How to run an AI citation gap analysis

What a citation gap analysis actually is

It is three steps: collect the domains assistants quote when they answer questions in your category, check which of those pages mention you, and decide which of the rest you could realistically get onto.

The unit of work is a page, not a keyword. Asked “who are the best bookkeepers for a small agency”, an assistant reaches for a handful of sources — a directory, a comparison post, a forum thread — and the names it recommends are largely the names it found there. The gap is the distance between that set of pages and the set your business appears on.

None of the three needs a tool. The first is copying and pasting, the second is searching each domain for your own name, and the third is judgement nobody can automate for you.

Why a citation gap is not a backlink gap

A backlink gap is a list of pages that link to a competitor and not to you. A citation gap is a list of pages an assistant was willing to read.

Those lists overlap by coincidence, and three differences explain why. A cited source does not have to link to you. A listing that marks every link nofollow, or carries no link at all, is worth nothing to a link index and plenty here: your name is on a page the assistant fetched. An assistant can quote a page that never mentions you, and can quote one that does and still recommend somebody else. And authority does not transfer: a strong page the assistants never reach for is not in the set, while a thin forum answer they keep quoting is.

A second distinction sits underneath that one. An assistant can say your name from training data without opening a page: nothing was fetched, and nothing you publish this month is the reason. That is being mentioned. Quoting is different — the answer carries an address it actually used as a source. This analysis is built from the second kind only.

How do you run one by hand?

A browser, a spreadsheet and an afternoon will get you a real citation gap list, and doing it once by hand is the only way to know what an automated version reports.

1. Write ten questions a customer types before buying — “best X for Y”, “how much does X cost”, “X alternatives”. Your brand name tests recall; these test the shortlist.
2. Put each question to the assistants you care about, five times each. Five is the smallest number that shows variation. Ten questions across two assistants is a hundred prompts, about two hours of copying and pasting.
3. Record one row per cited domain per question, with the date, the assistant and the run. Count a domain once per question even if one answer quotes it three times.
4. Then do the half a tool will not do: open each domain and search it for your own name. Mark each row on it, not on it or cannot tell. That is the hour that turns a list of sources into a list of gaps.

Keep the questions in a file and do not change them next month. The question set is your instrument, and an instrument adjusted between readings cannot tell you whether anything moved.

Reading the tally without over-reading it

The number worth sorting on is how many distinct questions a domain turned up in, not how many times it appeared in total.

A domain quoted while answering eight of your ten questions is in the set the assistant reaches for about your field. One quoted once might be in that set, or might be the oddity any sample of a hundred answers throws up. Sorting by distinct questions puts the stable sources at the top, and that ordering is the output of the analysis.

Two columns earn their place beside the count. What kind of site it is — directory, reference, regulator, very large platform, competitor — because the kind decides whether a route onto it exists. And which assistant quoted it, since a source only one of them reaches for is a narrower bet than one they both use.

Why one run is one observation and never a rate

Ask the same question twice and the cited sources can differ, so a single run tells you what happened once.

A domain observed on one run must not be written down as a percentage, a bar, or a figure like “0 of 1”. A rate needs repeated observations, and rendering an unrepeated one as a rate dresses a snapshot as a trend. Say it plainly: quoted on the one run it was asked, on this date.

Once you have repeats, state the denominator next to the number. Quoted in 6 of 10 questions, over 5 runs each is a sentence a reader can argue with. 60% citation share is the same figure with the sample hidden, and the hidden sample is where the error lives.

What “not measured” means, and why it is not a zero

A question that did not return an answer produced no data, and recording it as an absence invents a measurement nobody made.

Four ordinary things leave a row unrun: the call failed, the assistant refused, you hit a rate limit, or you reached the end of what you would spend that day. All four are the same state — not measured, with its reason — and all four stay out of the denominator. A failed check reported as a zero is the most damaging mistake available here, because the zero looks exactly like a finding.

A genuinely empty result is different again, and worth reading twice: a business whose own pages are the thing being quoted will see a short list for a good reason. An empty list and an unrun check look identical in a spreadsheet unless they get separate columns, so give them separate columns.

Which gaps are worth chasing?

A gap earns your time when the source is something you can get onto, which is a much shorter list than the tally.

The reachable kinds

• Directories and listing pages that take submissions. The dullest entry on the list and reliably the most actionable.
• Roundups and comparison posts whose author accepts a pitch or a correction. A checkable reason you belong beats a request to be added.
• Forums and question-and-answer pages anybody can post to. Slow, cheap, and the answers that get quoted are the ones that answered properly.
• Register and reference pages with a correction route — a trade body’s member list, a regulator’s register you qualify for but never filled in.
• A page nobody has written. Sometimes nothing in the set answers the question well. That is a brief for your own site, and the one gap you control.

The ones that are not

• A competitor’s own domain. Information, not a task. Record which rival the assistants reach for, and move on.
• A paywalled publisher, unless buying that placement was already a thing you intended to do.
• A very large platform — a marketplace, a social network, a video host. Presence there is a staffing decision, not an outreach one.
• Anything whose only route is paying for a link. A separate decision with its own risks, which this analysis did not find a reason to take.

What a gap does not tell you

Getting onto a cited source does not mean an assistant will then cite you, and nobody can honestly promise it will.

Three things have to happen after your name lands on that page, and none is yours to control. The page has to be fetched again. The assistant has to reach for it on that question rather than another. And your name has to survive a two-sentence summary that names three businesses at most. A source can also drop out of the set between your pitch and the next run.

The analysis says nothing about quality either: a page appearing in answers is not a judgement that the page is good. What it settles is narrow and useful — these are the pages the assistants reached for about your field, and on these ones your name is absent. The rest is ordinary outreach work with ordinary odds.

Doing the collection with our citation source finder

Our AI Citation Sources tool does the first of the three steps: it writes the questions, puts them to the assistants and lists every address the answers quoted.

You type a website address and nothing else. It reads the homepage to work out what you sell, writes up to five questions a customer of yours would ask, and puts each to ChatGPT and Gemini once. Back comes the list: each quoted domain, what kind of site it is, and how many of the questions it turned up in. Your own domain is dropped, and addresses that look like rivals go on a separate competitors list — a domain is on one or the other, never both.

It separates a short list three ways, for the reasons above: the domains it found, none recorded when the questions ran and carried nothing quotable, and not measured, with its reason, when a check could not complete. It does not do the second step — whether you appear on each domain is a judgement you make by opening them.

What a citation-tracking tool can and cannot do

A tool buys the collection and the repetition. It does not buy the judgement about which gaps are reachable, and nothing in this category can.

What one adds is volume and persistence: more questions than you will ask, on a schedule nobody keeps by hand, with the history kept — which is what makes a second month mean anything.

Four questions are worth putting to any of them, and all four are answerable from a vendor’s own pages. Which assistants does it ask, by name? How many times is each question asked, over what period? Does it keep quoted and merely named apart? And does every number carry the count of answers behind it? A rate printed without its denominator cannot distinguish a small sample from a real movement.

Tracking the same citation sources over time

One collection is a snapshot; the useful artifact is the same question set re-collected on a schedule, because the sources move on somebody else’s timetable.

A second collection gives you three things a first cannot. A domain that has entered the set — new, and easiest to get onto while it is. A domain that has left it, which may mean nothing, since one absence is a small sample, and means something once it persists. And the domains that survive every collection: that stable core is the only part worth a long, expensive campaign, because it is the only part with evidence it is durable.

Two rules keep the comparison honest. Keep the question set fixed, or you cannot tell whether the sources moved or your instrument did. And date every collection — a six-month-old source list in a field this young is a historical document, not evidence.

A monthly routine

Once a month is the right frequency, and after the first pass it fits in about an hour.

1. Re-run the fixed question set and record the date. Same questions, same assistants, same number of runs each.
2. Diff this month’s tally against last month’s. New domains, departed domains, the stable core. The third is the list to act from.
3. Do the work on exactly one reachable gap. One listing submitted, one correction sent, one page written. A month with four started and none finished has nothing to show.
4. Write down what you did and when — not because it proves anything alone, but because a dated change and a dated movement can be lined up.

Nothing in that routine demonstrates that your change moved a number. It buys a record in which cause and effect are both written down, which is more than a dashboard with no log beside it offers.

FAQs