Found

The score says whether AI can read you. Found says whether it names you. Every month we ask the assistants people actually ask — ChatGPT, Claude, Gemini, and Perplexity — what you want to be found for, and count how many times they name you. It is not a score. It is a count: each question goes to each assistant, and you see how many times you were found, where, and who was named instead.

Three things you tell us

You tell us three things: what you are, the problem you solve, and up to five sites you’re compared to, by their websites (honeybook.com, or any link to it), each with its name after an = if you’d rather say it yourself (hellobonsai.com=Bonsai). None may be you or carry your own name or domain. From them we build the questions a stranger would ask, in the same words for every site:

1. What’s the best {what you are}? “What’s the best studio management software?”
2. How do I {the problem you solve}? “How do I run a design studio without a bookkeeper?”
3. What are the alternatives to {name} ({host})? One real question per site you’re compared to, never a combined one, using the name that site gives itself and its host, so a shared name draws the right list: its og:site_name, else its JSON-LD Organization or Product name, else its title before the first separator, else its host’s first word; “What are the alternatives to HoneyBook (honeybook.com)?” A name you give wins.

Those three things are what you want to be found for. The first two are what Found counts: what you are, and the problem you solve. The sites you’re compared to widen the search: their questions are asked too, and every site their answers name joins the list of who gets named, but they are never part of your count.

A fourth, “What is yoursite.com?”, is asked of every assistant and never counted: it names you by construction.

Search and memory

Every question is asked twice. Memory is the model without search: what it carries from training, so a name there means the model has learned you. Search is the model with search: whether the live web puts you in front of it today. Perplexity always searches, so it has only the one. That makes seven answers for each question and fourteen for the two Found counts: “Found 9 of 14 times”.

A site whose three things are not set yet reads “Not asked yet”. Found 0 of 14 times is a result, not a blank, and the report shows what each answer said so you can see why.

Who gets named

Every site the answers name, from all the questions, the alternatives included, is ranked by how many answers named it. Yours is always on the list, in its own place, however far down. Each links to its own report.

How the judge decides

Each assistant is asked for a list, not an essay: the names and websites it would point to, up to ten, one per line. An answer names you if it carries your domain as a whole host (a subdomain is you; a look-alike isn’t), or your site’s name as a whole word, or cites an address whose host is your name run together. The name is the one your site’s own JSON-LD gives, else the host. The sites you’re compared to are known the same two ways, by their host and by their name.

Two rules keep it honest. From memory, a model can invent an address, so there a URL counts only when its host is yours; the name written as text still counts. And when your name is shared (Wikidata knows another site by it), the name alone isn’t enough: the answer has to carry your domain, or your name with a second signal from your site, such as a word from your own description or your city.

Each assistant runs what its free tier runs: the small, fast model, reasoning low or off, since that is what people are actually asked. Only the vendors’ official APIs are called, never a consumer app and never a scrape. Changing a model changes the rubric, so it takes a new version.

Why the questions are built, not written

The same wording for every site is what makes the counts mean the same thing from one site to the next, and from one month to the next: a change in your results is a change in the answers, not in how the question was put. What is yours is the substance, which no fixed list could supply: what you are, the problem you solve, and who you are measured against. Change any of them and we ask again at once, so the record shows the question that was asked.

Asked monthly: a month is long enough for a fix to reach the models, and short enough to catch them changing their minds.

On the plans

The Hide and Seek Score predicts Found: a site the machines can read and find vouched for is a site they can name. Found shows the outcome. Every plan counts your sites’ score every week and asks the machines about you every month. The plans.