How we count a citation
What a citation is here
A citation is one AI answer showing that it used one of your pages. Not that it ranked you, not that it liked you — that a page of yours was behind something it said. That sounds like a yes-or-no question and it is not, which is the whole of this article. An engine can hand back its own list of the sources it used. Or an answer can write your address out in the middle of a sentence. Or it can just say your domain name in passing. Those are three different observations of three different strengths, and for a long time we counted them as one thing.
The three tiers
Declared — the engine returned the page in its own source list. This is the engine's 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 URL on your site. Nobody declared anything, but the model is pointing a reader at the page, and that is worth something. Name in the text — your domain appears in the prose and nothing else. "Tools like yourcompany.co.uk exist" is this. 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 a score is involved, those count 1, 0.6 and 0.25 against each other. Those weights are a judgement. Nothing measured them — they encode 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 hide it inside a number.
Why we keep all three, and never add them up
Keeping only declared citations would be the safer-looking choice and it would be wrong. Some engines never publish a source list at all, so a declared-only count would quietly report those engines as having ignored you. We have made the opposite mistake too: for a period we ignored declared lists entirely, and a client's citation rate read 33% one week and 5% the next with nothing whatsoever having changed on their site. So all three are kept, and they are never summed into a single figure. When you see a citation count you can see what it is made of. If a project's whole score rests on the weakest tier, that is something you should be able to find out in one look — and before this existed, you could not.
The naming rate — the number behind the number
A citation says an engine used a page. It does not say the answer mentioned you. Across our own platform's measurements, of the third-party pages cited in AI answers about our customers, around 37% appeared in an answer that named the business at all, and around 4% in one that recommended it. Your own pages name you every time, by definition. Other people's pages usually do not. The consequence is 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 alongside "how many citations" we show "how many of those sat in an answer that named you", with the number it was counted out of — never a blended percentage, because 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, and anything already on your own competitor list. It will never guess "stockist" or "publisher", because 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. One rule underneath all of it: the kind is an INPUT, 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.
By subject, not just by domain
Every source view used to group by domain, and the subject was sitting on the row the whole time — each answer we probe is stamped with the topic it was asked about. Cut by subject, the same data answers a better question: for THIS topic, who owns it, and which rivals answer it instead of you. Account-wide, "stockists name you and rivals do not" is an interesting fact. Per subject it is an instruction. Topics with no citations at all do not appear, and answers belonging to no single subject — the discovery prompts and the brand-recognition questions — are left out rather than pooled under a heading nobody chose.
If a rate here ever appears without the count it was taken out of, treat it as a bug and tell us.