We wrote a while ago about the shift from search results to AI answers, and why being recommended matters more than ranking now. That piece deliberately stopped short of a recipe, because there isn't a single one. This is the follow-up people kept asking for: not a recipe, but the actual levers, and an honest account of which ones move the needle and which are a waste of your afternoon.
Here is the pattern we see over and over in the audits we run. The businesses assistants confidently recommend are almost never the ones with the most content or the cleverest markup. They are the ones whose core facts are clear, consistent, and confirmed in more than one place. That is the whole game, and most of the levers below are just different ways of getting there.
First, what does not work
Let's clear the hype out of the way, because there is a lot of it.
llms.txt. The idea of a file that tells AI models how to read your site sounds tidy. In practice the major assistants do not meaningfully rely on it, and no amount of it compensates for a site that is unclear about what it does. It is not harmful. It is just not the thing.
"AI schema" and magic tags. There is no secret markup that makes a model recommend you. Structured data helps machines parse facts, which is genuinely useful, but it is table stakes, not a cheat code. Anyone selling an "AI optimisation tag pack" is selling you 2011 SEO in a new hat.
Stuffing your site with question-and-answer text aimed at models. Assistants are built by people who are very good at spotting content written to manipulate them rather than to help a human. Writing for the model instead of the reader is the fastest way to be discounted.
The uncomfortable truth is that the durable levers are less exciting than the hacks, which is exactly why they work: they are harder to fake.
Lever 1: Say plainly what you are
An assistant can only recommend you for something if it can tell, without guessing, what that something is. This sounds obvious and is failed constantly. Sites lead with a clever tagline, bury the actual service three clicks deep, and describe what they do in language only an insider would decode.
Make the core facts explicit and early: what you do, who you do it for, where you operate. A model reading your homepage should be able to complete the sentence "this is a company that ___ for ___ in ___" without inferring anything. If it has to guess, it will often guess a competitor who made it easy instead.
Lever 2: Be the same everywhere
This is the single most underrated lever, and the one our audits catch most often. Models build confidence by cross-checking. If your site, your Google Business Profile, your LinkedIn, your directory listings, and the places that mention you all agree on your name, what you do, and where, that consistency reads as reliability. If they disagree, even in small ways, the model has a reason to doubt you and a safer alternative to reach for.
Practically: one consistent business name, one description of what you do, one location story, everywhere. The boring work of making the whole web agree about you is worth more than another blog post.
Lever 3: Get talked about by sources that are not you
An assistant weighs what independent sources say about you far more heavily than what you say about yourself. A mention in a respected publication, a genuine review, a case study on a client's site, a listing in a directory that matters in your field: these are the things that let a model recommend you without taking your word for it.
You cannot fake this at scale, which is precisely why it counts. One real, credible third-party signal is worth more than a hundred self-published pages.
Lever 4: Structure it so a machine can read the meaning, not just admire the layout
A site can look beautiful to a human and be nearly opaque to a machine. Meaning carried only in an image, facts implied by layout rather than stated in text, a single-page design that hides its substance behind animation: all of it makes you harder to parse, and a model that cannot parse you cannot confidently recommend you.
Clean structure, real text for the things that matter, headings that describe content rather than decorate it, and proper structured data where it applies. None of this is glamorous. All of it compounds.
Why this is a build decision, not a marketing task
Notice that almost every lever above is decided by how your presence is built, not by something you can sprinkle on afterwards. Whether your core facts are legible, whether your signals are consistent, whether your structure carries meaning: these are architecture choices, made well or badly at the start. By the time you are wondering why the assistants never mention you, the cheap moment to fix it has passed and you are retrofitting instead.
That is why we treat AI discoverability the same way we treat performance and SEO: as something built into a site from the first decision, not bolted on once it is live. The businesses that get recommended in a couple of years are the ones making themselves legible now, while most of their competitors are still treating their website as a brochure.
If you want your site to be the answer an assistant gives rather than the one it skips, start a conversation with us, or read about how we build search and AI visibility into everything we make.