Become the answer.

When some one asks an AI to find something it replies with a few names and a few links, chosen from thousands. This site asks one question: how does your site become one of them?

Tricks fail when we test them. What remains is seven conditions your website must meet, and a lot of measuring to learn what tips the choice.

The seven outcomes

Before any AI system can recommend you, seven things must be true. Each one is a section of the manual, written in plain English, with the evidence behind every recommendation. Where nobody knows the answer yet, the manual says so.

1. Reachable. A machine can fetch your pages. If the page needs a login, only appears once JavaScript has run, times out, or your robots file turns the AI crawlers away, nothing further down this list can save it.
2. Filed. It files the right version of your page, and can find it again. One page living at four addresses, an old page nobody redirected, a stray "do not index" left behind by a web designer: the system files the wrong copy, or files nothing. Google requires a page to be indexed before it can be used in an AI answer.
3. Understood. It can tell what your business is, does, and where. What you do, who you are and where you work have to be readable as words. A machine cannot get any of it from a photograph, a logo, or a heading that says "Welcome".
4. Relevant. Your pages answer the question being asked, and the questions behind it. Someone asks for a wedding photographer in Leeds, and the AI quietly asks itself several more questions of its own about price, availability, style and how far you travel. Your pages are judged on those too.
5. Worth choosing. They contain something better than the summary everyone else has. Your prices, your photographs, the job that taught you something, what you tell a customer who asks the awkward question. The generic version of your page already exists a thousand times over, and an AI can write it without you.
6. Trusted. The claims hold up, and someone real stands behind them. A named person rather than "the team", contact details that reach a human, dates that are true, and the same business facts wherever else you appear online.
7. Usable. An AI can quote you without error, and a customer can act on what they find. Prices, hours, areas covered and what is included, stated plainly and close to the claim they support, with an obvious way to book or ask.

The first three are plumbing. They are checkable, fixable, and mostly automatic; on the sites bernard hosts they run without anyone thinking about them. The last four come out of what you know and how you run the business, and nobody can do them for you. The fifth is probably the one that decides it, and it is the hardest of the seven to measure, which is much of why we are running the study. The manual covers all seven.

We'd rather count than guess

A measurement study sits behind the manual. We put real buying questions to the AI systems people actually search with, over and over, and record which pages they cite, which they ignore, and what the cited pages contain that the ignored ones lack.

The first study: 50 questions put to ChatGPT and Gemini — through both the developer interfaces and the apps a customer would really use — plus Google's AI answers. That makes about 2,400 recorded answers per round, across five kinds of business: local professionals, home services, artists and makers, coaches and course businesses, and small shops. Round one covers businesses where nobody does this for a living: no marketing department, no agency, no one whose job is search visibility.

Measuring the developer interface and the app side by side also lets us ask something the visibility industry assumes: whether the cheap way of measuring says the same thing as the real one. We registered that question, and the exact statistic that answers it, before collecting any data.

We published our method, and what we expect to find, before collecting a single result. You can check our findings against our predictions. Science treats that as routine; SEO treats it as unheard of.

Our first finding — and our first correction

Before the study proper had run, building the instrument produced a finding. We read the robots.txt of the 32 most-cited domains in our own data. Almost none is closed to AI crawlers — 28 of 32 permit all nine of the major agents — yet the most-cited source of all blocks every automated client in four lines, and is cited constantly anyway, because access to the biggest sources has moved into commercial agreements rather than crawl permission.

It also produced a correction, the same day. Our first reading of one publisher's policy was the opposite of the truth, because a hand-written script misread a robots file that our proper evaluator read correctly. We have corrected the page, named the cause, and withdrawn a second claim that re-measurement did not support. That is the arrangement: we publish what we find, and we publish it again when we get it wrong.

The folklore ledger

Add an llms.txt file to rank in Google. Break your pages into AI-sized chunks. Use special AI schema. Write a page for every question.

You have read the tips. So have we, and the evidence behind most of them is missing. We keep a public ledger of popular claims: what the platforms require, what they recommend, what the data supports, and what turns out to be folklore. When our measurements prove a claim wrong, it goes in the ledger. When they prove one right, that goes in too.

Who's behind this

The team behind bernard writes and runs Content is Everything. bernard is a platform that hosts and looks after small-business websites, including this one. That is our interest, stated plainly: we build the kind of sites this research studies, we test our own advice along with everyone else's, and every article says exactly what bernard can and cannot do about its subject.

Research by Spencer Thursfield. Edited by Spencer Thursfield and Lindsay Smith. Full disclosure, method and corrections policy live on the About page.