AI search optimization, surface by surface

Most advice in this category treats AI search as one thing. It is five surfaces reaching the web by different routes, and the thing that gets you into one of them is not the thing that gets you into the next.

AI search optimization is making a site retrievable, readable and quotable by systems that answer with generated text rather than a list of links. It shares a foundation with classic SEO and diverges above it: there is no position to hold, each engine reaches the web differently, and the same page can succeed in one and be invisible in another.

By Honeydmaxx, Content Writer

Published 13 August 2026

The five surfaces, and what actually decides each one

The blocking rates are from our own crawl of 1048 of the most visited websites in August 2026, with the method and the CSV published.

Google AI Overviews

Googlebot, blocked on 1.5%

How it reaches the web: The ordinary Google index, via Googlebot. No separate crawler, no separate index.
What decides whether you appear: Being indexed and allowed to show a snippet. Google documents no requirement beyond that, and says explicitly that no AI text file or special schema is needed.
The thing people get wrong: Google-Extended does not control this surface. It is blocked on 19.3% of the sites we crawled, and every one of those that did it to stay out of AI Overviews achieved nothing.

How AI Overviews work

Google AI Mode

Googlebot, as above

How it reaches the web: The same index, through a separate conversational surface with its own models.
What decides whether you appear: The same eligibility floor, then how well a page answers one narrow question. Google states the two features may use different models, so the links shown will differ.
The thing people get wrong: It is not reported separately in Search Console, so anything you are told about your AI Mode performance specifically is modelled rather than measured.

AI Overviews against AI Mode

ChatGPT search

OAI-SearchBot, blocked on 10%

How it reaches the web: The Bing index, plus its own crawlers for retrieval and user fetches.
What decides whether you appear: Bing coverage, which is the single most common surprise in a first report. A site that ranks respectably on Google and thinly on Bing can be strong in classic search and near invisible in the most used assistant there is.
The thing people get wrong: OAI-SearchBot is the crawler that decides whether you can be cited, and it is blocked on 10% of sites. GPTBot, blocked far more often, is training and does not affect this.

How to rank on ChatGPT, in full

Claude

Claude-SearchBot, blocked on 10.9%

How it reaches the web: Live fetches at the moment of answering, with sources cited explicitly.
What decides whether you appear: Whether the page can be fetched and read right now. Because it retrieves live and shows its working, it is the clearest read available on whether your own pages are reachable and quotable.
The thing people get wrong: Claude-SearchBot is the retrieval side and is blocked on 10.9% of sites. ClaudeBot, blocked on roughly twice as many, is training.

Check what Claude returns for you

Perplexity

PerplexityBot, blocked on 16%

How it reaches the web: Its own index and live retrieval, weighted heavily towards third party sources.
What decides whether you appear: Corroboration more than your own pages. Directory listings, reviews, comparisons and press show up here first and earliest, which makes it the cheapest place to see whether that work is landing.
The thing people get wrong: PerplexityBot is blocked on 16% of sites, the most blocked of the search and citation crawlers, and blocking it removes you from the assistant that shows its sources most plainly.

Answer engines and the zero click problem

The pattern across all five rows is the same and it is the most useful thing on this page: the crawler that trains a model and the crawler that retrieves for an answer are different agents, and blocking the first does nothing about the second while blocking the second removes you from the answer. The AI crawler checker lists them separately for exactly this reason, and the study has the rates for all thirteen.

What all five have in common, which is most of the work

Five surfaces is not five projects. The base below serves every one of them, and it is also the base classic search rewards, which is why doing AI search and SEO as separate engagements is the expensive mistake.

The crawler has to be allowed in

Per bot, not in general. 28.3% of the sites we crawled block at least one AI crawler while 1.5% block Googlebot, and the gap between those two numbers is almost entirely accidental.

The page has to survive without JavaScript

Of the sites in that crawl depending on client side rendering, 27.8% served almost no text to a fetch that ran no script. That is a total failure that looks from outside like a content problem.

The answer has to be liftable

Each section answering one question, in a self contained passage, with every figure next to the thing it describes. An answer spread across four scrolls of narrative survives none of these surfaces.

Somebody else has to describe you

Roughly 85% of the brand mentions we see in AI answers come from pages the brand does not own. This is the half no tool sells and the half that decides most outcomes.

The order to do the work in

Each step invalidates everything below it while it is unfixed, so the order is not a preference. The first two are free and take about a minute between them.

01
Confirm access, per bot
Not "are we blocking crawlers" but "which of these thirteen agents can fetch this page". Run the AI crawler checker. If a search and citation bot is refused, nothing below matters until it is not.
02
Confirm the page exists without JavaScript
The render gap checker shows what comes back before any script runs. An empty shell here makes every content decision downstream irrelevant.
03
Find out where you actually stand
Put your buying questions to the assistants and read which sources they used instead of you. The AI visibility checker does this live, free, across ChatGPT and Claude.
04
Edit the strong pages before writing new ones
Restructuring an existing page that already ranks is the highest return hour available. Writing a new page is the slowest way to test whether you understand the question.
05
Fix the entity and the dates
Structured data that matches the visible page, a date a machine can read, and named authorship. These are what let a model tell how current a claim is and whether the brand mentioned somewhere else is you.
06
Then work on pages you do not own
Listings, reviews, comparisons, press. Slow, unglamorous, and the half that moves the outcome once the mechanical work is done.

How it is measured, and how it is faked

The unit is a rate across a fixed prompt set, per engine, rerun on a cadence. Not a position, because there is no position, and not a single blended score, because blending four engines hides the only finding that tells you what to do next.

Record four things per question per engine: whether you were named, whether you were linked, who was named instead, and which sources the answer drew on. Named and linked are separate outcomes with separate fixes. The competitor named in your place is reliably the line that gets forwarded internally.

Three things to be suspicious of. A screenshot presented as evidence, since generative answers vary between runs and one ask is one sample. A single AI visibility score with no per engine breakdown. And any precise revenue figure attributed to AI search, because assistants pass very little referrer data and traffic that started in an answer usually arrives looking like direct. Anyone quoting that number is modelling, and the honest version of the claim names its assumptions.

Then hold the deterministic half separately, because it is the half that is actually verifiable. Crawler access per bot, render output, schema validity, freshness and attribution signals do not vary between two samples, any third party can check them on your live pages, and they are what you point at when somebody asks whether the work was real.

Questions people ask about AI search optimization

What is AI search optimization?

The practice of making a site retrievable, readable and quotable by the systems that answer questions with generated text rather than a list of links. It shares its foundation with classic SEO, since both need pages that can be crawled, indexed and understood, and diverges above it: there is no position to hold, several engines reach the web by different routes, and the same page can succeed in one and fail in another.

How do I optimize for AI search?

In this order. Confirm each engine's crawler is allowed, since 28.3% of the sites we crawled block at least one. Confirm your pages survive without JavaScript. Then restructure so each section answers one question in a passage that can be lifted whole, with the figure next to the thing it describes. Then earn third party corroboration, because most of what an assistant says about a company comes from pages the company does not own.

Is AI search optimization different from SEO?

It shares the base and diverges above it. Everything that makes a page crawlable, fast and well structured helps both, which is why doing them as separate projects is the expensive mistake. What is genuinely new is per crawler access control, rendering that has to survive without JavaScript, passages written to be lifted rather than read in sequence, and measurement as a rate across a prompt set rather than a ranking position.

Which AI search engine should I optimise for first?

Whichever one your buyers use, and you find that out by checking rather than guessing. If you have no signal at all, start with ChatGPT because it is the most used, and check your Bing coverage first since that is what it searches. The work has a large shared base, so the surfaces are not five separate projects.

Can I pay to appear in AI search results?

No. There is no placement to buy in any of these surfaces and any product promising one is selling something that does not exist. There is eligibility, which is technical and checkable, and there is being the best available source for a question you never see, which is editorial.

How is AI search performance measured?

As a rate across a fixed prompt set, per engine, rerun on a cadence, never as a position. Per engine matters because a blended score across four assistants averages away the one finding worth acting on, which is that a particular engine cannot see you. Hold the mechanical half separately, since crawler access, render output and schema validity can be verified by anyone on your live pages.

Do I need an llms.txt file for AI search?

Not for Google, which documents that no AI text file is required. For the others there is no operator documentation either way. It is cheap enough to be defensible and it is the last thing to do rather than the first: 11 sites in our crawl published one while blocking an AI crawler in the same robots file, which is decoration on a locked door.

Steps one and two, on your own site, free

The audit runs 36 checks on your live pages, including per bot crawler access and what a crawler without JavaScript receives. That is the mechanical half of this page, measured rather than guessed, and the report is yours whether or not we ever speak.

Run the free audit See our AI SEO services