What is AI SEO?

On 14 September 2026 we asked one AI engine eight ordinary marketing questions and then checked every source it named. For six of the eight it never searched the web at all. Of the fifty-three URLs it listed, twenty-one return a 404. The engine was not lying on purpose. It was answering from memory, the way a confident person quotes a book they read years ago and misremembers the page.

That is the thing worth understanding before spending a pound on AI SEO. Sometimes the engine looks, and what you publish can enter the answer this month. Sometimes it does not look, and nothing you publish this year will move it. Knowing which of the two you are facing is most of the strategy.

Below: what AI SEO is, how an answer gets assembled, where generative engine optimization and answer engine optimization fit, what makes a brand retrievable, how to measure any of it, and the parts I would skip.

Key takeaways

AI SEO has two meanings. Using AI tools to do SEO faster is a workflow question. Getting into AI answers is the one that changes a pipeline.

Engines do not always retrieve. In our 14 September 2026 study, live search fired on two of eight questions. On fresh, contested topics a new page can enter the answer within a cycle. On evergreen topics it cannot, at any quality.

Verify anything an engine cites you. Twenty-one of fifty-three listed sources in that study did not exist.

“AI SEO” is the biggest term in this cluster at 8,100 US searches a month, ahead of AI marketing at 5,400, generative engine optimization at 4,400 and answer engine optimization at 2,400 (Ubersuggest, 13 September 2026).

third party sources do more work than your own site copy. An engine counts how many independent places describe you the same way.

What AI SEO is, and how it differs from search engine optimisation

Search engine optimisation competes for a position on a list of ten links. AI SEO competes to be the sentence the model writes. The buyer sees one paragraph and two names, and the other eight are not seen at all.

The phrase itself carries two jobs, and people mix them constantly. The first is using AI tools inside the marketing team: clustering keywords, drafting, generating schema, mapping internal links. That saves hours and changes nothing about what a site deserves to be returned for. The second is optimising a site and a company’s public record so that assistants and AI search features surface it. Both get called AI SEO. Only the second one decides whether a buyer hears your name.

Ranking eleventh used to cost traffic. Being missing from an answer costs the meeting. A company described three different ways across its own site, its listings and its press coverage is not pushed down a ranking by a model. It gets averaged into something vague, or dropped. That is the problem a retrievable brand is built to solve, and it is a writing and record-keeping problem far more than a technical one.

The overlap with old-fashioned SEO is real at the base. Crawlable pages, clean site structure, legitimate coverage elsewhere. It separates above that. SEO is tuned to a query and a click. AI SEO is tuned to a category and a description.

How an AI engine builds an answer

First the retrieval

Before anything is written, the engine decides whether to search. If it does, it runs a handful of queries, pulls pages, and writes the answer with those pages in front of it. If it does not, it answers from what it absorbed in training, which is a compressed, undated memory of the web as it was when the model was built.

You can tell the difference from the outside. A grounded answer carries live citations. An ungrounded one carries confident source names that may or may not resolve.

Then the writing

Once the material is in hand, the model writes a paragraph. It leans on whatever is easiest to lift: a plain definition, a heading that matches the question, a number with a source beside it, a named framework it can attribute to someone. Vague brand language gives it nothing to grip. This is the least mysterious part of AI SEO and the part most companies skip, because writing a flat definitional sentence about yourself feels beneath a brand team.

The mini-story: an engine that invented twenty-one sources

Here is the study behind the number in the introduction. On 14 September 2026 we put eight questions to Gemini with Google Search grounding, one answer each, then requested every URL it named and recorded the status. Questions covered marketing principles, the history of advertising, neuromarketing, building a B2B brand for AI search, films and songs about advertising, whether a marketing encyclopedia exists, and the term retrievable brand.

Live retrieval fired twice. Both times it was a fresh, contested topic: AI search, and reference sites. The six evergreen questions were answered from memory, and that is where the dead URLs came from. Twenty-one of fifty-three listed sources returned not found or failed.

On the retrievable brand question, all sixteen grounded citations were real pages, and the mix was instructive. Two platform vendors, two trade publications, two software companies that use the word as a product name, seven small agency and consultant blogs, one Medium post. Small, recently published explainers sat beside the vendors. That is the opening. A new page on a contested topic can enter an answer inside an update cycle. The method and the full table are in our citation study, including the limits: one engine, one sample per question.

[VIDEO: a talk or explainer on how AI search assembles answers; Anne to choose]

Generative engine optimization and answer engine optimization

Two named practices sit inside AI SEO. They are used almost interchangeably now, and the work behind them is the same work.

GEO

Generative engine optimization, or GEO, is the practice of writing and structuring content so that generative engines cite it in the answers they produce. The term comes from a paper posted to arXiv in November 2023 by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. You can read the GEO paper on arXiv. It introduced the term and a benchmark for testing which changes to a source make it more likely to be used in a generated answer, and the authors reported visibility gains of up to 40 per cent on that benchmark.

The practical reading: a generative engine is not ranking pages for a click. It is assembling an answer and choosing which sources to lean on. Content that states a claim plainly, attributes it, and puts a quotable sentence next to a source gives the engine something to reuse.

AEO

Answer engine optimization, or AEO, is the practice of preparing content so it is used in a direct answer rather than a list of links. In practice it means answering one question per page, plainly and early.

AEO grew out of the years when search engines started answering questions on the results page itself through featured snippets and knowledge panels. The habit that built, one question and one clear answer per page, transferred straight to AI assistants. The mechanics are dull. Use the question as the heading, in the words people type. Answer it in the first two or three sentences. Keep the answer self-contained so it still makes sense when it is lifted out of the page. Put a date, an author and a source on it.

Where people separate the two terms, AEO is about the direct answer and GEO about the generated one. Search demand tells you which name to use rather than which practice to run. “AI SEO” measures 8,100 US searches a month, generative engine optimization 4,400, answer engine optimization 2,400 (Ubersuggest, 13 September 2026), and the AEO phrasing is the one climbing: “what is answer engine optimization” went from 210 a month in August 2025 to 2,400 by July 2026. We tracked all three lines in how the AI search vocabulary moved.

Retrievability, and the retrievable brand

Retrievability is not a marketing invention. It comes from information retrieval, the research field behind search engines, where Leif Azzopardi and Vishwa Vinay proposed it in 2008 as a measure of how easily a document can be surfaced by a retrieval system. Applied to a company it asks something blunter: when a buyer asks for the category, does the name come back.

A retrievable brand is one that AI assistants and search engines can find, name and describe correctly when a buyer asks a question. It is built from clear definitions, named frameworks, sourced facts and consistent descriptions repeated across every public source a model reads. Brave New, which publishes the Hive, uses retrievable brand as the name of its method, and that page carries the pillars and the measurement in full.

The mini-story: what a brand does inside a head

In 2004, Samuel McClure and colleagues published a study in Neuron, from Read Montague’s lab at Baylor College of Medicine. People drank Coke and Pepsi inside an fMRI scanner, blind and then labelled. Blind, preferences split. With the Coke label visible, stated preference shifted towards Coke and brain activity changed in regions associated with memory and self-image. The Coke and Pepsi study is the most cited demonstration that a brand is a stored pattern in the buyer rather than a property of the product.

I keep coming back to it because AI SEO is the same idea with a second head to fill. The brand has always lived as an association in a person. Now there is a second store of associations, assembled from public text, and a buyer consults it before they consult themselves. What is written about you is what gets stored, and most brand budgets are still aimed at the first head.

This is also why a rebrand costs more than it used to. The old description stays on the web, and it keeps being counted. Models do not retire your previous positioning because you did.

The machine-readable layer

Schema markup tells a machine what a page is rather than making it guess. Schema.org has the vocabulary. For AI SEO the useful types are Organization, Article, FAQPage, and DefinedTerm for anything you are defining. Put the same company description inside the Organization markup that you use on the About page, word for word.

llms.txt is a plain text file at the root of a site that gives AI systems a short, curated summary and links to the pages worth reading. Jeremy Howard of Answer.AI proposed it in 2024. It is a convention, not a standard, and no provider is obliged to read it. It takes about an hour. The honest benefit is a side effect: writing one forces a company to say plainly what it does and which pages matter.

The third piece is the one people ignore because it is boring. One entity statement, a sentence or two saying what the company is, who it serves and where, repeated without variation on the site, in the schema, in every directory, every profile, every author bio, every partner page. Marketers resist this because repeating the same sentence feels lazy. It is the opposite. Variation is the part the writer enjoys and the part the machine cannot use, which is an uncomfortable trade if you came up through copywriting. There is a whole entry on why brand consistency matters for AI search if you want the argument in full.

How to measure AI SEO

By asking the engines and counting. There is no dashboard handed down from the providers, and anyone selling certainty about the weighting is guessing.

Write the ten questions a buyer would actually type. Not your keywords. The awkward, specific phrasing a procurement lead uses at eleven at night. Ask each engine you care about, and record three things per answer: whether you are named, where in the answer you appear, and whether the description matches what you say about yourself. That is the baseline and it takes an afternoon.

Take the same count against named competitors, in the same prompts, on the same day. A number with nothing beside it means nothing. Then log every source the answers cite, because those pages are the ones shaping your category, and being described correctly on them moves an answer more than another post on your own blog. How models choose which brands to recommend goes further into that mechanism.

Repeat quarterly. Answers move as the sources move, and one reading tells you nothing about direction. And verify every citation before you repeat it, for the reason the study above made plain.

What not to do

Do not pay for volume. Publishing forty thin posts a month to feed the machine produces forty pages nothing wants to quote. The pages that got cited in our study were definition-first explainers with question headings and a visible date.

Do not buy a tool before you have a baseline. The afternoon of manual checks tells you more than a subscription, and it tells you whether the engines are even retrieving on your topics.

And I would not chase the term. AI SEO, GEO, AEO, LLM optimisation: the names have churned every few months while the work underneath them has not changed. Pick one category you intend to be returned for and defend it on every surface, which is positioning with a new set of readers.

The mini-story: the first banner ad

On 27 October 1994, HotWired ran a banner for AT&T. It asked whether the reader had ever clicked their mouse right here, and told them they would. Click-through rates for that first campaign are widely reported as far above anything the format managed afterwards.

The first banner ad is worth remembering for the shape of what followed. A new channel opens, the first movers get results that look like a law of nature, the numbers decay as everyone arrives, and what survives is whatever was a genuine asset rather than a trick of novelty. AI answers are somewhere in the early part of that curve. The tricks that work this quarter will be priced out by next year. A clear definition of what your company is, sourced and repeated, will still be doing work in five.

Frequently asked questions

Is AI SEO the same as SEO?

They share a base and separate at the goal. Both need pages a machine can reach and read, a clean structure and real coverage elsewhere. SEO competes for a position on a list of links, while AI SEO competes to be named inside a written answer.

How do I get my company mentioned in ChatGPT?

Publish a plain definition of what you do, keep that description identical everywhere it appears, and earn coverage on the sources the engines already cite for your category. Then check the answers monthly and record whether you are named and whether the description is right.

Does llms.txt actually work?

It is a proposed convention rather than a standard, and no model provider is required to read it. It costs about an hour, so the downside is small. The reliable benefit is that writing one forces you to state plainly what your site is and which pages matter.

How long does AI SEO take to show results?

It depends entirely on whether engines retrieve live pages for your topic. On fresh, contested subjects a new page can appear in answers within an update cycle. On settled evergreen topics the engine answers from training data, and nothing you publish moves it in the near term.

Can I pay to appear in AI answers?

Not directly, in the way you buy a search ad. What you can pay for is the work underneath: research worth citing, coverage on sources the engines already use, and a site clean enough to read. Anyone promising guaranteed placement in an assistant’s answer is selling something they do not control.

The short version

AI SEO is not a new discipline so much as an old one with a different reader. The reader is a machine assembling a paragraph, and it rewards what careful human readers have always rewarded: a plain statement of what you are, made early, made once, repeated without variation, with a source next to every number.

Start with the count. Ten questions a buyer would type, run against the engines your buyers use, recorded against named competitors on the same day. You will learn in one afternoon whether the engines retrieve on your topics at all, which decides whether publishing is a strategy or a waste. Then fix the entity statement everywhere it appears, which is unglamorous and finite work that most companies have never done.

Everything else is sequencing. The vocabulary will change again. Two years ago none of these terms existed, and the study that produced the numbers in this piece will need repeating on other engines before I would call any of it settled.

If you run the company rather than the marketing

AI SEO is the broadest term in this section and the one most likely to arrive as a proposal with a number attached. Three questions keep that conversation honest.

What do the engines say about us now, with today’s date on it. If nobody recorded the starting point, nobody can show you movement later, and you will be shown activity instead. This costs nothing and takes ten minutes.

Does the engine look things up for our buyers’ questions, or answer from memory. These need different work on different timescales. A plan built on page rewrites is addressed to the first case only, and in our own testing the engine retrieved for two questions out of eight.

Is our problem absence, confusion, or an out-of-date description. They look the same from a dashboard and they are not the same job. Confusion is the most urgent, because a confident wrong answer is being repeated to everyone who asks, and it is the one nobody notices.

What I would not do is treat this as a new discipline requiring a new budget line before those three answers exist. Most of what moves the needle here is the same work that made companies well known before any of this: being described accurately, in the same words, by people who are not you.