To improve brand visibility in AI search engines, your pages have to clear three gates, in this order.
- Retrievable: the assistant's crawler can reach the page at all.
- Quotable: one passage answers a question without needing the rest of the page around it.
- Citable: the answer names you and links to the page.
Almost all of the effort in this space goes to the third gate. Almost all of the failures happen at the first one. A blanket Disallow added a couple of years ago to keep training crawlers out is still sitting in a lot of robots.txt files, quietly doing exactly what it was asked to do.
An answer engine also behaves nothing like a results page. Classic search hands you ten links and lets you pick. An assistant retrieves a handful of passages and writes one answer over them, with two or three citations attached. Ten positions collapse into roughly three, and the runners-up get nothing.
The sections below start at the access layer, because nothing downstream matters if the crawler never arrives.
Why LLM SEO is a retrieval problem
LLM SEO is the work of getting your pages into the small pool of sources a model retrieves before it writes anything. The ranking happens inside a retrieval step you never see, and it scores passages.
The shape of it is worth knowing in detail. The assistant rewrites the user's question into one or more search queries and fetches candidate documents. Those get split into chunks, meaning passages of a few hundred words, and the best-scoring chunks go into the prompt as grounding. The model writes prose over that grounding and cites whatever it leaned on.
Two consequences fall out of that. The unit of competition is the passage, so a strong article whose useful part sits in paragraph forty loses to an average one that answers in paragraph two. And each chunk is scored alone, stripped of your navigation and the context you spent three sections building.
Google is unusually direct about the input side. Its documentation on AI features says there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that a page must be indexed and eligible to be shown in Search with a snippet. No AI text file to publish. No special schema to invent.
That is reassuring and mildly deflating in the same breath. The lever is the lever it has always been: pages a crawler can read, containing text that says something.
What strategies improve brand visibility in AI search engines
Start with the crawlers, because each vendor runs several of them and they do different jobs. Blocking the training crawler and blocking the search crawler are separate decisions with separate bills.
| Crawler | Operator | What it does | Cost of blocking it |
|---|---|---|---|
| GPTBot | OpenAI | Collects content for training foundation models | No effect on ChatGPT search results |
| OAI-SearchBot | OpenAI | Powers ChatGPT's search features | Your site stops appearing in ChatGPT search answers |
| ChatGPT-User | OpenAI | Fetches pages when a user asks it to | Robots rules may not apply, since a person initiated it |
| ClaudeBot | Anthropic | Collects web content for model utility and safety | No effect on Claude's search results |
| Claude-SearchBot | Anthropic | Improves search result quality for users | Your site stops feeding Claude's search answers |
OpenAI documents the split plainly. Disallowing GPTBot signals that your content should not be used to train generative models and does not affect ChatGPT search. Disallowing OAI-SearchBot prevents your site from appearing in ChatGPT search answers, though it may still surface as a navigational link. Anthropic separates its own bots the same way.
So the first real strategy is a robots.txt audit that somebody reads out loud, line by line. Refusing training while accepting search is a legitimate position, and the file has to state it one user-agent at a time. A single wildcard block makes that decision for every vendor at once, usually years ago, usually by someone who has since changed jobs.
The second strategy is being the kind of source these systems reach for. Retrieval favours pages that are indexed and that render their text without a JavaScript step. Put the brand name next to the claim, inside the same passage.
How to optimize content for AI search engines
Write so that any single passage survives being read alone. That is the entire discipline in one sentence, and it is harder than it sounds.
What that means in practice:
- Answer in the first two sentences under a heading, then expand. The expansion is for humans and the answer is for the retriever, and both get what they came for.
- Name the subject in every section. "It" and "this approach" turn into noise the moment a chunk is separated from its neighbours.
- Put genuine sets in a table. Tables chunk cleanly and tend to get quoted intact.
- Date the page, and keep the date honest.
- Keep structured data aligned with what a reader can see. Google's guidance is explicit that you should not add markup about information that is not visible to the user, even when the information is accurate.
Publish specifics while you are at it. A model has no way to verify that you are a leading provider of anything, and adjectives like that get paraphrased into the background. It can quote a latency number or a version.
One caution that deserves straight delivery. Do not put claims in structured data that the page does not support, and do not fabricate figures to make a passage more quotable. Search systems treat mismatched markup as a spam signal, and a fabricated number that gets quoted back to a customer is a problem no ranking gain covers.
Build a prompt set for AI discovery
You cannot watch AI search the way you watch rank. There is no position to check, so you have to ask the assistants directly and count what comes back.
Build a fixed prompt set: thirty to fifty questions your buyers genuinely ask, written the way a person would type them. Run the set against each assistant on a schedule, and record two things per answer. Whether the brand was mentioned, and whether it was linked.
Treat it as an evaluation suite. The prompts stay frozen between runs, because editing them invalidates every comparison you have. If you want to know why a specific phrasing changes the answer so much, the mechanics are covered in our piece on cause and effect AI prompts, and AI Prompt Engineering covers building prompt sets that hold their shape over time.
Expect variance. Personalisation and retrieval freshness both move the answer between runs, so a single query result is an anecdote. Track the mention rate across the whole set instead. Your brand appearing once in a lucky run is the AI search equivalent of refreshing until you like the number.
AI search visibility metrics and KPIs worth tracking
The useful KPIs measure presence in answers. Here is the set worth standing up first.
| Metric | What it measures | Where it comes from |
|---|---|---|
| Mention rate | Share of prompts where the brand is named at all | Your prompt set |
| Citation rate | Share of prompts where your domain is linked | Your prompt set |
| Share of answer | How much of the answer traces back to your page | Manual scoring of a sample |
| Search crawler hits | Whether OAI-SearchBot and Claude-SearchBot are reaching you | Server logs, by user agent |
| Assistant referrals | Sessions arriving from assistant hosts | Analytics, referrer report |
Crawler hits are the metric people skip and then regret skipping. It is the only one that separates "we are invisible" from "we are unreachable", and the fix is completely different in each case.
Referral traffic understates everything else, by design. The user got their answer inside the chat window, so a brand can be cited constantly and see almost no clicks. Judge the work by mention and citation rates, and treat referrals as a bonus.
What AI-powered SEO agents can do for you
AI-powered SEO agents are tools that run the loop above without a human driving each step. The honest summary is that they are good at the measurement half and uneven at the writing half.
The measurement half suits automation well. Running fifty prompts across four assistants every week and diffing this run against the last one: that is repetitive, well-specified work with a checkable output. If you are evaluating vendors here, our map of agentic AI companies covers the layers they sell at, and Agentic AI Engineering covers what the loop looks like underneath.
The writing half is where it gets expensive. Generating a hundred pages a month to blanket a topic produces exactly the kind of thin, near-duplicate content that search systems have spent two decades learning to demote, and answer engines inherited those defences along with the index. The agent will cheerfully keep publishing, since nothing in its loop tells it that page ninety-four said the same thing as page twelve.
Use the agents to see. Keep a human on what gets published.
If you do one thing this quarter, make it the audit: read your robots.txt aloud, then check your server logs to see whether the search crawlers actually came. Everything else in this article is an optimisation on top of a door that has to be open first.