Getting quoted by ChatGPT: an experiment on our own site

It is easy to write about answer engine optimisation in the abstract. It is more useful to try it on something you control and see what actually happens. So we did. We took our own site, made a set of deliberate changes, and over a few weeks tracked whether the major AI assistants started quoting us on the questions we care about. This is a field note on what we changed and what seemed to matter, offered in the spirit of showing our working rather than claiming a formula.

Which questions did we want to be the answer to?

We started by writing down the questions. Before touching anything, we listed the handful of things a buyer in our world might ask an assistant, the ones where we would genuinely want to be the answer. This turned out to be the most useful step, because it turned a vague ambition, be more visible to AI, into a specific target we could actually check against. You cannot tell whether you are being cited if you have not decided what you want to be cited for.

How did we make the site readable to machines?

Then we made the site legible. We went through the unglamorous checklist first, because it seemed foolish to optimise content a machine could not cleanly read. Structured data where it was missing, clean headings, the small files that tell a crawler what it is looking at. None of it was visible to a human reader. All of it made the pages easier for a model to parse without guessing, and it felt like the sort of thing that would matter quietly rather than dramatically.

Which change made the biggest difference?

The change that seemed to move things most, though, was editorial rather than technical. We rewrote a few key pages to answer their question directly and early, in plain language, instead of circling it. Where we had made a claim, we put the evidence next to it. Where we had been vague to sound impressive, we got specific. The pages became easier for a person to trust and, it turned out, easier for a model to lift a clean answer from. The two goals kept pointing the same way.

Did publishing more help?

What did not seem to help was volume. We had half-assumed that publishing more would raise our odds, and if anything the opposite held. A few clear, genuinely useful pages seemed to earn more citations than a larger pile of thinner ones. The models appeared to be choosing the source they could trust and repeat, not the one that had simply written the most, which is a reassuring thing to learn if you would rather write well than write constantly.

What did we learn?

We are still watching, and we would not pretend this is settled or that the engines will behave the same way next quarter. But the direction of travel matched the thing we already believed, which is that most of answer engine optimisation is the old discipline of being clear and being honest, done in a way machines can now read. If that holds, it is good news, because it is a game you win by being genuinely useful rather than by gaming anything. We will report back as we learn more.