Query. what is ai

Guides20 min read

What is AI?

Updated

Three nested circles with a row of small colored blocks flowing out of the smallest.

Artificial intelligence, or AI, is technology that lets computers do tasks that normally need human judgment. It has been part of everyday products like email and maps for years. Chatbots such as ChatGPT and Gemini are the newest and most visible kind.

Everyday AI

What is artificial intelligence?

Artificial intelligence is the branch of computer science that builds systems to recognize faces, predict outcomes, write text, or rank search results.

Most of the AI in everyday products does one narrow job and never talks to you. Products like ChatGPT, Gemini, Claude, Copilot, and Perplexity recently started putting several techniques behind one conversational screen. That’s why AI now feels like a single thing you can talk to.

What do the main AI terms mean?

The main AI terms name different layers of the same technology, from the whole field down to the app you open. People often use them as if they meant the same thing.

Term Where it fits
Artificial intelligence The whole field
Machine learning The main way AI systems learn from data
Deep learning A kind of machine learning that uses neural networks
Generative AI AI that produces new content
Large language model The generative model inside a chatbot
AI product The app you open, built around a model

What can AI do?

AI can recognize and classify, predict and recommend, generate new content, retrieve and organize information, and use tools to take actions. Most AI work falls into one of those five groups, and a product usually combines a few of them. An assistant like ChatGPT writes text and can search the web. Some assistants can also use a tool for you, such as filling in a form or booking a table.

What AI can do
Artificial intelligence
  • Recognize and classify.Identify objects, patterns, messages, or categories.
  • Predict and recommend.Estimate what may happen or what may be useful next.
  • Generate.Create text, images, audio, video, code, or other output.
  • Retrieve and organize information.Find useful information in documents, databases, or the web.
  • Use tools and take actions.Interact with software when the product has been given that ability.

An AI assistant combines several of these.

AI works in four steps

AI works in four steps that run from training data to the product you use.

  1. Data trains a model.
  2. The trained model takes in something new.
  3. The model produces an output.
  4. A product wraps search, files, memory, and tools around the model.

Steps one and two are how AI learns, and step three is where generative AI and its mistakes come from. Step four matters most if you run a website. The search a product runs for its answers is how your pages get into them.

AI learns from examples through machine learning

AI learns by being shown a large set of examples and adjusting itself until it picks up the patterns in them.

What is machine learning?

Machine learning is a way of building software that learns its patterns from data. Traditional software follows rules a programmer wrote line by line.

In spam filtering, a hand-written rule that flags “free money” stops working when spammers switch to “fr33 m0ney” the same afternoon. A learned filter studies thousands of spam and normal messages, works out what tends to separate them, and applies that to new mail.

What is training data?

Training data is the set of examples a model learns from. It might be text, photos, audio, code, rows of numbers, or examples people have labeled by hand.

A model only knows what its training data showed it. One trained only on well-lit photos of golden retrievers will do a poor job with a black lab in a dark garage.

Where does AI training data come from?

AI training data comes from a mix of sources, and the companies building models don’t publish the full list. The usual sources are licensed datasets, public web content, and data the company already owns. Companies also use examples people wrote or labeled for the purpose, and synthetic data generated to fill gaps. Data collected through a company’s own products can be used too, under whatever policy those products have.

How does AI training work?

AI training is a loop that runs until the model is good enough at its job.

  1. Show the model an example.
  2. Let it guess.
  3. Score the guess and nudge the model’s parameters, the adjustable numbers inside it, a little toward the right answer.
  4. Repeat with the next example, millions of times over.

After enough rounds, a language model can finish a sentence it has never read.

How a model is trained

Watching a model get trained

TrainingTrained model
Training data
OutputDog
Training
Model
New input
Confidence this photo is a dogNo prediction yet52%, barely a guess88%, improving96%, dog

Training data → Training → Model

New input → Model → Output

Confidence values shown here are illustrative.

  1. Examples enter

    Labeled examples go into training. Some are photos of dogs, and some are photos of other things.

  2. The model guesses

    The model makes a prediction. Early on it's barely better than a coin flip.

  3. Parameters adjust

    Each wrong answer shifts the numbers inside the model a little. Over many examples, the guesses get better.

  4. New input

    Training stops, and the trained model labels a photo it has never seen.

What is an AI model?

An AI model is the system that comes out of training. Give it new input and it produces an output based on the patterns it learned.

People often picture a model as a database that stores what it learned. A database keeps records you can look up. A model keeps patterns in its parameters and no copy of the examples. One trained on dog photos can describe a dog and has no photo to show you.

Ask a language model for a store’s return policy and it writes from patterns about what return policies tend to say. Unless a product fetches the store’s page, the answer is a plausible guess.

What is inference?

Inference is a trained model doing its job on new input. Every question you ask an assistant is inference.

Training Inference
Changes the model Uses the trained model
Adjusts the parameters Processes new input
Happens while the model is built or refined Happens every time the model is used
Can take enormous computing resources Produces one output for one request

Your message doesn’t change the model. The same model can answer thousands of people in a row and be no different at the end.

Where do neural networks and deep learning fit?

Deep learning is machine learning done with neural networks. A neural network is layers of connected units that transform information one step at a time. The large language models behind today’s assistants are deep learning models.

AI, machine learning, and deep learningDeep learning sits inside machine learning, and machine learning sits inside artificial intelligence, with a neural network of input, hidden, and output layers at the center.
Artificial intelligence
Machine learning
Deep learning

Input layerHidden layerOutput layer

Generative AI builds new content one piece at a time

Generative AI builds new content one small piece at a time, choosing each piece from patterns it learned in training. This is step three of the chain, the output. For text, the model doing the work is a large language model.

What is generative AI?

Generative AI is AI that makes something new, such as text, images, code, audio, or video. A chatbot drafting an email, an image tool turning a description into a picture, and a coding assistant finishing a function are all generative AI.

What is the difference between predictive AI and generative AI?

Predictive AI estimates an outcome or sorts things into groups. Generative AI produces new content.

Predictive AI Generative AI
Estimates whether a transaction looks fraudulent Explains why it may have been flagged
Recommends a song Writes a new song or a description of one
Classifies an image Creates a new image

Many products do both, such as a search product that estimates which pages match your question and then writes up what it found.

Predictive vs generativeOne new transaction goes to two kinds of AI. Predictive AI returns a verdict, flagged for review. Generative AI writes a sentence explaining why it may have been flagged.
New transaction
Predictive AI
Flagged for review

The output is a verdict.

Generative AI

This transaction was flagged because it is unusually large, was made in a new location, and doesn't match this account's recent activity.

The output is a written explanation.

What is a large language model?

A large language model, or LLM, is an AI model trained to work with language. One model can write, summarize, translate, sort, extract details, write code, and hold a conversation. An LLM is the engine inside products like ChatGPT, Gemini, Claude, and Copilot.

What are AI tokens?

Tokens are the small chunks of text a language model reads and writes. Depending on the model, a token can be a whole word, a piece of a word, a space, or a punctuation mark. “Unbelievable” might be three tokens, and “the” is usually one.

The context window is how much text a model can hold at once, counted in tokens. Everything in the current request counts against it. That includes what you type, attached files, the conversation so far, and any passages a product pulled in from the web. Anything past the limit is invisible to the model. This is one reason AI search hands the model selected passages from a page and skips the rest.

How does an LLM generate text?

A language model writes one token at a time. Given “Peanut butter and,” the likeliest next word is “jelly.” “Banana” or “chocolate” could also fit, depending on what came before. The model picks one, adds it to the text, and repeats until the response is done.

Your phone keyboard’s autocomplete does a rough version of the same thing. A large model has learned far richer patterns, so it can write whole essays.

Next tokenThe phrase is split into tokens, and the model ranks candidates for the next one, with jelly as the most likely continuation of peanut butter and.
  1. Pean
  2. ut
  3. space
  4. butt
  5. er
  6. space
  7. and

Candidates for the next token

  • jelly62%
  • banana14%
  • chocolate9%
  • honey6%
  • the4%

The candidates and percentages here are illustrative. No model or tokenizer produced them.

What is a prompt?

A prompt is whatever you give a generative AI system to respond to. It can be a question, a set of instructions, an example, pasted text, a file, an image, or the whole conversation so far.

The output depends on the trained model and on what’s in the request, and the request is the part you control. “Explain gravity to a 10-year-old using a playground example” gets a better answer than “Explain gravity” because the model has more to work with.

AI is sometimes wrong because it writes from patterns

AI is sometimes wrong because a generative model writes from patterns and from whatever is in the request. Nothing in that process checks a fact before a sentence is written.

What is an AI hallucination?

An AI hallucination is a confident answer with nothing behind it. It might be a statistic that doesn’t exist, a wrong date, a quote nobody said, or a study that was never published. The model is working as designed when this happens. It wrote the likeliest next words, and those words were wrong.

Why can AI sound confident and still be wrong?

AI can sound confident and be wrong because fluency and accuracy come from different places. The fluent writing comes from training. Whether a sentence is true depends on the information the system had when it answered, so each claim needs checking on its own.

Checking two claims in one answer
Question

When did the fictional Northbridge Museum open?

AI response

The Northbridge Museum opened in 1987 and moved into its current building in 2004.

  • Opened in 1987

    SupportedThe source says 1987.

  • Moved in 2004

    MismatchThe source says 2007.

The Northbridge Museum and its dates are made up for this example.

How do you check an AI answer?

To check an AI answer, pick the claim you’d act on and compare it against the source.

  1. Find the claim that matters. That’s usually a date, a number, a quote, or a recommendation.
  2. Open the original or cited source. A citation is the link an AI answer shows to a page it used, and a link on its own proves nothing.
  3. Check that the source says what the answer says, and that it’s current enough for your question.

If there’s no source, find one before you rely on anything important.

You can run the same check on answers about your own business. Ask an assistant about your services, prices, or opening hours and compare what comes back with your site. That shows you what AI systems currently say about you.

Why else does AI get things wrong?

AI also gets things wrong when the information behind an answer is out of date, uneven, or fetched badly.

  • A model’s training data stops at a fixed date. Anything that changed afterward is missing unless the product fetches it.
  • Missing or uneven examples shape what a model learns. Bias can enter through the data, the labeling, or the design of the system.
  • Retrieval can fetch the wrong page, or fetch the right page and misread it.

Out-of-date information and bad retrieval both depend on what a product can fetch. That’s step four of the chain, where your website comes in.

AI gets its information from four places

AI gets its information from training, the current request, retrieved documents, and connected tools. Which of the four a product uses depends on how it was built.

Source What it is
Training Patterns learned while the model was built, up to the date its training data ends
Your context Everything in the current request, including the prompt, the conversation so far, and any files, images, or spreadsheets you added
Retrieved information Pages, documents, or records the system fetched for this request
Tools and connected services Search, calculators, code environments, calendars, and business software the system can call

One answer can draw on several of these. If you ask an assistant whether a clinic near you is open on Saturday, the phrasing comes from training and your location from context. The hours have to come from retrieval or a connected service, because the model has no other way to know them.

Is the model the same as the product?

No, the model is the trained system, the AI system is the model plus everything wired around it, and the product is the app you open.

Memory is one of those wired-around parts. When a product remembers that you like short answers, it has saved a note about you, and the model itself is unchanged.

Model, system, productOne model sits at the center of all three views. It appears on its own, inside an AI system that adds prompt, context, files, memory, retrieval, tools, and instructions, and inside a product that wraps that system in an interface.

Model

Model

On its own, the model holds what it learned in training.

System

AI system
PromptContextFiles
Memory
Model
Retrieval
ToolsInstructions

The system adds context, instructions, retrieval, tools, and memory around the model.

Product

Product
AI system
PromptContextFiles
Memory
Model
Retrieval
ToolsInstructions

The product wraps the system in an app a person can open.

What is retrieval?

Retrieval is the step where a system finds outside information for the current request and hands it to the model.

A question like “What happened in the market today?” can’t be answered from training, because the day hadn’t happened when the model was trained. A search system finds current reporting, picks out the useful passages, and adds them to the context for that one request.

Does a citation in an AI answer mean the model was trained on that page?

No, a citation means the product retrieved the page when you asked. Whether the model saw that page in training is a separate question, and a page can go through one without the other.

Training changed the model months or years before anyone asked the question. Retrieval happens at the moment you ask, when the product fetches a page and hands it to the model. A business that changed its hours last week can appear correctly in an AI answer through retrieval. The model behind that answer may have been trained long before the change.

What is grounding?

Grounding is tying an AI response to outside evidence, such as a retrieved document, an uploaded file, a database, or a web search. A grounded answer can be checked, and it can be wrong when the source is wrong or the model misreads it.

What is RAG?

Retrieval-augmented generation, or RAG, is the pattern where a system finds relevant information first and gives it to the model while the model writes. The model works with that material open in front of it, like a student in an open-book exam.

What is the difference between grounding and RAG?

Grounding is the goal, and RAG is one way to reach it. Some products ground their answers in other ways.

Retrieval before generationA question triggers a search, the search returns three sources, one passage is selected from source 2, and the model writes the answer with that passage in front of it.
  1. Question

    What happened in the market today?

  2. SearchSearch
  3. Sources
    Source 1News reportSource 2Market dataSource 3Company filing
  4. Selected passagePassageFrom source 2
  5. AnswerAnswer

AI search fetches web pages and writes an answer from them

AI search fetches material from the web and builds an answer out of it. The system searches for sources, picks out useful passages, and writes a response, often covering several related questions at once.

AI search is a broad term for search that uses AI to interpret a request, retrieve information, and present a direct or combined response.

Google Search has AI Overviews, the generated summaries above some results, and AI Mode, a conversational view of search. Perplexity is built around web retrieval and citations. ChatGPT can search the web inside a conversation, and Microsoft Copilot does the same in the experiences that support it. Each one runs search, retrieval, and generation together.

AI search differs from traditional search in who puts the answer together. In traditional search you compare the results and build the answer yourself. In AI search the system does more of that work for you.

Step Traditional search AI-assisted search
1 You type a query. You ask a question.
2 You open the useful results. The system runs one or several related searches.
3 You compare the sources. Retrieval picks useful sources and passages.
4 You build your own answer. The system writes or organizes an answer, often with citations.

How does one question become several searches?

Query fan-out is when an AI search system turns one question into several searches behind the scenes. A question like “What should I look for when buying my first electric car?” holds six or seven smaller ones. They cover range, charging, purchase price, battery warranty, reliability, incentives, and what day-to-day ownership is like.

Fan-out widens the set of questions a page can be retrieved for. A page about battery warranties can end up in the answer to a question that never used the word “warranty.” The guide to Answer Engine Optimization covers what fan-out means for a page.

Where the product shows citations, each one gives the reader a path from the answer back to the source. For a publisher, citations are also something you can count.

How one question becomes several searches

How an AI search answer gets built

The question about buying a first electric car expands into several information needs, including range, charging, cost, reliability, and incentives. The system searches six sources of different types and keeps three of them. It adds those passages to the AI system's context and builds an answer. The answer's lines carry numbered citations, 1, 2, and 3, pointing back to the three sources it used.

This is a conceptual illustration, and the sources and answer shown are placeholders.
  1. Question

    A person asks one broad question about buying a first electric car.

  2. Fan out

    The system splits it into related searches about range, charging, cost, reliability, and incentives.

  3. Retrieve

    Each search finds sources, and the system keeps the passages that help answer the question.

  4. Answer

    The selected passages are combined into one response, with citations pointing back to where each part came from.

A webpage enters an AI answer when it can be reached and used

A webpage becomes part of an AI answer when a system can reach it, it matches the question, and it has a passage worth using.

Every product ranks sources its own way, and nobody outside those companies knows the full logic. Five conditions apply across all of them because they follow from how retrieval works.

Can a system reach the page?

A system can reach a page when nothing stops its crawler, the automated program that fetches pages. Crawler blocks, login walls, network blocks, and rendering problems all stop a page before its quality is judged.

Rendering problems are the easiest to miss. When content only appears after a page’s scripts run, a system that skips the scripts gets an empty page. A parts store that loads its prices by script shows that system pages with no prices. The AI agent readiness guide covers access in more detail.

Does the page answer the question?

A page gets used only when it answers the question being asked. A hospital’s knee surgery page is relevant to a question about recovery time only when recovery time is on the page. Publishing more pages doesn’t make any one of them relevant to more questions.

Is the writing clear?

Clear writing uses descriptive headings and direct explanations, so people and systems can tell what a passage is about.

Is the content worth using?

Content is worth using when it holds original information, firsthand experience, or clear evidence. Those give a system a reason to prefer it over a page that repeats common knowledge. A software company’s own documentation is the original source for that product’s limits. A page summarizing that documentation gives a system nothing it couldn’t get from the original.

Can the useful passage be retrieved?

A useful passage can be retrieved when it sits near the top of the page, in plain text, and makes sense on its own. AI search often works with selected passages and skips the rest of the page. An answer buried under three paragraphs of setup is hard for a system to pick out. So is an answer shown only in an animation with no text version.

“It is included in every plan” means nothing once it’s lifted out of the page. The same sentence with the feature’s name in it can be quoted as it stands.

Where the product shows citations, the source also has to be a page a reader can open.

How a page reaches an AI answerA webpage has to be accessible and hold a relevant passage before retrieval can hand it to the AI system, the answer uses it, and a citation sends the reader back to the page.

The page exists and publishes something worth using.

Crawl rules, login walls, rendering problems, or network blocks stop a page here.

The page holds a passage that answers the question on its own.

The passage is found and made available to the AI system for that request.

The retrieved information is used while the response is generated.

A citation links the reader back to the page.

What should you check on your site first?

Start with one page that matters to your business and run four checks on it.

  1. Load the page without scripts. Turn off JavaScript in the browser, or view the page source, and confirm the main content is there as text.
  2. Read your crawl rules. Open your site’s robots.txt file, the file that tells automated programs what they may fetch. Confirm it doesn’t block the systems you want to be found by.
  3. Find the answer. Name the question the page should answer, then check that a heading asks it and the first sentence under it answers it.
  4. Lift the passage out. Copy that answer into a blank document and see whether it makes sense with nothing around it.

None of these checks needs a paid tool, and they work across a larger set of pages too. I applied the same approach to 35 inventory pages for a used car retailer. Each page put the answer to a shopper question near the top and the facts in plain text. Of the 35 pages, 33 earned new AI citations within five months. Most had been live for less than two months when measurement closed.

The automotive inventory pages case study has the method and the figures. The AI search practice applies the same approach across enterprise accounts, and my expertise page lists what that work covers.

What does AI search mean for SEO and AEO?

AI search keeps most of search engine optimization, or SEO, the work of getting a site found in search. Pages have to be found and let in, and they have to be useful once a person arrives. Relevance, credibility, clear structure, and links between related pages count in both kinds of search.

AI search adds a second route for the same information to reach people, through passages quoted in generated answers.

Answer Engine Optimization, or AEO, is the work of getting content retrieved and cited in AI-generated answers. AEO builds on SEO and adds attention to answers stated in plain words, passages that make sense on their own, and information no other page has. The AEO guide covers what the evidence says about citations and how to measure them.

Some AI systems go beyond answers and act as agents

An AI agent is an AI system that uses tools, works with files, and carries a task through several steps on your behalf. Agents add a step after retrieval and act on what they read.

What an agent can do varies a lot by product. To a website, an agent is a visitor that may read a page and use its forms and buttons for the person it works for. The guide to AI agents covers how they work and what they need from a page.

The four-step chain explains how AI works

The four-step chain explains how AI works, and a website enters it at the last step, through retrieval. When a new AI product appears or an AI answer surprises you, four questions trace what happened back to the chain.

  • What came from training, and what came from the current request?
  • Can the product retrieve outside information, and from where?
  • Can it use tools or take actions?
  • How can the result be checked?
The full chain from training data to output
Training
Training dataTraining
Model
Outside the system
DocumentsDatabasesWebConnected services
Retrieval + Tools
AI system
PromptContextRetrievalTools
Model
Output
Which can be
PredictionContentAnswerAction

Common questions about AI

Is AI the same thing as ChatGPT?

No, AI is the field and ChatGPT is one product built with it. Spam filters, recommendation systems, fraud detection, Gemini, Claude, Copilot, and Perplexity all use AI.

What is the difference between AI and machine learning?

AI is the broader field, and machine learning is the main approach inside it. Machine learning lets systems learn patterns from data without being programmed by hand.

Does AI learn from everything you type?

AI usually doesn’t learn from everything you type, because the model isn’t retrained each time you send a prompt. Your conversation does become context for later messages, and some products save memories or use your data under their own policies. Those policies differ by product, so check the privacy and training settings in each one.

Why can two AI assistants give different answers?

Two AI assistants can give different answers because they may use different models, instructions, context, retrieval systems, tools, or sources. Two products built on similar models can behave differently because of everything wired around the model.

Does AI search read my whole page?

AI search often doesn’t read the whole page, because it tends to work with selected passages. The part of the page that answers the question matters more than the length of the page. A passage is easiest to use when it sits under a clear heading and makes sense without the text around it.