AI · · 7 min read

HOW AI DECIDES WHO TO RECOMMEND

Search gave your customer ten links and a choice. An assistant gives them one answer and a name. Being the name is a different job.

Read aloud by your browser
The short version
  • AI assistants recommend businesses they can read, corroborate and attribute. Those are three separate problems.
  • Most assistants answer from a mix of training data and a live retrieval step. You can influence the retrieval step this quarter; you cannot influence training data.
  • Being mentioned on sites an assistant already trusts matters more than anything you publish on your own domain.
  • The single most common blocker is a site an assistant cannot parse: content that only exists after JavaScript runs, or pages with no clear answer near the top.

When someone asks an AI assistant which consultant, agency or supplier to use, how the assistant chooses who to recommend is not mysterious, but it is different enough from search that most marketing habits do not transfer. Search returned a page of options and let the person choose. An assistant returns an answer, and often a name. If it is not your name, there is no second place to be seen in.

The mechanics vary by product, but the shape is consistent. There is what the model absorbed during training, and there is a live retrieval step where it fetches current pages before answering. Those two behave very differently and it is worth being clear about which one you can affect.

What happens between the question and the nameSimplified, but the stages are real. Most businesses fail at the second one and never reach the fourth.

It has to be able to read you

This is the unglamorous one and it disqualifies more businesses than anything else. A lot of modern sites send an almost empty HTML file and build the page with JavaScript in the browser. Search engines handle that reasonably well now. Many AI crawlers do not run JavaScript at all, so they receive a page with your navigation and nothing else.

If that is your site, you are not competing badly. You are absent. Check it the crude way: fetch one of your pages without a browser and read what comes back. If your argument is not in there, no amount of writing will help until it is.

It has to be able to corroborate you

Assistants are built to avoid asserting things they cannot support. A claim that appears only on your own website is a claim from an interested party. The same claim appearing in an industry publication, a client case study on someone else's domain, a directory, a podcast transcript and a conference listing is a claim with corroboration.

This is why the traditional advice to publish more on your own blog only goes so far. Your own domain establishes what you say about yourself. Other people's domains establish whether it is true. Both matter, and most businesses over-invest in the first.

It has to be able to attribute you

An assistant that wants to name a business needs an unambiguous entity to name. If your consultancy trades under one name, invoices under another, and appears on LinkedIn as a third, you have made attribution harder than it needs to be. Consistent naming, a consistent description, and structured data that says plainly who you are and what you do all help resolve you to a single thing worth naming.

1answer, not ten links. There is no page two to rank on
3separate problems: readable, corroborated, attributable
0influence you have over what is already in training data

What to actually do first

  1. Confirm your pages serve real content without JavaScript. Everything else is wasted until this is true.
  2. Put a plain answer near the top of pages that answer a question, before the argument for it.
  3. Fix your entity: one name, one description, consistent across your site, your profiles and your listings.
  4. Get mentioned somewhere you do not control. One credible third-party mention outperforms a month of your own posts.
  5. Add structured data that states what you do and where, so there is nothing to infer.

None of this is a trick and none of it is quick. It is the same work that made a business findable before assistants existed, with the ordering changed and much less tolerance for a site that cannot be read. If you want to see where you currently stand, the free AI visibility checker runs the readability and citation checks in about a minute, and the AI search readiness checklist is the longer version to work through.

Google's own guidance on this has been consistent and worth reading directly rather than through a summary. Their structured data documentation (opens in a new tab) explains what they use and what they ignore, which cuts through a lot of the advice being sold at the moment.

COMMON QUESTIONS

How do AI assistants decide which businesses to recommend?

They combine what the model absorbed in training with a live retrieval step that fetches current pages. A business gets named when its pages can be read without JavaScript, its claims are corroborated on sites it does not control, and its identity is consistent enough to attribute cleanly.

Can I get my business into an AI model's training data?

Not directly, and not on any timeline you can plan around. What you can influence is the live retrieval step, which is where most recommendation answers actually come from. That responds to changes within weeks rather than years.

Does my website need to work without JavaScript for AI search?

Yes for most AI crawlers. Google renders JavaScript, but many assistant crawlers fetch the raw HTML and nothing else. If your content only exists after JavaScript runs, those crawlers receive an empty page and cannot cite you.

Is being mentioned on other sites more valuable than my own blog?

For getting recommended, usually yes. Your own domain establishes what you claim about yourself. Third-party mentions are what let an assistant treat the claim as corroborated rather than as marketing copy.

Recognise any of this?