AI search optimization: a plain guide to being found in AI answers

What AI search optimization is

AI search optimization means making sure AI assistants can find you, and choose to name you, when a buyer asks them a question. It is SEO for answers rather than for lists of links.

A growing share of questions now gets a written answer instead of a page of links. Google puts an AI Overview above its results. ChatGPT, Perplexity and Gemini answer in full sentences and link a few sources underneath. In each case a model reads some pages and writes a summary. AI search optimization is about being one of the pages it reads, and one of the brands it names.

This guide covers the whole job in order: where these answers come from, what has been shown to change them, what to do in the first month, and how to tell whether it worked.

What AI search optimization is also called

It has at least four names, and they mean nearly the same thing. Google’s own position is that, for Google, it is simply SEO.

• AEO, answer engine optimization, stresses that assistants give answers rather than rankings.
• GEO, generative engine optimization, is also the title of a 2023 research paper on the subject.
• LLM SEO stresses the large language model that writes the answer, which our LLM SEO guide explains.
• AI SEO is the loosest name, used for almost any work aimed at AI answers.

Google’s guide to optimizing for generative AI features mentions AEO and GEO by name. It says “optimizing for generative AI search is optimizing for the search experience, and thus still SEO”. Pick the name your team already uses. The work underneath does not change, and our page on AEO, GEO and AI SEO traces where each name came from.

Where AI search happens

Each assistant runs its own search, or borrows one, and each documents a robot you must let in. By the providers’ own accounts, a site that blocks one of these robots is left out of, or harder to find in, that assistant’s search, whatever its content says.

The robot names come from each provider: OpenAI, Perplexity, Anthropic and Google’s AI features page. Anthropic, for example, says blocking Claude-SearchBot “may reduce your site’s visibility and accuracy in user search results”. Our engines page lists the robots for ChatGPT, Perplexity, Gemini and DeepSeek on one page.

How an AI answer gets built

The assistant decides whether to search, runs one or more searches, reads some of the pages it finds, and writes an answer from them. Your page has to survive every one of those steps.

1. A person asks a question, often a long one with follow-ups.
2. The model decides whether searching would help. Google’s Gemini API page says the model “determines if a Google Search can improve the answer”.
3. It writes one or more searches. Google calls the several smaller searches behind one question query fan-out.
4. The search returns pages from that assistant’s index. A page its robot never read cannot be in this list.
5. The model reads some of those pages and writes an answer from what they say.
6. It links some of the pages it used. Others it read but did not show, and some answers show no links at all.

Steps two and three come from Google’s page on Grounding with Google Search and its generative AI guide. Google gives a fan-out example: a question about a lawn full of weeds becomes searches about herbicides, weeding without chemicals and preventing weeds. That is why a page can be cited for a question it never ranked for. It ranked for one of the smaller ones.

What changes the outcome, and how sure we are

A few actions are documented by the providers or measured in research. Many others are popular but untested. Know which is which before you spend money.

The GEO paper, by Aggarwal and others, found that adding citations, quotations and statistics raised a page’s share of the answer by 30 to 40% in relative terms on its main measure. Keyword stuffing gave “little to no improvement”. The Chen preprint found that AI search showed “a systematic and overwhelming bias towards Earned media”, meaning third-party sources, over a brand’s own pages. Together they say: make your page checkable, and get talked about elsewhere.

What to put on the page

Google’s guide and the GEO paper point in the same direction for the page itself. Neither gives a template, so the list below is a summary of what each source says, not a formula.

• Answer the question in the first lines of the page or section, in a plain sentence a model could quote whole.
• Organise the page into sections with clear headings. Google’s guide says people appreciate pages “organized by paragraphs and sections”.
• Give each claim something checkable behind it: a source, a number, a date or a named example. This is what the GEO rewrites added.
• Say what only you know. Google contrasts first-hand experience with content that “simply restates information already available elsewhere”.
• Add real images or video where they help. Google says its AI features can bring those into answers too.
• Keep one good page per question rather than many thin variants. Google says pages made mainly to catch variations can breach its scaled content abuse policy.

None of this needs special markup or a new file. Google’s guide says structured data is not required for its AI answers, though it still helps with rich results in ordinary search.

Access first, then the page, then what others say about you. Skipping the first makes the other two worthless.

A first month of AI search optimization

You can do the groundwork in four weeks without buying anything. Each week has one goal.

1. Week one: check access. Read robots.txt and your firewall rules for Googlebot, OAI-SearchBot, PerplexityBot and Claude-SearchBot, and fix anything that blocks them.
2. Week one, also: decide about the training robots on purpose, and write the reason next to each rule.
3. Week two: write down 20 to 50 questions a buyer asks before choosing a product like yours, in the buyer’s words. Our guide to finding buyer questions shows where to look.
4. Week two, also: ask each question in two or three assistants and save every answer whole, with its links and the date.
5. Week three: for each question you lost, open the cited pages and list what they say that your site does not.
6. Week three, also: rewrite or add one page for the most valuable lost question. Put the answer near the top and back each claim with a source or a number.
7. Week four: contact the authors of third-party pages that were cited, where you belong there, with the facts they need.
8. Week four, also: ask the same questions again and compare with week two, question by question.

Do not expect week four to show a result. A rewritten page has to be read again by each robot, and then chosen. Treat the first month as building the list and the habit. The comparison that matters comes after several more rounds.

How to measure AI search optimization

Count how often you are named in answers to questions that do not contain your name. Traffic is a poor measure on its own, because AI answers send fewer clicks.

A Pew Research Center study of 900 US adults’ browsing in March 2025 found people clicked a traditional result in 8% of Google visits that showed an AI summary, against 15% of visits without one. They clicked a link inside the summary in 1% of those visits. That is Google only, and one month, but it shows that a brand can gain ground in answers while its clicks fall.

For Google’s own AI answers, Search Console now has a generative AI performance report. Google says it shows how often your URLs appeared in AI features in Search and Discover, which pages, and by country, device and date. For other assistants, there is no report from the provider, so you have to ask the questions and read the answers.

When you count, keep two numbers apart. Questions that name your brand will usually name you back, and they tell you nothing about a buyer who has never heard of you. Report those separately from the questions that do not name you, and always say how many answers a figure is out of. Our guide to reporting AI visibility honestly explains why, with a report you can copy.

Agentcites is one way to run this every week. Today it asks ChatGPT, Perplexity, Gemini and DeepSeek up to 50 buyer questions for each website, stores every answer word for word with its links, keeps branded and non-branded questions apart, and drafts a page for a question you lost. It costs $149 a month per website, as the pricing page shows.

What AI search optimization cannot do

It cannot guarantee a mention. No provider publishes how its model chooses, and the same question can get a different answer on a different day.

Google’s guide says no third-party tool “has access to our internal ranking or AI systems”. That is a fair warning about every vendor in this field. A tool can ask questions and record answers, which is useful. It cannot see inside the model, and a claim that it can is the first sign to look elsewhere.

It also cannot fix a product that buyers do not rate. If the reviews, forums and comparisons an assistant reads all prefer a rival for good reasons, a better page of your own will not outweigh them. In that case the answer is telling you something about the product, and that is worth knowing too.