- AI
- AI Search
- AEO
What is AI?
Artificial intelligence, usually shortened to AI, is a broad term for technology that allows computers to perform tasks that normally require some form of human intelligence. That can include recognizing what is in an image, deciding whether an email looks like spam, recommending a song, predicting traffic, translating a sentence, generating an image, or answering a question.
You probably interact with AI more often than you realize. Gmail filters unwanted messages. YouTube recommends videos. Spotify suggests music based on what you listen to. Google Maps predicts traffic and travel times. Products such as ChatGPT, Gemini, Claude, and Microsoft Copilot can interpret everyday instructions and generate responses.
All of these experiences can involve artificial intelligence, even though they solve very different problems.
AI is not one program, one model, or one product. ChatGPT is an AI product, but it is not the definition of AI. The same applies to Gemini, Claude, Copilot, Perplexity, and the growing number of products that use AI behind the scenes. Artificial intelligence is the larger field behind them.
AI also existed long before today’s generative AI boom. Machine learning has powered recommendations, fraud detection, image recognition, navigation, search systems, and other technology for years. Most of that AI was easy to overlook because it worked quietly in the background.
Generative AI made the technology much more visible. Instead of AI quietly determining which video you might enjoy, you can now ask a system to write an email, analyze a document, create an image, explain a difficult concept, search for information, write code, or help complete a task. That can make the subject feel more complicated than it needs to be.
You do not need to understand every algorithm or know how to build a neural network to understand AI. A better approach is to learn a few pieces, see how they connect, and gradually build the larger picture. This guide does exactly that.
What you’ll understand
By the end, you should have a practical understanding of:
- what artificial intelligence actually means
- how machine learning and AI models work
- what makes generative AI different
- how large language models generate responses
- where AI systems get their information
- why AI can hallucinate or make mistakes
- how retrieval, AI search, and AI agents fit into the larger picture
The basic idea
A useful starting point for much of modern AI is a simple process:
- Data provides examples or information.
- Training uses that data to adjust a model.
- The trained model receives new input.
- The model produces an output.
Depending on the system, that output might be a prediction, classification, recommendation, generated response, image, or another result.
Modern AI products can add much more around the model, including conversation history, uploaded files, search, retrieval, memory, tools, and connections to outside information. We will add those pieces later, once they are useful.
Chapter 01 — AI around you
AI you already use
- Before the chatbots
None of this is new. Long before an AI product could write you a paragraph, the same broad field was quietly sorting your inbox, ordering your feed, and estimating how long your drive would take.
- Classification
A mail service decides whether a message belongs in your inbox or in spam. It is sorting things into categories it has learned to tell apart.
- Recommendation
A video service picks what to put in front of you next, and a music service guesses which track you will want after this one. Both are ranking options against what they have learned about behaviour.
- Prediction
A maps application estimates how long your drive will take, then revises it while you are driving. It is not reading the future. It is projecting from patterns in traffic data.
- Recognition
A photo library finds every picture with a dog in it, and nobody labelled your library by hand. The system learned what the thing looks like.
- Generation
Only the last of these produces something that did not exist before. That is the part that became visible to everyone in the last few years, and it is why many people now assume it is the whole field.
- One field
These are different jobs, built differently and judged differently. What they share is a name. Artificial intelligence is the field they all sit inside, not any single product in it.
Artificial intelligence is a field, not a product. Classification, recommendation, prediction, recognition, and generation are different jobs, and you were already using most of them.
What artificial intelligence actually is
Artificial intelligence covers a wide range of technologies. A system that recommends a video is solving a different problem from one that creates an image. A fraud-detection model has different requirements from a conversational assistant. Software analyzing a medical image is doing something different from software generating computer code.
They can still fit under the larger AI umbrella because they give computers capabilities related to recognizing patterns, processing information, making predictions, generating outputs, making decisions, or performing other tasks associated with intelligence.
AI is a category, not a product
The prominence of products such as ChatGPT has created an understandable shortcut in how people talk about AI. Someone says “AI” when they really mean “an AI chatbot.” The terms are related, but they describe different things.
| Term | What it means |
|---|---|
| Artificial intelligence | The broad field of technologies designed to perform tasks associated with intelligence. |
| Machine learning | One major approach used to train systems to learn useful patterns from data. |
| Generative AI | AI designed to create new outputs such as text, images, audio, video, or code. |
| Large language model | A type of AI model designed to process and generate language. |
| AI assistant or chatbot | An interface through which someone interacts conversationally with an AI system. |
| ChatGPT, Gemini, Claude, Copilot, Perplexity, and similar products | Individual products that combine models with interfaces and other supporting systems. |
There is no need to memorize the table. The distinctions become clearer as we work through the article. For now, the simplest version is enough: AI is the larger field. The applications people use are products built with AI.
Chapter 01, continued — AI is bigger than a chatbot
Why AI is not the same as an AI chatbot
- The picture in your head
For most people, the mental image of artificial intelligence is now a chat window. It is a reasonable shortcut, because it is the part of AI that started talking back.
- Pull back
But a chat window is an interface, not a technology. Something has to be behind it, and that something is shared with software that looks nothing like a conversation.
- The neighbours
The same broad field decides what you get recommended, reads what is in a photograph, estimates a delivery time, ranks a page of search results, flags a suspicious transaction, and turns speech into text. None of them chat.
- Same field
All of it sits inside artificial intelligence. The chatbot is not the centre of the field, and it is not a summary of it. It is one application among many, and a fairly recent one.
- Why it matters
Keeping the two apart makes the rest of this article much easier. When you read that a product 'uses AI', the useful question is which capability it is using, and what that capability is actually good at.
An AI chatbot is one application of AI. The field it belongs to also covers recommendation, vision, prediction, ranking, detection, and speech — none of which involve a conversation.
What can AI actually do?
AI can perform several broad types of tasks:
- Recognize information. Computer vision can identify objects, handwriting, road signs, faces, medical patterns, and other visual features.
- Classify information. Spam filters, fraud systems, and moderation tools can place information into useful categories.
- Predict outcomes. AI can recommend content, estimate travel times, forecast demand, or calculate risk.
- Generate content. Generative AI can create text, images, code, audio, video, and other outputs.
- Retrieve information. AI systems can find relevant information in documents, databases, search indexes, and websites.
- Take actions. Systems with access to tools can interact with software or complete parts of a workflow.
These capabilities can overlap. One AI assistant might retrieve a webpage, analyze it, generate a summary, and use another tool during the same task.
Is AI the same thing as automation?
No. AI and automation are related, but they describe different things. Automation means using technology to perform work with less manual intervention.
Consider a basic workflow:
When someone submits this form, send them a confirmation email.
That process does not need to interpret the message or learn anything. It follows a predefined instruction. Now consider a system that reads the form submission, determines what kind of request it contains, summarizes it, routes it to the correct department, and drafts a response. AI can contribute to several parts of that workflow.
Automation describes how work gets carried out. AI describes capabilities that can help a system interpret information, make predictions, generate outputs, or decide what to do. Automation can use AI, but it does not have to.
What types of AI are there?
There is no single universally accepted list of “types of AI.” Different explanations classify AI according to different characteristics.
One common distinction is based on capability. Narrow AI describes systems built to perform particular types of tasks, and today’s AI systems fit within this broad category even when one product can handle many different activities. Artificial general intelligence, or AGI, generally refers to the theoretical idea of AI with broadly general abilities across many intellectual tasks. The exact definition remains debated, so AGI is better understood as an open technical concept than as a clearly defined product milestone.
AI can also be described according to what a system does:
- Predictive AI estimates outcomes.
- Generative AI creates new outputs.
- Conversational AI interacts through dialogue.
- Agentic AI describes systems that can work toward goals and potentially use tools to take actions.
Those labels can overlap. A conversational assistant can use generative AI, retrieve information, and have agentic capabilities at the same time. Thinking in terms of overlapping capabilities is more useful than trying to place every AI product into one exclusive category.
Chapter 01, continued — Four capabilities, not four categories
What AI systems are built to do
- The same field
This is not a new diagram. It is the field from a moment ago, divided by what a system is actually built to do.
- Predict and generate
A recommendation engine mostly predicts. An image generator mostly generates. Those are genuinely different problems with different data, different objectives, and different ways of being judged.
- The overlap is the point
An assistant such as ChatGPT, Gemini, Claude, or Copilot sits across generate and converse at the same time. The regions overlap because real products do.
- Reaching into action
Connect that same assistant to tools and it extends into acting: running a search, executing code, filling something in. The capability changed without the category changing.
- Not a ladder
It is tempting to read this as a progression, as though systems graduate from predictive to generative to agentic. They do not. A system is built for the region the job needs.
- Narrow to general
One more distinction is worth keeping separate. Everything shipping today is narrow AI, built and trained for particular kinds of work, however wide the range looks. Artificial general intelligence sits at the far end as a contested idea rather than a product you can buy.
Predict, generate, converse, act. Overlapping capabilities, not a sequence a system moves through.
Why understanding AI matters
You do not need to become an AI engineer to understand AI well enough to use it intelligently. You do need enough context to judge what these systems can do and where their limitations begin.
AI increasingly affects how people find information, what gets recommended to them, how software responds, how businesses analyze data, how creative work gets produced, and how repetitive work is automated. A basic understanding also helps answer practical questions:
- Why can an AI system confidently say something that is wrong?
- Why can one assistant access information another cannot?
- Does AI learn everything you type into it?
- What is the difference between a model and the product built around it?
- How can an AI system work with information created after the model was trained?
- How can an AI move from answering a question to completing part of a task?
Those questions become much easier once you understand how AI learns.
How AI learns
Much of modern AI can be understood through the relationship between data, training, models, and new inputs. The technical details can become extremely complex, but the basic idea does not have to be.
What is machine learning?
Imagine trying to build an email spam filter entirely through manually written rules. You could tell the software to flag certain phrases, suspicious links, or sender patterns. That might work for a while, but the number of possible situations would quickly become enormous. A spam message might avoid the exact words you expected. A legitimate message could accidentally match one of your rules. Attackers could change their tactics.
Machine learning provides another approach. Machine learning is a way of building computer systems that learn useful patterns from data and apply those patterns to new information. Machine learning is part of AI. It is not another word for artificial intelligence as a whole.
What is training data?
Training data is the information used to help a model learn. That information is often organized into one or more datasets, which are collections of data used for training, evaluation, analysis, or another purpose.
Depending on the model, training data might include text, photographs, video, audio, numerical records, computer code, labeled examples, or behavioral information.
Suppose we want to train a model to recognize dogs in photographs. We could show it many images containing dogs and many images that do not. A real dog does not have one exact combination of pixels — dogs appear in different breeds, sizes, poses, lighting conditions, and backgrounds — so the model needs to learn patterns that generalize beyond any one training example.
How does AI training work?
The exact mechanics vary between models, but a simplified training process looks like this:
- The model receives training examples.
- It produces an output or prediction.
- The training process measures that result against an objective.
- The model’s internal parameters are adjusted.
- The process repeats across many examples.
Over time, those adjustments can improve the model’s ability to perform the task it is being trained for. The model is not simply reading each example once and placing the correct answer into a giant filing cabinet. Training changes a mathematical system.
Chapter 02 — Data becomes a model
Watching a model get trained
- A few examples
Training starts with examples: a set of pictures that contain a dog and a set that do not. Shown a photograph it has never seen, a model at this stage is essentially guessing.
- Repetition
Each pass compares what the model produced against what the training objective asked for, adjusts the internal parameters, and goes again. Nothing is being filed away. A mathematical system is being tuned, slowly, across an enormous number of passes.
- Variety
As more varied examples move through — different breeds, angles, lighting, backgrounds — the adjustments accumulate into something that generalises. The model starts getting new photographs right, including ones no training example resembled closely.
- What is in the data matters
Now narrow the examples so they cover one kind of dog in one kind of light. The process still runs and a model still comes out of it. It is simply less reliable the moment something unfamiliar arrives. The quality, coverage, and construction of training data shape what a model can do.
- Not a filing cabinet
This is the part most often pictured wrongly. A database stores records so that one exact record can be retrieved again. Training leaves behind learned relationships spread across parameters, which is a genuinely different kind of object — and the reason asking a language model a factual question is not the same as looking something up.
- Take the data away
The training data is not part of the finished product. It did its work by changing the model.
- What is left
One object remains. Everything else in this article happens around it.
Data → Training → Model. Training changes the model; what survives the process is not the data but a trained mathematical system.
What is an AI model?
An AI model is the learned mathematical system created or adjusted during training. Once trained, the model can process new information and produce an output based on the relationships it learned.
One common misconception is that a model works like a database containing every piece of information it has encountered. It does not. A database is normally designed to store specific information so records can be retrieved in a structured way. A trained AI model instead encodes learned relationships through mathematical parameters.
That is one reason asking a language model a factual question is different from looking up one exact row in a database. The model may have learned enough relationships to produce the correct answer, but that does not mean the answer exists inside it as one perfectly stored entry.
What is inference in AI?
Once a model has been trained, it can be used. That use is generally called inference.
| Training | Inference |
|---|---|
| Changes or improves the model. | Uses the trained model. |
| Happens while building or refining a model. | Happens when a new request is processed. |
| Adjusts model parameters. | Normally does not retrain the model. |
| Can require very large amounts of computing power. | Produces outputs from new inputs. |
When you ask an AI assistant a question, generate an image, or use another AI feature, you are usually seeing inference rather than the model retraining itself in real time.
Same object, two different jobs. Training data flows into the model and changes it. A request passes through the model and leaves it exactly as it was.
When you ask an assistant a question, generate an image, or run an AI feature, you are almost always watching inference — not a model retraining itself on what you just typed.
Does AI learn from everything you type?
Not in the simple sense people often assume. Your current conversation can become part of the context an AI assistant uses to answer later messages, and some products can also store memories or other information for future interactions. That is different from the underlying model instantly retraining itself on every sentence you type.
Companies also have different policies regarding how interactions may later be stored or used to improve products and models. Those practices need to be evaluated at the product level rather than assumed to work the same way everywhere.
What is a neural network?
Many of today’s most capable AI systems use a form of machine learning called deep learning, which commonly relies on neural networks. A neural network is a mathematical system made up of connected layers that transform information.
Information enters the network, mathematical operations process it through a series of layers, and those transformations contribute to an output. During training, many of the parameters controlling those transformations are adjusted.
The name “neural network” was inspired by biological neurons, but artificial neural networks are not digital replicas of a human brain. They are mathematical systems.
What is deep learning?
Deep learning is a form of machine learning that uses neural networks with multiple layers to learn complex patterns. These techniques have helped drive advances in computer vision, speech recognition, language processing, and generative AI.
The relationship between several common terms can be summarized as artificial intelligence → machine learning → deep learning. Many modern generative AI systems are built using deep-learning models, although AI as a whole is broader than this simplified hierarchy.
Chapter 03 — Inside the model
What a neural network is
- Going in
The box has been opaque so far. Inside a great deal of modern AI is a neural network: layers of simple mathematical operations wired together, with the strength of each connection set during training.
- One pass
Information enters at one side. Each layer transforms it and passes it along. What comes out the other side is the output. That is the whole mechanism, repeated at scale.
- Wrong, then adjusted
During training, when the result misses, the values controlling those transformations are nudged. No single adjustment means much. The accumulation of an enormous number of them is what learning looks like here.
- Better
Run it again and the same input lands differently. That is all improvement is at this level: the same arithmetic with better-tuned numbers.
- Scale
This diagram has a few dozen connections so that it fits on a screen. A production model has billions of parameters. The mechanism does not change with size. What changes is the complexity of the relationships it can represent.
- Not a brain
The name was borrowed from biology, and it has caused a lot of confusion. An artificial neural network is not a digital copy of anything — no neurons, no thoughts, no inner life. Deep learning simply means networks with many layers, and it is what drove the recent advances in vision, speech, language, and generation.
- Where it sits
Which gives you the nesting worth remembering. Artificial intelligence contains machine learning. Machine learning contains deep learning. Most of the generative systems you have heard of are built on the innermost box.
A neural network is arithmetic, not biology. Layers transform information, training tunes the numbers that control those transformations, and depth is what makes complex patterns learnable.
Where do algorithms fit?
An algorithm is a procedure or method used to solve a problem. AI systems can use many algorithms, and the terms algorithm and model are sometimes used loosely in everyday conversation, but they do not mean the same thing. An algorithm describes a process or method. A trained model contains parameters and relationships learned through training.
So far, we have mostly discussed AI recognizing, classifying, or predicting something. Generative AI adds another capability: creating something new.
How generative AI creates something new
Suppose two AI systems receive information about a customer. One predicts that the customer has an 82% chance of canceling a subscription. Another uses the same information to write a personalized retention email. Both can involve AI. The first predicts an outcome. The second generates something new.
Predictive AI and generative AI
Predictive AI estimates, classifies, ranks, or forecasts an outcome. Generative AI produces new content based on relationships learned during training.
| Predictive AI | Generative AI |
|---|---|
| Predict whether a transaction may be fraudulent. | Explain why it may have been flagged. |
| Estimate future demand. | Generate a summary of the forecast. |
| Recommend a movie or song. | Generate a description, image, or script. |
| Classify an image. | Create a new image. |
| Predict customer behavior. | Create personalized content. |
These capabilities can exist inside the same product.
What is generative AI?
Generative AI is artificial intelligence designed to create new outputs from patterns learned during training. Those outputs can include text, images, computer code, speech, music, audio, video, and structured information.
Products such as ChatGPT, Gemini, Claude, and Microsoft Copilot are familiar ways people interact with generative AI, but the technology is not limited to conversational assistants. Image generators, coding assistants, video-generation systems, design tools, and other applications can also use generative AI.
Modern AI is also increasingly multimodal, meaning one system can work across more than one kind of information. A product might process text and images together, analyze an uploaded document, or work with audio and video as part of the same task.
Chapter 04 — The same model, a different kind of output
What generative AI changed
- Same object
This is the model from the training sequence. Nothing about it has been swapped out, upgraded, or replaced for this section.
- Small outputs
For years, what came out of AI systems was small: a label, a score, a category, a ranked list. Enormously useful, and almost invisible — you experienced the result, not the system.
- Room to fill
Give the output space room, and a model built for it will fill that space. A paragraph. An image. A waveform. A block of code. The output is where the change happened.
- Both directions
Modern systems increasingly work across formats going in as well as coming out — text, images, audio, documents, video. That is what multimodal means, and it is why one assistant can read a spreadsheet and describe a photograph in the same conversation.
- Two jobs, one machine
Predictive and generative AI are not rival technologies. One system might estimate an 82 percent chance that a customer cancels; another writes the email that tries to keep them. Both are learned patterns being put to different work, and the same product often does both.
Generative AI is a kind of output, not a different kind of machine. A trained model can be built and used to score something, or to produce something.
What is a large language model?
A large language model, usually shortened to LLM, is a type of AI model designed to work with language. LLMs learn complex statistical relationships involving language during training, and those learned relationships can support tasks such as answering questions, summarizing text, translating between languages, extracting information, classifying text, generating code, rewriting content, and participating in conversations.
An LLM is one type of AI model. It is not another word for artificial intelligence. Likewise, products such as ChatGPT, Gemini, Claude, and Copilot are not simply “an LLM.” A product can combine one or more models with a user interface, search, files, memory, tools, safety systems, and other capabilities.
What are AI tokens?
People see words and sentences. Language models process text differently. Before a language model processes text, it is divided into smaller units called tokens. A token may represent an entire word, part of a word, punctuation, or another unit of text.
The sentence:
Artificial intelligence is changing how people find information.
is therefore not necessarily processed as seven clean word-sized units. The text is converted into representations the model can work with mathematically. This is also why you may hear terms such as token limits or context windows when people discuss LLMs.
How does an LLM generate text?
A useful simplified explanation is that a language model generates text one token at a time. If the model receives:
Peanut butter and…
several continuations are possible. “Jelly” may be likely. Other words may also make sense depending on the context. An unrelated word would usually be much less likely. The model selects a continuation according to its generation process and then repeats the calculation with the expanded sequence. One token becomes another, and repeated generation gradually becomes a sentence, paragraph, code sample, or longer response.
This is why LLMs are often described as next-token prediction systems. That description is useful, but incomplete. Modern language models learn complex relationships that support translation, summarization, code generation, classification, planning, and reasoning-like tasks, and comparing the entire system to phone autocomplete misses much of that complexity. Next-token prediction explains an important part of how language is generated. It does not explain everything a modern AI system can do.
Chapter 04, continued — How language actually gets made
Tokens and next-token generation
Tokenisation differs between models. This is a conceptual demonstration, not a literal tokenizer, and the identifiers are illustrative.
- You see words
People read words and sentences. A language model does not start there.
- It sees units
Before anything else happens, text is split into tokens. A token might be a whole word, part of one, a space, or a piece of punctuation, depending on the model and the tokenizer it uses.
- Then numbers
Each token maps to a number. From that point the system is doing arithmetic over sequences of numbers, not reading English. This is also why token limits and context windows are measured the way they are.
- One step forward
Generation works forward from whatever context the model currently has. Given a few words, it estimates which continuations are plausible and takes one.
- Then again
Then it repeats the calculation with the longer sequence. And again. Each individual decision is small and local.
- It compounds
Repeated quickly enough, those small decisions become a sentence, then a paragraph, then a working piece of code. Nothing about the output was planned in advance as a whole.
- Useful, but not the whole story
Next-token prediction explains how the text comes out. It does not explain everything a modern system can do — translating, summarising, extracting, planning, working through a problem. Comparing the whole system to phone autocomplete gets the mechanism right and the capability badly wrong.
Text in, numbers through, one token at a time out. The mechanism is simple and local; what it adds up to is not.
What is a prompt in AI?
A prompt is the information or instruction provided to a generative AI system. A prompt can be simple:
Explain gravity.
It can also provide additional context:
Explain gravity to a 10-year-old using a playground example and avoid mathematical equations.
A prompt can include a question, instructions, background context, examples, formatting requirements, pasted text, uploaded information, or previous conversation content. The output therefore depends not only on what the model learned during training, but also on the information available when the request is processed.
Does AI actually understand what it is saying?
This question gets complicated quickly. Modern AI models can learn sophisticated representations involving language, concepts, relationships, images, code, and other information, and they can display behavior that resembles reasoning, planning, explanation, or problem solving. That does not automatically mean their internal experience is equivalent to human understanding, nor does it establish that an AI system is conscious.
For practical purposes, it is more useful to ask what information the system has, what the model was trained to do, which tasks it performs reliably, where it struggles, and how the output can be verified.
Those questions lead directly to another important part of AI. Where does the information in a response actually come from?
Where an AI system gets its information
Several different sources can contribute to an AI response, which is why it can be misleading to describe everything as something “the AI knows.” At a high level, an AI system can receive useful information from four places:
- Training. Patterns and relationships were learned while the model was being trained.
- Current context. The user provides prompts, conversation history, files, images, or other information for the task.
- Retrieval. The system finds information in external documents, databases, search indexes, or websites.
- Tools and connected services. Outside systems can provide information or perform operations when the AI needs them.
Those sources behave differently.
Information learned during training
The first source is the trained model itself. Training can encode enormous amounts of information about patterns, language, images, code, concepts, and relationships into model parameters. That learned information can help the model answer questions or perform other tasks, but it can also be incomplete, outdated, inconsistent, or difficult to reproduce reliably.
Information you give the system
An AI product can work with information supplied directly by the user. If you upload a report to Gemini, Claude, ChatGPT, or another assistant and ask for a summary, the model does not need to have seen that exact report during training — the document becomes part of the information available for the current task. The same principle applies to pasted text, images, spreadsheets, code, previous conversation messages, and other files supplied to the system.
Information retrieved from somewhere else
An AI system can also retrieve information from an external source. That source could be a database, an internal document collection, a search index, a website, an API, or another connected system. The retrieved information can then become additional context for the model.
Information from tools and connected services
Some AI systems can use tools to perform operations outside the model. Depending on the product, those tools might allow the system to search the web, perform a calculation, run code, inspect a file, work with a calendar, query a database, or interact with other software.
The model can help determine what should happen. The connected tool performs the outside operation.
Chapter 05 — The model is one component
Model, system, and product
- Back out
We have been inside the model for two chapters. Pull the camera back far enough and it is a box again — the same box that came out of training.
- What arrives with the request
Around it, first, the things that show up when you ask something: the prompt you wrote, the conversation up to that point, and any file you attached. None of this was in the model. All of it can shape the answer.
- What it can go and get
Then the things the system can fetch on its own: information retrieved from documents or the web, data from connected business systems, and whatever the product has been allowed to remember about you.
- What it can do, and what it must not
Then tools — a search index, a code runtime, a calendar, a database — and the system instructions and safeguards the developer wrapped around all of it. The model helps decide what should happen. The tool is what actually does it.
- That is the system
Model plus context plus information plus tools. Look at the composition rather than the centre: the model is one component of it, and frequently not the component that changed most recently.
- And then the product
The thing you actually open is a further layer again — an interface, an account, a pricing tier, a set of decisions about what to expose. Chatbot, search engine, coding assistant, or a feature you never see running in the background.
- Why this matters
This is what explains two products built on similar models behaving differently, and one product gaining a genuinely new capability without anyone retraining anything. Developers added a tool, connected a source, changed the instructions, or switched which model sits in the middle.
Model ≠ AI system ≠ product. Most of what a product can do was decided by what got assembled around the model.
Model, AI system, and AI product are different things
These terms are closely related, but they describe different layers.
- The model is the trained mathematical model.
- The AI system can combine that model with prompts, instructions, retrieval, memory, tools, outside information, and other components.
- The product is the experience a person actually uses.
The product might be a chatbot, search engine, coding assistant, recommendation service, enterprise application, or feature running quietly in the background.
This explains why two products using similar underlying models can behave differently. It also explains why a product can gain substantial new capabilities without completely retraining the model. Developers can add tools, connect new information sources, enable search, change system instructions, or switch the model being used.
Why AI can be impressive and still be wrong
Generative AI can produce remarkably polished output. That does not make every output reliable. A response can be fluent, confident, detailed, and wrong at the same time.
What is an AI hallucination?
An AI hallucination occurs when an AI system generates information that sounds plausible but is incorrect, unsupported, misleading, or invented. Examples include inventing a statistic, creating a quotation that was never said, providing the wrong date, describing a study that does not exist, misrepresenting what a source says, producing a citation that does not support the claim, or combining real facts into an incorrect conclusion.
This makes more sense once you remember that a generative model produces an output from learned relationships and available context. It is not necessarily checking a perfect internal database before writing each sentence.
Confidence is not verification
People naturally use confidence as a signal. A hesitant speaker may seem uncertain, while someone providing specific details confidently can appear more trustworthy.
Generated language breaks that shortcut. The style and confidence of an AI response do not necessarily indicate whether the underlying information is reliable. Important information still needs verification.
Chapter 06 — Polish is not evidence
Why a confident answer can still be wrong
- Two answers
Two responses to the same question. Same length, same tone, same clean formatting, same air of knowing what it is talking about. Nothing on the surface separates them, and nothing ever will.
- What is underneath
So draw in what each sentence is resting on. Every specific claim — a number, a date, a study, a quotation — either connects to something or it does not.
- One holds
On the left, each claim lands on something that can be checked: a warranty page, a testing body, a rate card. You may still want to verify it. You can.
- One does not
On the right, the same polish sits on top of a statistic with no source behind it, a study that does not exist, and a real source that says something different from the sentence attached to it. This is what a hallucination looks like from the outside: correct-sounding, specific, and confidently wrong.
- Why the shortcut fails
People use confidence as a signal for reliability, and it is usually a reasonable one. Generated language breaks it, because fluency is what the system produces either way. The style of an answer tells you nothing about the information underneath it.
Fluency is not evidence. A generated answer is produced from learned relationships and whatever context is available — not checked against a perfect internal record of facts before each sentence.
Information can become outdated
Training happens during a period of time. The world continues changing afterward. People change jobs. Prices change. Laws change. Products launch. Companies close. Scientific understanding evolves. News happens. A model relying only on what it learned during training may therefore lack current information.
That does not mean every AI product is permanently limited to old information. The product may be connected to web search, databases, company systems, or other current sources. Once again, the model and the system around it are different things.
Bias and other limitations
AI systems can struggle for several reasons:
- Training data. Missing, inaccurate, or unevenly represented information can influence what a model learns.
- Bias. Historical or social biases can enter through data, labeling, training objectives, system design, or deployment decisions.
- Privacy. Sensitive information may be inappropriate to provide to certain AI systems or services.
- Security. AI products and agents can introduce new ways for systems, accounts, or information to be misused.
- Misinformation. Generative tools can make incorrect or fabricated information easier to produce at scale.
- Overreliance. People can place too much trust in outputs that still require judgment or verification.
- Computational impact. Developing and operating large AI systems can require substantial energy, hardware, water, and data-center infrastructure.
AI is neither universally good nor universally bad. Its value depends on the task, the system, the information available, and how the output is used. For many tasks, the practical approach is to use AI where it is helpful, understand where it can fail, and verify what matters.
One way to give an AI system better access to factual or current information is to let it use information from outside the model.
How modern AI systems use outside information
Suppose you ask an AI system:
What happened in the market today?
A model relying only on information learned during training has an obvious limitation. Today happened after that training. You could paste current market information into your prompt yourself, but that does not scale to every current question someone might ask. Instead, an AI system can retrieve relevant information when it needs it.
What is retrieval?
Retrieval is the process of finding information relevant to a request. A simplified retrieval process looks like this:
- The system receives a question.
- It searches an available collection of information.
- Relevant information is selected.
- That information becomes additional context.
- The model uses the expanded context to produce a response.
The collection being searched might contain a small number of internal documents or billions of webpages.
What is grounding?
Grounding refers broadly to connecting an AI response to external information or evidence that helps support the output. Several things can provide grounding: retrieved documents, uploaded files, company data, databases, web search, and other external information sources.
Grounding does not guarantee correctness. The wrong source can be retrieved. The source itself can contain an error. The system can misinterpret it. The generated response can still introduce a mistake. Grounding improves the information available to the system, but it does not remove the need for judgment.
What is RAG?
RAG stands for retrieval-augmented generation, an approach that combines retrieval with generative AI. A simplified RAG process is:
- A person asks a question.
- The system searches an information collection.
- Relevant information is retrieved.
- The retrieved information becomes part of the model’s context.
- The model generates a response using the question and that additional information.
It is similar to the difference between taking a closed-book test and being allowed to consult relevant material before answering.
RAG is useful, but it should not be treated as a universal explanation for every AI system that uses search or retrieval. Different products can use different architectures, ranking systems, indexes, models, and data sources.
Chapter 07 — The system opens
Retrieval, grounding, and RAG
- A question it cannot answer
Ask what happened in the market today. Training finished before today did, so nothing encoded in the model covers it. You could paste the information in yourself, but that does not scale to every current question anyone might ask.
- So it opens
The alternative is for the system to stop being closed. The boundary opens and it becomes possible to reach information that was never part of training.
- Finding
Retrieval is just this: find what is relevant to the request. It might mean searching a handful of internal documents or an index covering much of the web. Question, then find, then use.
- Grounding
The passages that come back become part of the model’s context for this one request — the same slot your uploaded file occupied two chapters ago. Connecting a response to outside evidence like this is what grounding means. It does not guarantee correctness: the source can be wrong, the wrong source can be picked, and the generated sentence can overstate what the source actually said.
- Open book
Put those steps together and you have retrieval-augmented generation, usually shortened to RAG. Closed book becomes open book. It is a genuinely important pattern, but it is not a universal explanation for every AI system with a search box — products differ in their indexes, ranking, models, and sources.
- Same parts, different builds
Once the pieces are separate, products stop being mysterious. A conversational assistant is mostly a prompt and a model. Add an uploaded file and it is a document assistant. Add retrieval and sources and it is search-connected.
- And an agent
Add tools and a feedback loop and the system can keep taking steps toward a goal rather than returning a single answer. The difference between these products is mostly which of these parts got assembled, not which one has the smartest model.
- Retrieval-augmented generation, in five steps
- 1Question. The request arrives.
- 2Search. An index of sources is searched for what is relevant.
- 3Retrieve. A few passages come back.
- 4Into context. Those passages are put into the model’s context for this request.
- 5Generate. The response is written from the request and those passages.
An AI product is a system assembled from capabilities. Retrieval is one of them — the one that lets a system answer with information that was never in the model.
How different AI products combine these pieces
Modern AI products can combine the same underlying concepts in different ways:
- AI assistant. Combines a prompt, model, and conversational interface to generate a response.
- Document assistant. Adds files or other supplied information to the context available to the model.
- Enterprise assistant. May connect the AI system to internal company information, databases, or software.
- Search-connected AI. Adds web search or another retrieval system so the model can work with external information.
- AI agent. Combines a model with tools and a feedback loop so the system can continue working toward a goal.
Widely used products illustrate different combinations of those capabilities. Gemini can work with uploaded information and search-connected experiences. Microsoft Copilot can combine AI with eligible work information and web results in supported contexts. Perplexity is designed heavily around web retrieval and citations. ChatGPT can combine conversational generation with web search. Claude can work across conversation, files, coding workflows, and connected tools depending on the product.
The exact feature lists will change. The underlying concept is more durable: an AI product combines models with other capabilities to create a particular experience.
An AI agent is worth its own explanation because tools and continued actions introduce a different kind of system. If you want to go deeper, read What Is an AI Agent?
One other combination has become particularly important because it changes how people find information. Search.
What happens when AI meets search?
Search is one of the clearest examples of the concepts in this article coming together. A person asks a question. The system needs to understand what they want. It may need information that does not exist inside the model. Search or retrieval systems find potentially useful sources. Relevant information becomes context. An AI model helps organize or synthesize the result.
The response may combine a generated answer with citations, traditional links, images, products, maps, videos, or other information.
What is AI search?
AI search is a broad term for search experiences that use artificial intelligence to help interpret a request, retrieve relevant information, and present a direct or synthesized response. There is no single universal AI-search architecture.
Google Search integrates generative experiences such as AI Overviews and AI Mode into a traditional search engine. Perplexity approaches search as an answer-oriented product built heavily around web retrieval and citations. ChatGPT Search adds web retrieval to a conversational AI product. Microsoft Copilot can use Bing web search in supported experiences.
These products differ in implementation, but they share a broader pattern: search, retrieval, and AI-generated responses are becoming more closely connected.
How is AI search different from traditional search?
A simplified traditional search experience works like this:
- The user enters a query.
- The search engine returns relevant results.
- The user opens useful pages.
- The user compares the information.
- The user builds their own answer.
An AI-assisted search experience can shift more of that work into the system:
- The user asks a question.
- The system interprets what information may be needed.
- Search or retrieval finds potentially useful sources.
- Relevant information is selected.
- That information becomes context for the AI system.
- The system generates or organizes a response.
- Sources or citations may appear alongside the answer.
Neither model completely replaces the other. Sometimes a generated answer is useful. Sometimes the best result is a website, product page, local business, forum discussion, original study, video, or source document. Modern search increasingly combines both experiences.
One question can represent several information needs
Consider:
What should I look for when buying my first electric car?
That single question can involve several different considerations: driving range, charging, purchase price, ownership costs, battery warranties, reliability, incentives, everyday driving, and long-distance travel.
Some modern AI search systems can explore several related aspects of a question rather than treating the user’s wording as one literal keyword query. That changes the relationship between the exact wording of a question and the information ultimately retrieved.
Chapter 08 — One question, the whole machine
How an AI search answer gets built
The questionWhat should I look for when buying my first electric car?
- One question
Here is a question a person actually asks. It looks like one question. It is not.
- Several needs at once
Inside it: how far it goes, where it charges, what it costs to buy, what it costs to run, how the battery holds up, whether the thing is reliable, what incentives apply, and whether any of this survives contact with a normal week of driving.
- More than one page can hold
A modern search system can work on those related needs rather than treating your wording as a single literal keyword string. Which means it is not looking for one page. It is looking across a lot of them.
- What is out there
Each need points somewhere different. A government schedule for incentives. A manufacturer specification for warranty terms. An editorial long-term test for reliability. A research paper for degradation. A charging network for coverage. A forum thread where owners argue about winter range, which is often where the honest answer lives.
- Most of it is not used
Almost none of what exists gets used. A small number of passages are selected as relevant to this particular request, and note what is being taken: a passage, not a page.
- Into the system
Those passages move into the model's context — the same slot the uploaded file occupied three chapters ago, and the same slot the retrieved market data occupied in the last one. Nothing new has been introduced. The machine is just running at full size.
- The answer forms
The response is assembled from what came back, organised around what was asked rather than around what any single source happened to cover.
- And the way back
Then the citations. They mark which claims rest on which sources, and they run backward — from a sentence in the answer to the page it came from. They do not certify anything. What they do is make part of the chain visible, so that a person can go and check, read more, or disagree.
The answer the system assembles: Range is the number most first-time buyers over-weight: most current models comfortably cover normal daily driving, and cold weather matters more than the rated figure. Charging is the decision that actually shapes ownership — home charging is cheaper and easier, and public fast charging is what long trips depend on. Battery health is covered by a long warranty on most new vehicles. Purchase incentives change frequently and are worth checking against current government sources rather than a review from last year. Each of those claims carries a citation marker pointing back to the source it was drawn from.
A conceptual model of how these systems are put together. It does not reproduce any one platform's architecture or interface.Question → search and retrieval → sources → AI system → answer with citations. Every part of that was introduced earlier in this article; this is the first time they all run at once.
Why citations matter
A citation gives the reader a path back to a source that contributed to an AI-generated response. That can help someone verify a claim, inspect the original information, read additional context, compare perspectives, or continue researching.
Citations do not make an AI response automatically correct. A source can be wrong. The system can misunderstand it. A generated sentence can overstate what the source actually supports. What citations do is make part of the information chain visible.
For someone who owns a website, that creates a different question. So far, we have followed the user’s question toward an answer. What happens if we look at the same system from the other direction?
How information from a website becomes part of an AI answer
This is where the article changes perspective.
I spend much of my time thinking about how websites become visible when people search. AI search changes the interface and adds retrieval and synthesis layers, but the underlying information problem remains familiar. A system still needs useful information.
Instead of asking how an AI system finds an answer, we can ask how information from a website becomes useful enough to potentially contribute to one.
Chapter 09 — The same machine, from the other side
How information from a website reaches an AI answer
A framework for understanding AI visibility, not a universal ranking algorithm.
- Turn around
Everything so far has run in one direction: person, question, system, answer. The rest of this article looks at the same machine from the other end.
- Start at the page
Now the beginning is a webpage somebody published, and the question changes shape. Not how does an AI system find an answer, but what made this information useful enough to end up inside one.
- Discoverable, accessible
Nothing can be retrieved that cannot first be found and fetched. Crawler restrictions, authentication walls, rendering failures, and network protections all settle this before quality is ever considered.
- Understandable, relevant
Then it has to be interpretable, and it has to match. Clear writing, sensible headings, descriptive context, and appropriate structured data help with the first. Nothing helps with the second except genuinely being a good source for the question being asked — a strong page can still be the wrong source.
- Useful, retrievable, usable
Original information, firsthand expertise, supporting evidence, and specific examples give a system a reason to prefer you over a dozen pages repeating each other. Burying the answer two thousand words deep gives it a reason not to. And a source that is right for a definition may be wrong for current pricing or original research.
- Citable
At the end, some systems surface the source beside the answer. That is the same citation marker from the previous chapter, seen from the other direction.
- What it is and is not
This is a way of thinking about the journey between publishing something useful and that information becoming visible inside an AI-powered experience. It is not a ranking algorithm, and no framework can represent the internal process of every platform.
Discoverable · accessible · understandable · relevant · useful · retrievable · usable · citable. Eight conditions, each attached to a stage of the system the previous chapters built.
A useful framework for thinking about AI visibility
No single framework can represent the internal process of every AI platform. Still, a useful way to think about AI visibility is to ask whether information is:
- Discoverable. Can the relevant crawler, index, or system find the page?
- Accessible. Can the system retrieve the important content without technical barriers?
- Understandable. Is the information clearly written and structured so its meaning is easy to interpret?
- Relevant. Does the page actually address the user’s question or information need?
- Useful and credible. Does the source offer reliable information, evidence, expertise, original insight, or something meaningfully different from generic alternatives?
- Retrievable. Can the relevant information or passage be identified when it is needed?
- Usable. Can that information help support the response being created?
- Citable. Can the platform expose the original source to the person receiving the answer?
This is a framework for thinking about AI visibility, not a claim that every AI platform uses these eight items as a ranking algorithm.
What does AI search mean for SEO?
For someone who works in SEO, much of this should sound familiar. Search optimization has always involved helping systems discover, access, interpret, evaluate, and surface useful information. Several fundamentals remain the same:
- Search systems still need to discover and access the page.
- Content still needs to satisfy a real information need.
- Relevance, credibility, technical accessibility, and page experience still matter.
- Internal linking still helps users and systems discover related information.
- Important information still needs to exist in accessible textual form.
- Publishing more content does not automatically make that content more useful.
AI search introduces additional ways that information can surface:
- A page can contribute information to a generated answer in addition to appearing as a traditional search result.
- An individual passage can matter because it answers one part of a larger question.
- Retrieval can occur across several related searches or subtopics.
- Citations can create a direct path from a generated answer back to the original source.
- Follow-up questions can continue the information journey without requiring a completely new search.
That changes how I think about search visibility. The question is no longer only whether a page ranked for a keyword. It is also whether the page contains information that meaningfully answers what people are trying to understand, whether that information can be retrieved, whether it makes sense outside the context of the entire article, and whether the source provides something more valuable than dozens of interchangeable alternatives.
Where does AEO fit?
Answer engine optimization, or AEO, focuses on improving how information can be discovered, understood, retrieved, and used within answer-based and AI-powered search experiences. It does not require abandoning SEO. The two overlap heavily.
Strong SEO and AEO share many fundamentals: technically accessible pages, clear information architecture, useful content, strong topical relevance, credible sourcing, original expertise, self-contained answers, sensible internal linking, and a good experience for the person using the page.
The interface is changing. The need to publish information worth finding is not.
If you want to go deeper into this side of AI search, continue with Answer Engine Optimization (AEO): The Complete Guide.
Where AI goes next
The future of AI is difficult to predict in detail because the products and models are changing quickly. The broader direction is easier to see.
AI systems are becoming better at working across multiple forms of information. Search, chat, software, and other interfaces are becoming less distinct. Models are gaining access to more tools and external information. AI assistants can work with larger amounts of personal, enterprise, and web context when users permit it. Agentic systems are moving from one-step responses toward longer sequences of work.
The boundaries between a search engine, chatbot, assistant, software application, and agent are likely to become less obvious. That is why understanding the concepts in this article is more useful than memorizing the feature list of any individual AI platform.
When you encounter a new AI product, ask:
- What model or models are involved?
- What was the system designed to do?
- What information does it currently have access to?
- What came from training, and what came from the current context?
- Can it retrieve outside information?
- Can it use tools or take actions?
- What is it predicting, generating, retrieving, or deciding?
- How can the result be verified?
Those questions will remain useful even as individual products change.
Chapter 10 — The whole machine, then the part worth keeping
The complete picture
- Where it started
This is the diagram from near the beginning: data trains a model, an input goes in, an output comes out. Everything since has been an expansion of one of those five boxes.
- The model
Training data flows in and changes the model. That has not moved since chapter two, and it is still the only part of this where learning happens.
- The system
Around the model: what you supply, what the product remembers, what it retrieves, what it can act with. Model plus context plus information plus tools.
- The world outside
Beyond the boundary: documents, databases, websites, search indexes, connected services. The system reaches out to them when what it needs is not already inside.
- What comes out
And out of all of it: a prediction, a piece of generated content, an answer with sources beside it, or an action taken through a tool. Different products, different combinations, the same underlying parts.
- Then put it away
Which is the thing actually worth carrying. Not the full map — the habit of asking which part of it you are looking at. What model is involved, what information does it have right now, what came from training and what came from you, can it retrieve, can it act, and how would you check the result.
The whole inventory, in words: training data → training → model. Around the model: prompt, conversation context, uploaded files, memory, retrieved information, tools, and system instructions — together, the AI system. Outside it: documents, databases, websites, search indexes, and connected services. Out of it: a prediction, generated content, an answer with citations, or an action. Wrapped around all of it: a product.
So, what is AI?
Artificial intelligence is a broad field of technology that allows computers to perform tasks associated with intelligence, including recognizing patterns, making predictions, generating content, retrieving information, supporting decisions, and increasingly using tools to take actions.
AI is not one program or one type of model. Machine learning is one major way modern AI systems learn from data. Training changes a model, while inference uses that trained model on new information. Generative AI uses learned patterns to create new outputs. Large language models are one type of AI model designed around language.
The products people use add more around those models. Prompts provide instructions and context. Uploaded files provide information for a specific task. Retrieval can bring in outside knowledge. Tools can allow the system to perform actions. Search can connect the system to information across the web.
Products such as ChatGPT, Gemini, Claude, Microsoft Copilot, Perplexity, and the AI experiences built into search engines combine these capabilities in different ways.
The technology behind all of this can become extremely complicated. The basic mental model does not have to be. Once you understand which part of the system you are looking at, AI stops feeling like one mysterious black box.
Common questions about AI
What is AI in simple terms?
Artificial intelligence is technology that allows computers to perform tasks associated with human intelligence, such as recognizing patterns, making predictions, understanding language, generating content, or making decisions.
Is AI the same thing as ChatGPT?
No. AI is the broad field. ChatGPT is one product built with AI. Gemini, Claude, Copilot, Perplexity, recommendation systems, computer-vision systems, and many other technologies can also use AI.
Is ChatGPT generative AI?
Yes. ChatGPT is a product that uses generative AI models to create responses from prompts and other context. It is one example of generative AI rather than the definition of the category.
What is the difference between AI and machine learning?
Artificial intelligence is the broader field of building systems capable of tasks associated with intelligence. Machine learning is one major approach within AI that allows models to learn useful patterns from data.
What is the difference between training and inference?
Training changes or improves an AI model using data. Inference is what happens when the trained model processes new information and produces an output.
What is an AI model?
An AI model is the mathematical system produced or adjusted through training. It contains learned relationships that can be applied to new information during inference.
What is generative AI?
Generative AI is artificial intelligence designed to create new outputs such as text, images, audio, video, or computer code based on patterns learned during training.
What is an LLM?
A large language model, or LLM, is a type of AI model designed to process and generate language.
What is RAG in AI?
Retrieval-augmented generation, or RAG, combines retrieval with generative AI. Relevant information is found outside the model, added to its context, and then used while generating a response.
What is AI search?
AI search is a broad term for search experiences that use artificial intelligence to help interpret a question, retrieve relevant information, and provide or organize a more direct answer.
What is the difference between an AI model, an AI system, and an AI product?
The model is the trained mathematical system. The AI system combines that model with prompts, instructions, retrieval, memory, tools, and outside information. The product is the experience a person actually uses. This is why two products built on similar models can behave differently, and why a product can gain new capabilities without retraining anything.