How do you get your company cited by AI assistants?

Assistants do not rank, they synthesise

A search engine returns a list and lets the person choose. An assistant reads several sources and writes one answer, citing a handful. That difference changes the economics completely.

Position seven on a results page still gets clicks. There is no position seven in a generated answer. You are in the paragraph or you are not, which makes this closer to winner-take-few than to a gradient. It also means the useful question is not how to rank. It is what makes a passage safe for a model to lift.

Test one: can a crawler read you without JavaScript?

Several of the crawlers feeding assistant answers do not execute JavaScript. If your content is assembled in the browser, what they fetch is an empty shell with a loading spinner in it. No amount of good writing survives that.

How to check yours in thirty seconds: open the page, view source, and search the raw HTML for a sentence from the middle of your copy. If it is not there, it does not exist as far as a non-rendering crawler is concerned. Do it for your service pages, not just the homepage, because they are usually built differently.

For reference, we measured our own site the same way: between 97% and 100% of the text inside the main element of every page is present in the raw HTML before any JavaScript runs. That is not a boast, it is the floor. It is also the single most common reason a well-written site is invisible to assistants, and it is invisible to the people running the site because their browser renders everything perfectly.

Test two: does your company resolve to one entity?

Assistants corroborate before they cite. A company that appears under three variations of its name, with two phone numbers and an address that differs between the website and the Google Business Profile, does not consolidate into one confident entity. It looks like three weak ones.

We shipped this bug on our own site. Our schema.org markup declared the organisation under one name, our Open Graph tags used a longer version of it, and the domain was a third string. Three names for one company, in our own machine-readable markup, on a site that sells this as a service. It is an easy thing to miss precisely because every individual page looks correct to a human.

What consolidation actually requires:

• /One name, spelled one way, in the schema.org organisation node, in og:site_name, in the page copy, in the Google Business Profile and in every directory listing.
• /A stable @id on the organisation node. Without one, the same organisation object repeated across thirty pages reads as thirty unrelated declarations rather than one entity seen thirty times.
• /sameAs pointing at the profiles you control: LinkedIn, the Google Business Profile, Instagram. This is the corroboration mechanism. It is how a model confirms the entity on your site is the same one it has seen elsewhere.
• /Identical phone number and address everywhere, down to the formatting.

None of this is exotic and all of it is checkable in an afternoon. It is also the part most agencies skip, because it produces no screenshot.

Test three: have you published anything worth quoting?

This is the one that actually decides it, and the one no technical fix reaches.

A model quoting you needs a sentence that is specific, self-contained and attributable. Look at your own site honestly and ask which sentence a model could lift into an answer without embarrassing itself.

Hedged copy is worse than vague copy. A page that says a figure is illustrative and not an average is explicitly instructing a model not to cite it, and the model will comply. That hedge may be the right call, ours are deliberate, but be clear about the cost: every caveat you add is a passage you have removed from the citable pool.

Which leads to the uncomfortable part. If everything specific you know is reserved for a sales conversation, there is nothing on your site to cite, and you will not appear in answers no matter how clean the markup is. Getting cited requires publishing something you would rather charge for.

The plumbing, in order of how often it is broken

1. 01Server-render the content. Everything else is decoration if the crawler sees an empty page.
2. 02One correct canonical URL per page. We found every page on our own site declaring the homepage as its canonical, inherited from a single line in a layout file. A crawler reads that as a site of duplicates, and it is silent: nothing looks wrong in a browser.
3. 03schema.org as a graph with a stable @id, not a loose object repeated per page, so the publisher and author references resolve to one node.
4. 04FAQPage markup on real questions, and never the same question on two URLs. Duplicated FAQ markup means neither page gets the result.
5. 05An llms.txt naming what the site actually answers, which is the file an assistant reads to orient itself.
6. 06A robots.txt that names the answer engines. Note that Google-Extended and Applebot-Extended are separate tokens governing AI grounding rather than classic search, so allowing Googlebot does not allow them.

Treat that list as hygiene rather than a growth lever. It does not get you cited. It removes the reasons you would be excluded.

What nobody can promise

There is no submission form, no ranking API, no inclusion guarantee, and no agency relationship that changes what a model says. Answers vary between assistants, between phrasings of the same question, and between one week and the next.

So treat a promise of guaranteed AI citations the way you would treat a promise of a guaranteed Google position: as a claim about something the seller does not control. What is controllable is being parseable, being one entity, and being quotable. All three are necessary. None of them is sufficient, and we would rather say that than sell the other thing.