How to Get Recommended by AI: We Tested Our Own Site

Someone has probably told you that buyers are asking an AI assistant for recommendations instead of typing into a search box. The advice that follows is always the same list of ten things to do, and it is almost always written by somebody selling the service. This is the version we could not find anywhere: we ran the check on our own site on 16 August 2026, and we are publishing what it returned, including the parts where we came up short. If you want to know how to get recommended by AI, the honest starting point is that nobody outside those companies can show you the function that picks the names. So the useful question is not how to game it. It is what you actually control, and how you check whether you have done it.

The Short Answer

There is no ranking report for this. You cannot log in anywhere and see where you sit inside an assistant’s answer, and any page implying otherwise is selling something. What you can do is make your business legible to software that reads pages rather than looks at them, and then verify that it worked. That is a real job with a short list of steps, and almost all of it is free. We did exactly that on this site. Every article published here, checked on 16 August 2026, carries a machine-readable description of itself. None of it describes what we sell. We do not publish llms.txt, the file some sites add for AI crawlers. Neither gap is fatal, and neither was visible to us until we went and looked. If that is all you needed, the ten-minute check below is the part worth keeping.

What Actually Changed, and What Has Not

Where a Machine Gets Its Description of You

When a person searches, they land on your page and read it. When software assembles a shortlist, it reads a description of your page — and that description is something you either supply deliberately or leave to be guessed. Two public specifications govern that description. Schema.org is the shared vocabulary: an agreed set of type names such as Article, Organization, Service and LocalBusiness, so that “we do emergency plumbing in Pittsburgh” can be written in a form software already understands. Google Search Central documents how that vocabulary is embedded in a page and what Google does with it. Most sites running a modern content plugin already emit some of this without anyone deciding to.

What Is Documented, and What Is Not

This is the line the sales pages walk straight past, so we will mark it clearly. What is documented is the mechanics: what the vocabulary contains, how markup is embedded, and what a search crawler is permitted to read. Those are published by the companies that operate the crawlers, and they are reliable for mechanics and policy. What is not documented in any source this site will cite is efficacy. Our own sourcing rule admits platform documentation for mechanics and policy only, never for efficacy or comparative performance. That rule exists for a reason a reader can check without trusting us: a vendor’s blog post claiming their work produces recommendations is an interested party marking their own paper, and so is a platform’s. So we are going to be careful with our verbs for the rest of this article. Legible, checkable and consistent are things you can verify. Recommended is not something anyone can currently promise you.

We Ran the Test on Our Own Site, and Here Is What We Found

The method was deliberately dull, and you can reproduce every step. We fetched every article published on this site as an ordinary browser would, read the machine-readable block each page emits, and then checked two files at the root of the domain. No private tooling, no vendor dashboard.

What Our Published Pages Emit, Measured 16 August 2026

What Is Missing, and What We Can Honestly Conclude

Three findings, in the order they matter. Our pages describe themselves as writing, and say nothing machine-readable about our business. Every article announces itself as an article by an organization with a named author. Not one of them carries a Service or LocalBusiness type. A system reading this site learns that we publish; it does not learn what we sell or where we work. We do not publish an llms.txt file. Requesting it returns HTTP 404. We checked that against a control: a file name we invented on the spot returned the same 404 with the same response body, which means the absence is genuine rather than an artifact of how this server answers. To be clear about what that file is, llms.txt is a proposed convention rather than a published standard from any body we would cite — a plain-text file at the root of a domain, meant to give AI crawlers a readable guide to what a site holds. We are not going to tell you how widely it has been adopted, because we have not measured that and we are not repeating somebody else’s figure for it. Our crawler access is clean. robots.txt returns HTTP 200, disallows nothing, and points at the sitemap index. That one we got right, and we got it right by accident rather than by decision, which is worth saying out loud. Now the limit, stated plainly because it is the sentence this article most wants to skip. We can show you what our pages say. We cannot show you whether any AI assistant has ever named us, because we have not run that test. We are not equipped to run it in a way we would be willing to publish, and we are not going to imply otherwise. Anyone who tells you they have measured this should be asked which assistant, on which date, with which prompt, and what happened when they ran it again a week later.

What You Can Actually Control About How to Get Recommended by AI

Four things. Each one is checkable by you, today, at no cost. None is a guarantee, and we are not going to write one. Whether software can read your site at all. This is the floor. A misconfigured robots.txt or a missing sitemap can make the rest of the work invisible, and it is the cheapest thing on this list to get wrong without noticing. Google documents what those files control. Whether your pages state what you sell in a vocabulary software already knows. This is the gap we found on our own site. A services page written for a human is not automatically a services page a machine can classify, and the Service type exists precisely to close that distance. Whether your claims are specific enough to be quoted. What you do, who you do it for, where you work, what is included and what is not. A summary is only ever as specific as the page it was made from, and a summary you did not write is the one that travels. Whether the same facts agree in every place you appear. Your site, your business profile, your directory listings. Contradictory hours, service areas or business names leave any reader, human or otherwise, with no way to tell which version is current. Notice what is not on that list: nothing about publishing more, and nothing that requires a subscription.

How to Check Your Own Site in Ten Minutes, Without a Developer

Every step below happens in a browser tab. You need no plugin, no account, and no code. Two of the six checks are simply typing a URL.

The Ten-Minute Check

Write down what you find. Six answers is a complete picture of your machine-readable surface.

What to Fix First, Before You Spend Anything on This

The order matters more than the work, and it is the opposite of the order most quotes are written in. Buy the plain-language version first. Markup is a translation layer. If your site does not yet say clearly, in ordinary sentences a person can read, what each of your services is and who it is for, then translating that into a machine-readable format produces a precise description of something vague. The first thing worth paying for is a page per service that states what it is, who it suits, and what happens next. Then build the machine-readable layer on top of it. With real service pages in place, the markup follows directly from copy that already exists, and a competent developer or a well-configured plugin can add it in an afternoon. Then make your listings agree. This is administrative rather than technical, it costs nothing but attention, and it is the step that gets skipped. Everything above is work you can commission with a fixed scope and check yourself afterwards using the six-row table. That is a better position to be in than buying an outcome nobody can currently measure.

Where We Would Fail This Test

We would fail two of the six rows on our own check, as of 16 August 2026. We carry no Service or LocalBusiness markup. We publish no llms.txt. We have known both facts for exactly as long as this article has existed, because running the check is what produced them. We have not fixed them yet, and we are telling you that rather than quietly closing the gap before publishing. Two reasons. The first is that a fix applied the same morning we found the problem would leave you no way to verify anything we said here. The second is more honest: we do not know how much either fix is worth, and adding markup we cannot demonstrate the value of, in an article arguing that nobody should sell you unmeasurable work, would be its own kind of failure. We also cannot tell you whether this article will make an assistant more likely to name us. We can tell you it made our own gaps visible, which is the only claim we have earned.

How to Decide Before the Next Invoice

Run the six checks. They cost you ten minutes and they turn a vague worry into a short list of specific, boring, fixable things. If somebody has already quoted you for AI visibility work, the useful questions are which of those six rows they will change, what they will show you afterwards that you could not have seen yourself, and what they will do that the ten-minute check does not already cover. A good answer to that is easy to give. We publish our own mistakes on this site with the same detail, including the ones we made with our own marketing, and the broader question of where AI belongs in your marketing is covered in our guide to using AI in marketing. If you would rather talk it through against your own site, book a strategy call — thirty minutes, and we will run the six checks against your domain while you watch.