How to Get Your Website Cited by AI: What We Changed on Ours

Search used to end with a list of links. Increasingly it ends with an answer, and that answer either mentions you or it does not. Ranking third for a question an assistant resolves in a single paragraph is worth considerably less than it used to be.

We rebuilt this site around that problem. Below is what we changed, what we deliberately did not change, and what we still cannot measure. All of it is live on this site right now, so you can check any claim here against the thing itself.

The part most people get wrong

"Optimising for AI" sounds like it should be a new set of tricks. It is not. An assistant can only use text it can fetch, parse and trust, and those three verbs are the entire job.

Fetch means the words exist in the HTML, rather than being assembled by JavaScript afterwards. Parse means the structure of the page says what is a question, what is an answer, what is the business and what is a service. Trust means the facts hold up when checked against everything else that is known about you.

Most of the work is unglamorous. None of it is clever.

What we changed

Answering the question in the first two sentences

Our comparison pages open with a block called "the short answer": two or three sentences that still make sense with everything else on the page stripped away.

That is the shape an assistant quotes. A page that builds carefully to its point over eight paragraphs offers nothing clean to lift, so it gets paraphrased instead, and a paraphrase is exactly where the attribution goes missing.

Write the answer first. Then write the page around it.

Publishing questions as questions

Every service, industry and comparison page on this site ends in an FAQ section, and each one is published as FAQPage structured data, so a question is labelled a question rather than inferred from some bold text.

The rule we hold ourselves to is that the marked-up answer is identical to what a visitor reads. Structured data that says something different from the visible page is the fastest way to have all of your markup ignored.

Saying what the business is, once

Spread across a website, one business can easily look like several. The contact page gives one phone number, the footer formats it differently, a service page implies a second company entirely.

We publish a single business entity with a stable identifier, and every other piece of structured data on the site points at it rather than restating it. The ABN, the phone number, the service area and all six services hang off that one node. The name, phone and locality you see in the footer are read from the same source that generates the structured data, so the visible version and the machine-readable version cannot drift apart.

Dating pages honestly

Our sitemap reports when each page's content last changed, worked out from the commit dates of the files that actually produce that page's visible copy. Shared furniture like the header and footer is deliberately excluded from that calculation.

That exclusion is the interesting part. It would be trivial to stamp every page with today's date on every deploy. But Google asks that this date reflect a significant change, and treats a sitemap it finds unreliable as noise. A footer tweak that claimed every page on the site had changed would be a small lie that costs you the benefit of telling the truth later.

Shipping a plain-text version of the site

We publish an llms.txt file: a plain-text map of the site listing every page, its description, and all 76 of our published question-and-answer pairs in full. Ours is at durbacti.com/llms.txt if you want to see what one looks like.

Being straight about this one: llms.txt is an emerging convention, not a standard that every AI crawler honours today. We publish it because it costs nothing, because it is generated automatically from the pages themselves rather than maintained by hand, and because the downside if it is ignored is precisely zero. That is a sensible bet, not a guarantee, and anyone who tells you otherwise is guessing as well.

Prerendering every page

Every page here is real HTML sitting on disk before anyone asks for it. Nothing that matters waits for JavaScript.

This is the least fashionable item on the list and probably the most important. Plenty of crawlers, including some feeding AI systems, either do not execute JavaScript at all or give up long before an app finishes rendering. If your content only exists after hydration, you are betting your visibility on the crawler being patient.

Making it fast, and stopping it moving

Cumulative Layout Shift measures 0.0000 on every page we have tested. The fonts are self-hosted and served from this domain rather than fetched from a third party as the page loads.

Speed is a well-worn topic and we will leave it there. The reason it earns a place in a post about AI is blunter than the usual argument: a page that times out is a page nobody reads, human or otherwise.

Saying where the competition wins

We publish pages comparing our AI receptionist to a human answering service and to hiring a receptionist. Both carry a section on when the alternative is the better choice, and both mean it.

Partly that is about being straight with buyers. It is also practical. An assistant asked to compare two options needs a source that describes both of them, and a page that only argues one side reads as marketing and gets handled as marketing.

What we deliberately did not do

This is the shorter list, and probably the more useful one.

The reasoning is the same every time. Structured data you cannot substantiate is the kind that earns a manual penalty, and a model that catches one wrong fact about you has a reason to discount everything else. Being incomplete costs far less than being caught.

What we cannot tell you

We are not going to show you a graph.

Search Console went on this site only recently, so there is no meaningful before and after yet. And there is a harder problem underneath that one: there is no rank tracker for "did an assistant mention us". Citations happen inside conversations nobody shows you, they differ between models, and they move when the models move.

So this is a bet with reasoning behind it rather than a proven return. The reasoning is that every item on the list above is either good practice for ordinary search anyway or costs almost nothing, which keeps the downside small even if the AI part turns out to matter less than expected.

Anyone selling you a guaranteed AI citation strategy cannot measure it either.

If you are doing this yourself

Roughly in order of value:

That last one is not a moral point. It is the one that compounds.

If you would rather not do all this yourself, it is most of what we do when we build a website. Book a free strategy call and we will go through what your current site actually gives a crawler to work with.