How to Get Found in AI Search

A client asked me last month why his competitor kept coming up when he asked ChatGPT for a supplier in his own industry, and he didn't. Same city, similar company, and his site was the better of the two.

The answer took about ten minutes to find and had nothing to do with AI. His competitor ranked on page one for the query. He was on page three. The model had run a search, read the top handful of results, and written its answer out of those.

That is the part most of the advice gets wrong. There is no separate discipline where you optimise for the robot instead of the search engine. Being findable the ordinary way is the entry condition, and almost everything else follows from it.

What "AI search" actually refers to

Four different things get called by the same name, and they behave differently enough that the distinction matters:

Only the last one involves the model knowing anything about you in advance. It is also the one you have least control over and the least reason to chase. The first three are a search engine with a writer attached, which is good news, because they respond to the same work that has always worked and that work can be checked.

Ranking is the prerequisite, not the alternative

If a model runs a search and reads the top results, then not being in the top results means not being read. No amount of AI-specific tuning routes around that. Indexing is the stricter version of the same gate. Seven weeks after launch, Google had indexed 22 of our 77 pages and hadn't yet crawled the other 48, and a page that isn't indexed can't be in any result a model reads.

I want to be blunt about this because an industry is forming around the opposite claim. The pitch is that AI search is a fresh start where small sites leapfrog established ones through some technique the incumbents haven't found yet. It isn't. The ranking is the mechanism. A site that ranks nowhere gets cited nowhere.

What is different is which of your pages gets pulled into an answer once you are in contention. That part is winnable in ways ordinary ranking isn't.

Can a crawler read the page at all

Before anything else, the content has to exist in the HTML your server sends.

Most AI crawlers do not run JavaScript. Googlebot does. So a page that assembles its content in the browser can rank perfectly well in Google and still be close to blank for ChatGPT, Claude, and Perplexity, which read the raw HTML and pull the text out of it.

This catches good sites as well as bad ones. Tabbed sections, carousels, "load more" buttons, and anything that swaps content on click will often put one item in the HTML and generate the rest on interaction. We had this on our own homepage. A services panel rendered one service and produced the other five only when you hovered over them. Fixing it took an afternoon and put five services worth of copy into the page source that had never been there.

Accordions and FAQ dropdowns are usually fine, as long as they are built with real <details> elements, because collapsed content is still in the HTML. It is the JavaScript swap that costs you. The full test is here and takes about a minute per page.

Say things plainly enough to be quoted

Answer engines lift sentences. So a sentence has to survive being pulled out of its paragraph and still be true and useful on its own.

"We're an AI digital marketing company in Mississauga, Ontario, founded in 2019" can be extracted and repeated. "We're passionate about crafting digital experiences that inspire" cannot, because there is no fact in it. Marketing voice is the biggest single obstacle to being quoted correctly, and most business sites are written almost entirely in it.

Three habits do most of the work here. Put the answer in the first sentence under each heading, so that if the heading asks a question the sentence below it answers the whole question without depending on the paragraph before or after. Name things concretely: cities, dates, prices, technologies, numbers. A model reproduces specifics and paraphrases vagueness into nothing. And write your headings as the question a customer would actually ask. "What does a website cost?" gets found. "Investment" does not.

Hand over your facts in a form that needs no interpretation

Structured data is the difference between a model inferring your address and being told it. Schema.org JSON-LD covering LocalBusiness, Service, FAQPage, and Article states your facts in a format that requires no reading comprehension at all.

It isn't glamorous and it isn't optional. Which types actually earn their keep is a shorter list than most guides suggest, four or five rather than thirty.

The companion to it is consistency. Your address, phone number, and hours should be identical across your site, your schema, your Google Business Profile, and every directory you appear in. Two of anything is worse than one that's out of date. A conflict nothing can settle gets left out rather than guessed at, and a recommendation with your address missing from it isn't much of a recommendation.

Publish an llms.txt

A plain-text file at /llms.txt summarising what the business is and what each page covers. It is a young convention, unevenly supported, and takes an afternoon to write. That combination will not last much longer. Here is what goes in one.

Then check you aren't blocking the crawlers

Plenty of sites disallowed AI user agents during the training-data backlash and never went back to the decision. Open your robots.txt and look for these:

ClaudeBot
Claude-User
GPTBot
OAI-SearchBot
PerplexityBot
Google-Extended

If any of them are disallowed, you have opted out of the channel you are trying to win. There are legitimate reasons to block some and not others, and they are worth deciding on rather than inheriting from whatever template your site came with.

What doesn't work

Anything sold as an "AI backlink", an "LLM citation placement", or a service that will get you "indexed by ChatGPT". An AI citation is generated inside one private conversation and disappears with it. There is no URL, so there is nothing to buy, sell, or link from. The longer version of that argument is here and is worth reading before you spend money on it.

The exception, and it's a narrow one, is that ChatGPT now runs actual advertising. That's a real product you can really buy, sitting below the answer rather than inside it. What it is and whether it's worth testing is a separate question from everything above, because it buys attention rather than the mention.

Where to look for results

Not in Search Console. AI citations will never appear in the link report, so if that is where you are watching, you will conclude the work failed while it is quietly working.

AI referrals arrive in analytics as referral or direct traffic. Track them as their own line next to organic. Then once a quarter, open ChatGPT, Claude, and Perplexity and ask what a customer would ask, and follow it by asking each one what it knows about your business by name. Write down what comes back, including the sources it names, which are the part worth staring at. Measuring AI search traffic covers what to do with them.

Most of this is a week of unglamorous work that compounds. If you would rather it were done properly the first time, that is what our AI SEO and analysis service covers. The technical SEO underneath it, which is what makes any of it hold, is part of the same programme.