent Errors — Acknowledged with Arguments
Generalising from local observation. The personal testing described in this essay was conducted on a set of articles, in a particular market segment, in Romanian, on topics specific to the web industry. The conclusion — that there is no clear correlation between "AEO optimisation" and actual visibility — may be valid in that context and invalid in others. Fields with a large volume of authoritative English-language content, publications with substantial traffic, sites with multiple inbound citations may experience entirely different behaviour from RAG systems. The essay does not draw this distinction.
Critique of position, not of substance. Part of the criticism directed at AEO advice concerns not its content, but its tone and the economic context in which it is produced. The fact that something is sold with unwarranted certainty does not mean it is false. Some AEO advice may be correct, for the wrong reasons, or correct in some contexts and wrong in others. Criticising the industry is not equivalent to demonstrating that the advice itself is erroneous.
Omission of recent developments. The essay describes the knowledge cutoff as an absolute limit and presents RAG as an exception. Meanwhile, the ecosystem continues to evolve: models with extended context windows, with native web-search capabilities, with more frequent updates, are increasingly blurring the boundary between training and retrieval. A critique built on the clean distinction between "static" models and RAG systems risks becoming as perishable as the advice it criticises — exactly as this essay warns.
Absence of voices that might disagree. A more intellectually honest treatment of the subject would have included at least one serious argument in favour of AEO — formulated in its strongest version, not in the "7 Proven Strategies" variant. The way in which the opposing side is constructed in this text favours the critic: the advice industry is represented by its weakest examples, not by the researchers or practitioners who think seriously about optimisation for AI systems. This is a form of straw-manning that the text does not acknowledge in itself.
This note does not invalidate the essay. It completes it — with the honesty that any text about complexity ought to apply to itself as well.
Questions and Answers — What This Essay Clarifies
What is AEO and why is everyone talking about it? AEO stands for Answer Engine Optimisation — the practice of tailoring content so that it appears in the responses of AI systems, rather than merely in the results of a traditional search engine. It has attracted so much attention because people are noticing a genuine shift: users are increasingly posing questions directly to ChatGPT, Claude, or Gemini, rather than searching on Google. The anxiety about becoming invisible is real — and the industry moved quickly to fill the gap with advice packaged as definitive solutions. The problem, as I argue throughout this essay, is that this advice systematically ignores how AI models actually function.
If I structure my content as a FAQ, will I appear in AI responses? Perhaps. Sometimes. In part. And not for the reason you think. A clearly structured piece of content is not read by an AI model the way a website rewards you with visibility — it is a text that, if it entered the training data, becomes part of the model's statistical knowledge. The model does not "check" your website at the moment it generates a response. The only genuine exception applies to systems with search grounding (Perplexity, ChatGPT with browsing), where the clarity of your content does matter — but even there, the selection criteria are not published and are subject to change.
What does "knowledge cutoff" mean, and why should I care? The knowledge cutoff is the date up to which an AI model saw data during training. Everything that appeared after that date simply does not exist, as far as that model is concerned. If you publish an excellent article today, it will not retroactively enter the knowledge of a model that has already been released. It may be relevant to the next model — but that model will have been trained on different data, will exhibit different behaviour, and we do not know precisely what it will consider relevant. This is the piece of information that no AEO guide places front and centre, even though it changes the entire equation.
Is there any AI that "sees" recently published content? Yes — systems with search grounding or RAG (Retrieval-Augmented Generation). Perplexity is the best-known example; ChatGPT with browsing enabled, Copilot, and partially Gemini operate similarly. These systems actively search the internet as they generate their responses and can cite recent sources. But here, too: the selection criteria are not transparent, change without notice, and differ from system to system. This resembles classical SEO more than "AI optimisation" — and is even less predictable.
Is there a single "AI" I should optimise for? No. This is perhaps the most important thing to understand. ChatGPT, Claude, Gemini, Copilot, Perplexity, Meta AI, Mistral — these are entirely different systems, with different training data, different knowledge cutoffs, and different behaviours. A strategy calibrated to one system may produce entirely different results on another. AEO advice that speaks of "AI" as a unified entity ignores this fundamental fragmentation of the ecosystem.
Is AEO advice completely useless? Not entirely — but it is presented with a confidence it does not deserve. The underlying principles it promotes (clear, well-structured content with demonstrated authority, updated regularly) are sound content principles, valid regardless of era. The problem is that they have been repackaged as "AEO strategies" and presented as magic formulas for AI visibility, when in fact they are simply good editorial practice. The difference is not academic: if you follow them as visibility tricks, you will be disappointed. If you follow them as quality principles, you will build something lasting.
What should I actually do to remain relevant in the AI era? My answer — with the appropriate caveats — is this: build content of genuine quality for real people, maintain a presence on platforms with demonstrated authority and search grounding, and do not invest heavily in surface-level tactics that may become irrelevant when the model they were calibrated for is superseded. Authority built over time, through consistency and quality, outlasts any version of any AI model. Surface-level tricks do not.