Hide and Seek Rubric
Everything that earns the Hide and Seek Score, handed out in advance: every criterion, in four categories, each something you can see for yourself and something you can do, so any site can reach 100. Authority, who vouches for you, is counted beside the score and is not part of it. Found, whether the assistants name you, is counted beside it too: a results grid, question by question. Criteria are tested in public before they count: the ones still testing are measured on every site and shown on every report, and never scored.
The score, from the criteria up
Every criterion gives a finding a score from 0 to 1: the fraction of what it asks for that your site does. A criterion that looks at several things (the home page and ten sampled pages, say) scores the share that pass. The marks are bands on that score, for reading: ✓ from 0.9, ! from 0.4, ✗ below.
A criterion that could not be judged (a page that didn’t load, nothing to look at) is skipped, with the reason, and left out of both the numerator and the denominator. A criterion whose source didn’t answer (a timeout, an error) is asked once more; if it still doesn’t, it is unknown: its category’s number is over the criteria that answered, with the count that didn’t beside it, and the score counts it at the midpoint, so a source’s bad minute never raises a score. The report says so, and can be scanned again. A criterion still testing has no weight: measured on every site, shown at the bottom of the report, never counted.
Each category is the weighted mean of its counted criteria, 0 to 100. The score is the weighted mean of the categories:
| Category | Weight | Asks |
|---|---|---|
| Access | 35 | Can they get in? |
| Readability | 25 | Can they read it? |
| Identity | 15 | Do they know what you are? |
| Coverage | 25 | Is the whole site findable? |
Access is scored by who you let in: every other Access criterion’s points are earned in proportion to the robots.txt criterion’s score.
Within a category, a criterion’s weight is relative to the others’: the page for each criterion says what it is out of. A category with nothing counted has no score and is left out of the total. The grade is a banding of the score: A from 90, B from 80, C from 70, D from 60, F below. Only a true 100 shows as 100.
Versions
Every report and every weekly count carries the version of the rubric it was scored under, named for the day it took effect; this is Rubric 20261116. It moves whenever a criterion is added or dropped, a weight changes, or an assistant’s model changes. Scores made under different versions are not compared: your account shows no change across the boundary, the weekly digest counts none, no drop alert fires, and a thin tick on the trend line marks where it happened. An old report keeps its number and its version; nothing is scored again. Every version, and what it changed, is at the end.
You and others
All of it is yours to do: a robots.txt that lets answer bots in, pages that read without JavaScript, JSON-LD that says what you are, a sitemap that is true, a Wikidata item and directory listings that point back at you. The report lists what falls short under “Fix first”, worth the most first. Until November 5, 2026 a fifth category, Authority, scored what others had said about you — a Wikipedia article, a Hacker News thread, a place in Common Crawl — and no site could reach 100 without them. That broke the score’s promise, so they left it. The claims you make about yourself elsewhere — a Wikidata item, directory listings, profiles that link back — are yours to make, and moved to Identity.
What Authority counts
Authority is the pages on other sites that name you, one row each, ranked by how widely read the site is — ten ticks filled from its place in Common Crawl’s web graph — never a number out of 100, and never in the score. Authority isn’t scored: we show it because who names you plausibly affects whether AI does, and Found will tell us which of these sources matter. For each, what we read, why it’s there, and what we’re trying to learn:
- Wikipedia. What we read: any article that links to your domain or a subdomain of it, asked through the API’s list of external links over both protocols; a page titled with your name alone is not a link to you. Why: models learn heavily from Wikipedia, and assistants cite it. What we’re learning: whether being linked from Wikipedia goes with being named, especially from memory.
- Hacker News. What we read: stories and comments that link to or name your domain as a whole, the newest hundred read; a loose match on a word of it is not a mention. Why: the largest public discussion of software and the web, read by the people who build AI. What we’re learning: whether it matters for software and developer sites; we expect it to matter little for others.
- Web search. What we read: pages a web search finds for your domain or your name, each opened and checked: it links to you, or writes your domain in what it says. Why: it covers every kind of site. What we’re learning: the general shape of who names you.
- Reddit. What we read: Reddit threads that name your domain, as a web search finds them: Reddit doesn’t let us read its pages, so we go by the search result’s title and snippet: your domain, or your name beside a word from your own description; a thread that names only your name is not a row. Why: answer engines cite Reddit often, especially for products, places, and local services. What we’re learning: whether being discussed on Reddit goes with being named by the assistants.
- News. What we read: news articles from the last three months, worldwide, that name your domain or your name (the GDELT Project’s index of online news), checked before they count: your domain in the headline, or your name there beside a word from your own description, or the article opened and found to link you or write your domain; what we found earlier is kept, each with its date. Why: press coverage is how places, shops, and organizations get talked about, and assistants cite news. What we’re learning: whether being in the news goes with being named.
A model reads each result’s title and excerpt and decides whether it is about your site. Your own pages, your own profiles, and directory listings are set aside by rule before it sees them.
Only others count. A page you made is not someone else naming you:
- Your accounts. Your profiles and what’s under them — the ones your sameAs and rel="me" links name on a platform, and your handle on a social site — are Identity’s, never a row; so are your own videos and files on a platform, by the owner the page names.
- Your listings. A directory’s or a store’s page for you (Trustpilot, the App Store, G2, Crunchbase, a stock quote, a publisher’s author page) is a listing, never a row; directory listings counts the ones your site links to.
- Your own properties. Your help center, your shop, your admissions site, your Wikidata item, and your platform’s own hosts are yours, never a row — a host named for you whose pages link you in their chrome is one of them; your own writing elsewhere still counts.
A profile page is never a row, whoever’s it is: an account isn’t someone naming you, though a post on the same site can be. And a page that writes your address counts when it’s in what the page says, not in its title or an account’s name.
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 question you want to be the answer to, and up to three sites you’re up against, 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: the first starts with a stem you pick, the second you write whole, as someone would ask it before they know what you are:
- What’s the best · Who’s the best {what you are}? “What’s the best studio management software?”
- {the question you want to be the answer to}? “How do I stop chasing clients to pay their invoices?” Your own question, asked as you write it: the one someone asks before they know your category exists.
- 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 question you want to be the answer to. 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”. An assistant that doesn’t answer still counts among the fourteen — its cell says so, and we ask it again next month. Fourteen answers is a small sample: a move of one is noise, not a trend.
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 AI recommends
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 question you want to be the answer to, and who you are measured against. Change any of them and the new ones get asked at your next count, or at once from your report, twice a month, 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 and Authority predict Found: a site AI can read, and that other sites vouch for, is a site it can name. Found shows the outcome. Every plan counts your sites’ score every week and asks the AIs about you every month. The plans.
What we deliberately don’t score
- Speed, beyond a floor. A home page that answers in under two seconds is fast enough for anything that crawls; what is faster than that is for people, and the score is not a performance audit.
- “Answerability”. Whether your prose is phrased the way an assistant likes to quote it is writing advice, not a criterion anyone can verify, and the outcome it promises is what Found measures directly. The one piece of it that can be measured, whether your home page says what the site is in its first sentence, is still testing.
- Whether a page answers your question. Found notes, under the question you want to be the answer to, which of your pages answers it, and that note never affects the score.
What we can’t ask
Whether Bing or Google has the site. ChatGPT search answers from Bing’s index; AI Overviews, AI Mode, and Gemini from Google’s. We wanted to ask each whether it has a site, since a site missing from either is invisible to the largest answer surfaces whatever else it does right. But Microsoft retired the Bing Web Search API in August 2025, and Bing’s results page ignores a site: search from anything it takes for a bot (DuckDuckGo, which reads Bing’s index, answers the same fetch with a captcha). Google closed whole-web search to new Programmable Search Engines in January 2026, and an engine restricted to a list of sites returns nothing for site: searches on other hosts. So neither index can be asked whether it has a site, not without pretending to be a person, and we won’t. Google will still answer the site’s owner directly: search for site:yoursite.com and see what comes back.
Whether Common Crawl has the site. Common Crawl is the open copy of the web that much of what AI models know came through, and until November 5, 2026 a criterion asked its index whether the latest crawl held your domain. It was dropped: being in a crawl is not a site that names you, there is nothing an owner can do about it but wait, and the index timed out on most of the sites we tried. Common Crawl stays in the product as the web graph Authority ranks by.
What the score isn’t
The score is not a ranking, not traffic, and not a judgment of what you’ve written. It doesn’t say who vouches for you; that is Authority, counted on a plan’s report and never in the score. It doesn’t say whether an assistant names you when people ask; that is Found, which is counted beside it and is not a score. A 100 means one thing: nothing the rubric asks for is left undone. What remains after that is what your site says and who repeats it.
Access 35 of 100
Can they get in?
Access is the robots.txt criterion’s points, plus its share of the points the other Access criteria earned. A site that turns away ChatGPT, Claude, Perplexity, and four of the smaller assistants keeps 5 of 15 by assistant, a third: 10 of 30 for robots.txt and a third of the other 70, so at most 33. A site that turns every assistant away has an Access of 0, and one that turns nobody away loses nothing. That is a choice, not a fault, and the report says so.
- Search and answer bots are allowed30
Does robots.txt let in the bots that search and answer questions?
- Pages can be indexed and quoted25
Do any pages tell search engines not to index them or not to quote them?
- Your firewall lets bots in20
Does the home page reach a bot the same way it reaches a browser?
- AI crawlers get the pages browsers get15
Do AI’s crawlers get the same sampled pages a browser gets?
- One https address5
Does every version of the address lead to one https site, quickly?
- A fast home page5
Does the home page arrive in under two seconds?
Readability 25 of 100
Can they read it?
- Readable without JavaScript45
Is the text in the page itself, or does it only appear after JavaScript runs?
- Headings and landmarks20
Does each page have one main heading, headings in order, a main area, and a language?
- No walls before the content20
Does a cookie banner, age gate, or sign-up box come before the content in the page?
- Images have alt text10
Do the images in your content say what they show?
- Language and encoding5
Do the pages say what language they’re in and how their text is encoded?
Identity 15 of 100
Do they know what you are?
- Structured identity25
Does the home page say, in a form a machine can read, who or what the site is?
- Structured data says what the page says10
Does the structured data say what a reader of the page sees?
- One name, everywhere10
Does the site call itself by the same name everywhere a machine looks?
- Profiles that name you and link back10
Do the profiles the site claims as its own exist, name it, and link back to it?
- Pages cite their sources10
Do the site’s articles link to other sites as sources?
- Every page points home5
Does each page say whose it is, pointing back to the home page’s identity?
- Title, description, and canonical5
Does each page have a title, a description, and a canonical link?
- Link previews5
When someone shares the home page, does the preview have a title, a description, and a picture?
- The identity’s details5
Does the site’s structured identity carry the details its type needs?
- An about page5
Is there a page that says who you are, linked from the home page?
- A Wikidata item5
Is there a Wikidata item whose official website is this site?
- Directory listings5
Does somebody official list the site, in a directory that fits what it is?
Coverage 25 of 100
Is the whole site findable?
- Sitemap pages load30
Do the pages your sitemap lists actually load, with readable text?
- A sitemap20
Does the site have a valid sitemap, and does robots.txt say where it is?
- Recently updated15
Is there any sign the site has changed in the last few months?
- A feed10
Can a reader subscribe to the site’s posts?
- Articles are dated10
Does each article say when it was published?
- Linked from the home page10
Does the home page lead somewhere, and is the rest of the site mapped?
- Missing pages are 404s5
When someone asks for a page that doesn’t exist, does the site say so?
Still testing
The rubric isn’t finished, and it says so. Every criterion is a guess at what makes AI likelier to name a site, and we test the guesses. The criteria below are still testing: we measure them on every site, show them at the bottom of each report, and don’t count them, and each one’s page says what we expect before the data says otherwise. Found tells us, across every site we watch, whether passing a criterion goes with being named. When the evidence is strong enough, by the rule here, written in advance, a criterion starts counting, a weight moves, or a criterion goes. Changes ship monthly, dated, with the reason.
A criterion starts counting, or a weight moves, only when all of these hold:
- Enough sites: at least 100 sites on each side of the comparison, passing and not.
- Two kinds of evidence agree in direction: sites that fix it, compared with similar sites that didn’t, and sites named versus not named for the same question.
- More than one assistant: the effect shows on at least two of ChatGPT, Claude, Gemini, and Perplexity.
- It lasts: the result holds in two monthly analyses in a row.
- It’s the owner’s to do: nothing outside the owner’s control is ever scored, whatever the data says.
A criterion is dropped, or moves back to still testing, when two monthly analyses in a row find no effect with enough sites to see one. Published evidence from others counts beside our own, and contrary evidence is always shown.
- Content signalssince November 8, 2026
Does the site say what AI may do with its content, in robots.txt or an ai.txt?
We expect: No effect on being named: it states permission for training and search, which engines may or may not honor.
- A plain-text versionsince November 8, 2026
Does the site offer its pages as Markdown or plain text, for AI that prefers it?
We expect: No effect on being named, for the same reason as llms.txt: agents that write code ask for it; answer engines don’t.
- Pages long enough to answer, short enough to citesince November 8, 2026
How long is the main content of each page we sample?
We expect: A small effect: pages of roughly 500 to 2,000 words are cited most in two studies (AirOps, Seer Interactive); very long pages less.
- Says what it is firstsince November 8, 2026
Does the home page say what the site is in its first sentence?
We expect: A small effect on memory asks (an assistant that has read a plain first sentence about you can repeat it), and little on search asks.
- FAQ in structured datasince November 8, 2026
Does a page mark its questions and answers up as FAQ structured data, and are they on the page?
We expect: Little or no effect on being named. Many vendors push FAQ markup, and the one study (Seer Interactive, 2026) found that pages heavy with FAQ and HowTo markup won fewer AI Overview slots.
- Articles have authorssince November 8, 2026
Do the site’s articles say who wrote them?
We expect: No effect, or a slightly negative one: the one study (Seer Interactive, 2026) found that the sites with the most author bios had the lowest AI Overview share.
- An llms.txt filesince November 8, 2026
Does the site publish an llms.txt file?
We expect: No effect on being named. Answer engines don’t fetch it (Ahrefs’ logs, 2026: about 97% of the files were never requested); it helps coding agents, which isn’t what Found measures.
Every version
- Rubric 20261116November 16, 2026
Corrections from an audit of forty sites. A page is an article only by its own content: a date or byline of its own, an entry in your feed, or a place among your posts. An article’s links and length are read from its own text, not from menus, footers, author boxes, or share buttons. A one-page site isn’t asked for a missing-page answer. When a scan can’t read enough of your sitemaps to know your newest date, it says so and scores nothing. And a dozen smaller fixes, each a criterion that said something untrue of one site.
- Rubric 20261115November 15, 2026
Directory listings counts the listings your site links to: in sameAs, or among the links your site shows. A listing we could only find by searching no longer counts, so the score is measured from your site alone, and is the same whenever it is scanned. Authority is no longer gathered on a free scan.
- Rubric 20261110November 10, 2026
The scan reads up to ten pages, and makes up to 40 requests. An image marked alt="" is believed to be decoration, whatever its size: only an image with no alt attribute at all is missing its alt text, and large images marked decorative are counted in a note, never the score. A short page whose words are in the HTML passes content: a page is charged only on a sign that words are missing, such as an empty shell, a notice asking for JavaScript, script that carries the words, or a feed item much longer than its page. Your newest posts are read by every criterion that reads your pages, not only the ones about articles: they join the sample in the places of the pages drawn last, still ten at most, so a site that is mostly posts is judged on its posts. A post under 300 words is not judged for its sources, and a post that links only your own subdomains is neither credited nor failed for it. A Wikidata item that describes another kind of thing than your site is isn’t yours, even when its website redirects to you; we name it, and say what it describes. A feed and an archive page map your site whether or not robots.txt declares a sitemap; a declared sitemap that doesn’t answer costs a fifth beside them, and a Sitemap line that gives a path instead of a full address is said to be invalid. An address, a maps listing, and a Yelp listing are asked only of a site that says it has premises: a LocalBusiness or opening hours, and neither an address alone nor a link to a page of locations says so; a bare Organization is asked for a name, a url, and a description. HideAndSeekBot speaks HTTP/2, as browsers and AI fetchers do: a home page or a robots.txt that refuses HTTP/2 is asked for once more over HTTP/1.1, and an address only an older client gets through is said to fail for real clients. Access is scored by who you let in: every other Access criterion’s points are earned in proportion to the share of the assistants your robots.txt lets in, so a firewall that lets bots through no longer counts for an assistant robots.txt turns away. That is your choice to make, and the report says it that way. A sitemap is the map: a site with a sitemap we could read has mapped its pages, and links no longer asks that the sampled pages be one click from the home page, or for a feed beside the sitemap. A small site’s sample is just smaller: a login, a form and little else, or a thank-you page is never read as one of its pages to make up ten. A figure’s caption is its image’s text: an image with no alt attribute inside a figure whose figcaption has text is not charged.
- Rubric 20261109November 9, 2026
The scan reads more of your sitemap, and samples it better: up to four sitemaps, an index inside an index followed, posts first; five pages when you have five, never a login, a form, or a thank-you page while another is left, and never the same page twice; your newest posts told by their address or your feed when the sitemap gives no dates. And what we couldn’t read of a sitemap is said as that, never as its fault. The kind of site is told from what you declare first: an Organization is never taken for a person, software comes before a person we only guess, and a kind we guessed is never stated as fact; directory listings read what the scan’s search found, and say “couldn’t check” of a listing that didn’t answer. Your about page is never judged as an article, a feed of pages that show no date and no byline is not a feed of posts, an “Updated” date never dates a post, and a byline’s link, your own profiles and repository, and a footer marked only by its class are never counted as sources. A page’s text is counted where it is: no hidden text, a sidebar that holds the page’s heading, and nothing subtracted that the text never held; an image with no size and alt="" is a picture, not decoration. A feed of a test page is no feed, and a feed’s dates are read in more forms; a date is printed in the time zone it was written in; structured data is compared with the page’s own article; and a sentence says what was read: which agents robots.txt was tested for, how many profiles were fetched, where a name came from. When a scan’s 25 requests run out, what wasn’t asked for is said as not checked and nothing more, and what is left goes to the criteria that count toward your score before the ones that don’t; a site that refuses every request is told so once, by every criterion.
- Rubric 20261108November 8, 2026
The scan reads your pages first: after the home page, robots.txt, and the about page come the five sampled pages, then the feed and llms.txt, each with a request kept for it, then the rest; and nothing is scored that wasn’t read — a criterion whose evidence the budget refused says it couldn’t check. One test says what a post is, for every criterion: never a press release, reference docs, a login, a landing page, or a date found in the prose. A firewall refusing bots is charged once; a sitemap that robots.txt doesn’t name earns full credit, a one-page site needs none, and noindex on a listing page isn’t charged; structured data is compared with the node that is the page, and a description may be your meta description. Robots scores by assistant: ChatGPT, Claude, Perplexity, Google, and Microsoft count double, and the answer bots of Apple, Amazon, Meta, and Mistral are judged too. Each criterion’s finding now keeps the numbers its score was made from, so a future rule can be applied to past scans. The criteria that were information are still testing: measured on every site, shown at the bottom of the report, never counted, each with what we expect written down first; two are new, FAQ in structured data and whether the home page says what the site is first.
- Rubric 20261107November 7, 2026
The Hide and Seek Rubric, by name, every criterion with the published evidence for it where there is some; and the weights now follow the evidence: the criteria with the strongest evidence (content, robots, directives, pages) carry the most, and Identity weighs 15 of 100. Three criteria are new (AI crawlers get the pages browsers get, structured data says what the page says, and page length, as information), four fold into two (content, profiles), IndexNow and the Internet Archive leave, and author is information.
- Rubric 20261106November 6, 2026
A byline in a site’s footer or navigation signs no page: a long page is an article only when its own content is signed or dated.
- Rubric 20261105November 5, 2026
Authority leaves the score: four categories, with the claims you make about yourself elsewhere scored under Identity and Common Crawl dropped. A site that refuses every request is one sentence and no score.
- Rubric 20261104November 4, 2026
Wikidata is looked up by official website, the Internet Archive by its capture index, and every criterion vetted against a real site: if a person can plainly see it, the criterion must too. Unknown never raises a score: a source that didn’t answer counts at the midpoint, and a home page that isn’t HTML means no score at all.
- Rubric 20261103November 3, 2026
Found counts two questions, what you are and the problem you solve, seven asks each; the criteria are unchanged.
- Rubric 20261102November 2, 2026
Found’s questions are built from the three things you tell us, in the same words for every site; the score is unchanged. One score, the criteria alone, its letter always beside it; Found is counted beside the score, never folded into it.
- Rubric 20261101November 1, 2026
IndexNow and Common Crawl weigh less; the metadata and overlay criteria read the sampled pages, not the home page alone. A Google index criterion was built and withdrawn before any scan ran it; the same forty criteria.
- Rubric 20261025October 25, 2026
The launch rubric: forty criteria in five categories.