AI Citation Tracking

A backlink is a static link on another website's HTML. A citation is a link an AI generates dynamically inside its answer, only when it decides your page is worth sourcing. The same page can have thousands of backlinks and zero AI citations, or vice versa.

Perplexity, Google AI Overviews, Gemini, ChatGPT Search, You.com, and Copilot in Bing show inline or grouped citations most of the time. Claude and ChatGPT (without web) mention brands but often do not link out — track those as 'uncited mentions'.

It depends on the prompt set. On branded prompts, mature brands see 40–70%. On category-defining prompts ('best X for Y'), 10–25% is strong; 30%+ makes you the default source. Track the delta over time, not the absolute number in isolation.

Models re-crawl, re-rank, and re-weight sources constantly. Competitor content ships, PR events shift authority signals, and retrieval indices refresh. Weekly measurement is the minimum to catch drift before it becomes a category loss.

Fix a prompt set, run it on a schedule against every model that matters, and record for each answer whether your domain appeared in the source list, in which position, and alongside which other domains. An AI answer is generated once for one person and never published, so there is nothing to crawl afterwards: a citation you did not capture at the time is gone. That is the whole reason this is a measurement job rather than an audit you can run retroactively.

Citation rate is the number of answers that cite your domain divided by the number of answers you collected, expressed as a percentage. If you run 100 prompts across 5 models you have 500 answers; cited in 62 of them gives a citation rate of 12.4%. Always compute it per model as well as overall, because a blended number hides the surface you are actually losing.

A tool that runs your prompt set against AI models on a schedule and resolves, per answer, which domains were cited. The useful ones record citation position and the competing domains, not just a yes or no, because being cited fifth in a list behaves very differently from being cited first. The thin ones only report what a single model happens to display in its interface, which makes coverage depend on that vendor's UI rather than on your brand.

Yes, and you should start there. Ask ten buying questions in each model by hand, write down which domains appear in the sources, and repeat it next week in the same order. That costs an hour and tells you whether you have a problem worth paying to solve. Free one-shot checkers, ours included, give you a snapshot; what you cannot do free is the repetition, which is where the signal actually lives.

No, and conflating them is the most common measurement mistake. A mention is the model naming your brand in the text. A citation is the model linking your page as the source it drew on. You can be mentioned constantly and cited never, which means the model learned about you somewhere you do not control. Track them as two numbers with two different fixes.

It helps a model parse a page it has already retrieved, which is not the same as helping the page get retrieved. Clean Article, FAQPage and Organization markup makes your definitions, answers and entity unambiguous, so the passage a model lifts is more likely to be the one you intended. It does not compensate for a page that is not authoritative on the topic, and no published study puts a reliable multiplier on any particular schema stack.

Enough that one answer changing does not move your number. Below about 30 prompts a single model update looks like a trend. Most teams settle between 50 and 150, split across branded, category, comparison and problem-led questions. What matters more than the count is that the set is frozen: change the prompts and you lose the ability to compare this week to last.

Because retrieval is not ranking. A page gets cited when a clean, specific passage can be lifted out of it: a plain definition near the top, visible dates, structure a model can chunk without guessing. A thinner site whose pages are better shaped regularly out-cites a stronger one whose answers are buried in prose. Your fix is usually the shape of the page, not more pages.

Manually, budget an hour to build the prompt set and roughly two hours a month to re-run it and record results, assuming 30 to 50 prompts across four models. The cost curve bends when you add models or want per-country and per-language views, because the number of answers to read multiplies rather than adds. That is the point where teams stop doing it by hand.

Yes. Citations are the only AI answer surface with a clickable link — Perplexity and Gemini source cards regularly deliver measurable referral sessions, and cited-page CTR compounds because the model shows you as the trusted source, not one of ten blue links.