Getting found by machines

Some people contacting Morf Research already knew a surprising amount about us. They hadn't arrived with a blank slate and a list of questions. They'd asked an assistant first.

Around the same time, we noticed fewer clicks from search. I couldn't tell how much the two observations explained each other. They did raise a question we needed to answer: what were people being told about our metabolic health programs before they spoke to us?

On our site, a claim sits beside an explanation. A visitor can read who a program is for, what it includes and where its limits are. In an assistant's answer, that material may become a sentence or two. The sentence can sound complete even when an important qualification has disappeared.

For a healthcare program, that gap matters. Being mentioned is of limited value if the mention creates an expectation we can't meet.

Make the answer easier to assemble

We started with the questions a person considering a program would ask. Who is it for? What does it cost? How long does it run? What doesn't it do? Who reviews the medical side?

Then we looked at how much work our pages required someone to do to assemble those answers. A fact in an image, a qualification elsewhere on the page, an old price on a forgotten URL: information can be available and still be easy to get wrong.

We added a block of facts to the foot of program pages, including the results in numbers. We wanted units and context written beside the figures. We rewrote headings around ordinary questions and checked for retired programs and stale prices.

We also checked that the page text, JSON-LD and llms.txt described the same thing. JSON-LD is structured information on a page; llms.txt is a separate file that points to information for language models. I can't say how much weight a particular assistant gives either. Keeping them consistent at least prevents us from supplying conflicting accounts ourselves.

Check what people hear

Once a week, we ask assistants the questions our customers ask and record the answers. Who gets mentioned? What is said about us? Which details are wrong? That gives us a record to compare as we make changes. It doesn't tell us what every customer will see.

It's too early to report meaningful results. We haven't established which changes affect an assistant's choice of sources, or how reliably.

The useful discovery so far has been closer to home: we'd expected people to piece together more than we realised. The same work that makes a program easier to describe also makes it easier for a prospective customer to understand.

This site has an API, llms.txt and a sitemap for a similar reason. They offer different ways to reach the information. I still have to check what happens when someone else uses it to answer a question about me.