How we count a citation
What a citation is here
When we say one of your pages was cited, we mean an AI answer showed that it used that page. Not that it ranked you, and not that it mentioned you. Those are separate things, and we count them separately. The difficulty is in "showed that it used it", because that arrives in three quite different strengths. An engine may hand back its own list of the sources behind its answer. Or the answer may write one of your web addresses out in the middle of a sentence. Or it may just say your domain name in passing. Those are not the same claim, and we do not report them as one.
The three tiers
Declared — the engine returned your page in its own list of sources. This is the engine's own account of what it used, and it is the only one of the three that is ground truth. Address in the text — the answer wrote out a web address on your site. Nothing was declared, but the model is pointing a reader at the page, which is worth something. Name in the text — your domain appears in the prose and nothing more. "Tools like yourcompany.co.uk exist" is this, and so is "unlike yourcompany.co.uk". The model typed a domain. It did not say it read the page, and it may not have. Where the three are combined into a score, they count 1, 0.6 and 0.25 against each other. Those weights are a judgement rather than a measurement: they say that an engine's own record beats an address in prose, which beats a domain name somebody typed. We would rather write that down than bury it inside a number.
Why all three are kept, and never added together
Counting only declared citations would look like the careful choice, and it would mislead you. Several engines never publish a source list at all, so a declared-only count would report those engines as having ignored you when they had not. Leaning on the loosest tier is worse in the other direction. A domain name mentioned in passing is thin evidence, and a figure built mostly from those describes something far weaker than it appears to. So the three are kept apart and never summed into one number. Whenever you see a citation count you can see what it is made of — and if a figure is resting mostly on the weakest tier, you can find that out in one look rather than take it on trust.
The naming rate — the number behind the number
A citation says an engine used a page. It does not say the answer mentioned you. Most third-party pages cited in AI answers about a business appear in an answer that does not name it at all, and fewer still in one that recommends it. It depends on the kind of page: an answer that cites your stockist usually names you, one that cites a competitor's guide rarely does. Your own pages name you every time, by definition. Other people's pages usually do not. That has a consequence worth sitting with: a rising citation count can describe a business becoming LESS visible. More rivals' pages cited for your subject is more citations and less of you. So beside "how many citations" we show "how many of those sat in an answer that named you", each with the number it was counted out of. Never a blended percentage: a topic's citations fall to single figures quickly, and a rate over four citations is not a measurement.
What kind of source it is — and why we ask you rather than guess
The kind of third party predicts naming better than anything else we measure. In one account a stockist named the client in 189 of 213 citations, while a competitor's guide managed 1 of 163. In an unrelated trade, shops named the client 21 times out of 21, and sixteen rival businesses managed 0 out of 17. That is why we ask you to label the domains that come up often. Five citations or more is the point at which it is worth your time. A rule can guess some of them from the domain alone: encyclopaedias and government sites, forums and social platforms, directories and review sites (a directory's site for another country too), company and domain registers, and anything on any competitor list you entered - a watchlist's or a campaign's. It will never guess "stockist" or "publisher" — whether a marketplace stocks YOUR products is not something a domain name knows, and filing a rival's blog under "publisher" would hide the one distinction that matters most. Where a kind came from a rule's guess rather than your answer, the count says so. A platform anyone publishes on - YouTube, LinkedIn, Medium, GitHub, a Google Doc or Site, a hosted blog - gets no kind for the whole site, because whose page it is decides: open the row to see the pages, with ours marked. Your own sites come from your own records: your projects' sites, their subdomains, and the sites your entity lists as the same organisation. One of them with no label is read as Ours. If your account has several projects, we ask once whether they are all your own businesses - an agency's clients are other people's businesses, so their sites stay third parties unless you label one Ours. A label you gave always wins: if you labelled one of your own sites something else, the row says it is also one of your projects and leaves your label as it is. A competitor or a stockist is a relation to one of your businesses, so with one project chosen at the top of Sites AI cites, a Competitor or Stockist label holds for that project only; with every project shown it holds for all of them, as every label you gave before did. A project's own label wins over the all-projects one, and that project's figures, rival list and writers' sources read only its own and the all-projects labels. Every other kind is what the site is, in every project. A label covers the site's subdomains too: label seranking.com a competitor and visible.seranking.com is one. It never reaches across a host where different people publish - labelling wordpress.com says nothing about someone's blog on it, and labelling google.com says nothing about a Google Doc. One rule underneath all of it: the kind is something you tell us, and is never inferred from behaviour. Deciding a site is a competitor because it never names you, and then reporting that competitors never name you, is a circle. It would always look like a finding and never be one. Two kinds are for a site the answer took for something it is not. "Not related" is a different business or subject - most often one that shares your name - cited on a question that named you. It is left out of the rates about other people's sites and is never offered as a page to write to. "Neighbouring field" is a business that sells to your customers without selling what you sell. It has a rate of its own, so it no longer blurs the competitors' or the publishers'.
By subject, not just by domain
Each answer we probe is stamped with the subject it was asked about, so the same citations can be read a second way: for THIS topic, who owns it, and which rivals answer it instead of you. That is a more useful question than the account-wide one. "Stockists name you and rivals do not" is an interesting fact about your account. Per subject it is an instruction. Topics with no citations at all do not appear. Answers belonging to no single subject — the discovery prompts, and the questions that ask whether anyone has heard of you — are left out rather than pooled under a heading nobody chose.
The sentence an engine quoted
Open a site on Sites AI cites to see the pages cited on it. Where Claude, asked directly on your own Anthropic key, quoted a sentence from a cited page, the sentence shows under that page with the date; other engines hand back only the address, so most pages have none.
Where to earn a place: the outreach list
Below the table, Sites AI cites ranks other people's pages a second way: by how many of the answers that cited them recommended someone - you, or a competitor you track - rather than by how often they were cited. A roundup cited beside "the best tools are X, Y and Z" is worth one honest note to its author. A definition cited in answers that recommend nobody is not, and a citation count ranks the two side by side. Each page shows how many of its answers sat beside a recommendation, how many recommended a competitor and not you, which competitors those were, and one sentence on how to approach it. A competitor is only read on the answers that were read for your competitors, so that count is out of those answers, and older answers understate rather than overstate. Your own sites, anything you labelled Ours, your competitors' sites and the pages we published for you are left out. Whether a page already names you is not checked, so open it before you write. The whole list downloads as a spreadsheet from the same place. We contact nobody for you, and nothing on the list suggests paying for a place or swapping links.
Earned placements: the pages you got onto
When you get onto somebody else's page - a review site, a directory, a press article, a forum thread or a guest post - record it on Sites AI cites. Use "I got onto this page" on a row of the list, or "Record a page" to type any address. Say what kind of site it is, the day you got onto it, and who wrote it. We read the page once to see whether it names you. It counts either way, because some pages load slowly or hide their text. From that day on, every AI answer we read is compared with the page, and the first one to cite it is shown with its date and engine. A page someone else wrote counts as independent. A page you wrote, such as a guest post or a directory entry you filled in, counts as controlled, like a page we published. Placements have their own line on What's working, and that line is never added to the pages we published. You cannot record your own site or a page we published for you.
Conversations to join, and pages to pitch
On each project page's Off-site tab, "Conversations and pages to pitch" starts with forum threads on the questions where AI answers of the last 30 days did not name you: Reddit, Quora, Stack Exchange and community forums, and an X thread only for a question where no forum thread passed. Threads AI answers already cite come first, then the ones Google ranks for the question, then the newest. Every thread is checked before you see it. It must take replies (Reddit's and Stack Exchange's own APIs say so where they answer; otherwise the page is read, and a page that could not be read is never assumed open), be from the last 6 months, and be judged by your own AI to be someone choosing, comparing or asking for a recommendation for what you sell, or struggling with the exact job it does. General how-tos are left out even when AI answers cite them. One line says how many were left out and why, with "Show them"; when nothing passes, the panel says "Nothing worth joining this week". Each thread shows why it was picked, how old it is and the angle to take from your own work. "I replied" records your reply as a placement, watched for citations like any other. "Not for me" and "Comments closed" set a thread aside for this project; it can be brought back. Refresh searches at most 5 questions, one search each and one more on X where no forum thread passed - at most 10 searches - and a question is searched again only after 7 days, at most 10 questions a week per project. The searches use your Serper or DataForSEO key (ours where you have neither, on Refresh only), and judging uses your own AI key: the Refresh button says what judging costs before you press it. Without an AI key of your own, nothing is searched or judged, and the panel says why. The same refresh can also run by itself every Monday, but only for a project where you switch on "Find new conversations every Monday", under Refresh on this panel. It is off for every project until you do, and it runs only where your account has both its own search key and its own AI key - never ours. Each Monday run makes at most 10 searches and judges at most 30 threads, within the project's 20 searches a week, and the switch shows the same judging cost as the Refresh button. Refresh works whether it is on or off. A thread judged once is not judged again for 30 days unless its own date changes. Below the conversations, Pages to pitch lists the pages other people published that those answers cited, ranked by how many of them cited each page, each with one action - get listed, get reviewed, ask to be mentioned or pitch an update - and your own page's Google position for the topic. A publisher can be owned by a competitor without saying so: "It's a competitor" labels the whole site Competitor, the same label as on Sites AI cites, and every page from it leaves the list (Undo puts back the label it replaced). It is not offered on sites where anyone publishes, such as YouTube, LinkedIn, Reddit or Quora. Your own sites, competitors' sites, reference sites such as Wikipedia, pages we published for you and pages you recorded are left out, and so is a site no one has labelled when an answer that cited it put that site's own business forward by name. Each measured topic is also read as Winnable, Contested or Long shot, with the reasons and the counts behind them: whether your page is in Google's top 10, what Google's AI Overview did on that search, whether you were named, whether the questions name your town, and what kinds of sites the answers cite. It is a suggestion only; your topics are never changed.
If a rate here ever appears without the count it was taken out of, treat it as a bug and tell us.