The short answer

When someone asks ChatGPT, Perplexity, or Google's AI Mode for "a good bookkeeper in Boise," the assistant doesn't read your homepage and form an opinion. It builds an answer from what other sources say about you — local roundup posts, directory listings, review sites, forum threads, the regional paper, a supplier's partner page.

So the honest answer to "how do I get recommended by AI?" is this: most of the work happens off your own website.

That finding turned up again in seoFOMO's 2026 State of AI Search Optimization report, a survey of people doing this work daily. Much of the report is industry shop talk. This part isn't — it changes what you should do on Tuesday morning.

Your website is the confirmation, not the recommendation

Your site still matters. It's how an assistant confirms you exist, what you do, where you are, and whether you're open. If your site is vague about your services or your service area, you'll get skipped even when you deserve the mention.

But confirming is a different job from recommending. The recommendation comes from elsewhere.

This flips the usual to-do list. Adding a fourth service page to your site is worth less right now than getting named once on a page someone else owns.

For agencies, this is the harder conversation and the more valuable one. Clients are used to buying deliverables that live on their own domain. The work that actually moves AI visibility often produces nothing you can screenshot in a CMS — a corrected directory record, an email that lands you in a roundup, two new reviews on the site that keeps getting cited.

The four-step check you can run today

This takes about thirty minutes and needs no tools.

1. Ask the question a customer would ask.

Not your business name — the need. "Best [what you do] in [your town]." "Who should I call for [specific problem]?" Try it in two or three assistants, because they don't all pull from the same places.

2. Write down the sources it cites.

Most assistants show their links. Note the top five or six. That's your list: the pages an AI actually reads when it answers questions in your category.

3. Check whether you appear on them.

Often you won't. It might be a "12 best X in [town]" post from a local blog, a chamber of commerce directory, or a niche listing site you've never heard of. Some of them will be genuinely obscure. That's fine — obscure to you isn't obscure to the model.

4. Pick three and act.

Do these, properly:

Three of those, done properly, will move you further than a month of website tweaks.

Nobody has clean measurement yet — start counting anyway

The same report makes a second point worth knowing: measurement in this category is still rough for everyone.

AI referrals usually land in your analytics as plain "direct" traffic. Someone asks ChatGPT for a recommendation, sees your name, types your business into a browser, and arrives with no trackable source attached. Your dashboards will lag behind reality for a while yet.

Don't wait for a perfect report. Add one question to your intake form or your phone script:

> "How did you hear about us?"

Then count the people who say an AI assistant sent them. Log it in a spreadsheet with the date. That's your baseline, it costs nothing, and in three months it'll tell you more than most analytics setups will.

Agencies: build this into client onboarding now. It's the cheapest way to have a number to point at when a client asks whether any of this is working.

What this means for the next few months

The practical shift is from publishing to being cited. The old instinct — write more pages, tune more headings — hasn't stopped working, but it's no longer where the leverage is. The leverage is in the handful of pages in your category that assistants already trust, and whether your name appears on them accurately.

Start with the four-step check. You'll likely find two or three obvious gaps, and closing them is unglamorous, quick work.

---

The pages an AI cites in your category are the ones worth being on. A SimpleAIO audit shows which of them already mention you, which ones don't, and what to fix first — in plain English, with no dashboard to learn.