Glossary
Citation Gap
A citation gap is the set of sources an AI engine relies on for answers in your category that you neither own nor appear in. Closing the gap — earning coverage or accurate mentions on those third-party domains — is often the fastest way to change what AI says, because those pages already feed the answers.
When an engine answers a category question, it leans on a fairly small set of retrieved sources: review aggregators, comparison articles, community threads, industry blogs. If your brand is absent or misrepresented on the sources that dominate those citations, no amount of on-site optimization fully fixes the answer — the engine is reading someone else's pages.
A citation-gap analysis ranks the non-owned domains by how many relevant answers they influence, then checks your presence on each. The output is effectively an outreach target list ordered by AI impact: update the G2 profile, get added to the two comparison posts engines keep citing, answer the recurring Reddit thread accurately.
This reframes digital PR for the AI era. Link equity mattered for rankings; citation presence matters for answers. The domains worth pursuing are not the highest-DR ones in the abstract but the ones engines demonstrably retrieve for your buyers' questions.
How to run a citation gap analysis
A citation gap analysis answers one question: which pages that you do not control are feeding the AI answers your buyers see, and where are you missing or wrong on them? It takes a focused afternoon, needs no special tooling, and the output is an outreach list ordered by how much each source actually moves answers in your category.
-
Write 15–30 prompts your buyers would actually type
Not category keywords — questions. Mix four kinds: category ("best X for Y"), comparison ("A vs B for a 20-person team"), pricing ("is A worth it over the free tier"), and trust ("has A ever had a breach"). Write them the way a person types, with the qualifiers that make the answer specific. The quality of everything downstream is set here: a prompt set of generic category terms produces a citation list of generic category pages.
-
Run each prompt in a clean session and save the sources, not the answers
Use a fresh session or incognito window per prompt so personalisation and conversation history do not contaminate the result, and run the same prompt in each engine you care about. Most people record the answer text; for this analysis the citations are the payload. Capture every cited URL, its domain, and which prompt and engine produced it. Run the set twice on different days — AI answers are non-deterministic, and a source that appears both times is a real pattern rather than a coin flip.
-
Build a domain frequency table
Collapse the URLs to domains and count how many distinct prompt-and-engine pairs each one influenced. This is the whole analysis in one table. What usually emerges is a short head — typically five to fifteen domains carrying most of the citations, often review aggregators, one or two comparison posts, a community thread, and a handful of industry publications — and a long tail that appears once and can be ignored.
-
Split the table into owned and non-owned
Your own domain's citations tell you which of your pages engines trust; keep that list, it is useful separately. The non-owned rows are the citation gap candidates. Resist the urge to act on your own pages first — if engines are answering from someone else's comparison post, rewriting your homepage will not change the answer.
-
Score your presence on each non-owned source
Open each one and classify it in three buckets: absent (you are not on the page at all), present but wrong (listed with outdated pricing, a missing feature, a stale positioning line), or present and accurate. "Present but wrong" is the most valuable bucket and the most commonly skipped — those pages are already feeding answers, so a correction propagates without you needing to earn anything new.
-
Weight by how close the prompt sits to a decision
A source cited on "what is generative engine optimization" is worth far less than one cited on "A vs B pricing", even if the first appears more often. Tag each prompt by intent — definitional, category, comparison, pricing, trust — and weight each domain by the intent of the prompts it was cited on. Ranking purely by citation count sends you after encyclopedic sources that buyers read long before they are choosing.
-
Order the outreach list by impact over effort
Rank each non-owned domain by weighted citation frequency divided by how hard it is to change. A review profile you can edit today outranks a journalist's article you would have to pitch. A recurring community thread you can answer honestly outranks a competitor's blog post you will never appear in. The output is a ranked list of concrete actions, not a score.
-
Re-run the prompt set in four to six weeks and diff the citations
This is the step that makes the analysis a methodology rather than an audit. The only valid proof that closing a gap worked is that the answer changed — a new source cited, a corrected claim, your name appearing where it did not. Track the same prompts, same engines, and compare citation sets. Engines re-crawl and re-weight on their own schedule, so give it weeks, not days.
Everything above is runnable by hand with a spreadsheet and a couple of afternoons. Perciva automates the capture-and-diff loop — steps two, three and eight — but the method does not depend on it, and the weighting in steps six and seven is a judgement call no tool should make for you.
Frequently asked
What are citations in AI answers?
Citations are the sources an AI engine retrieved and used to compose its answer, usually shown as linked references beside or beneath the text. They matter because they reveal where the engine's claims about your product actually came from — which is often a third-party page rather than your own site.
How is a citation gap different from a backlink gap?
A backlink gap asks which domains link to your competitors but not to you, and it is about ranking equity. A citation gap asks which domains AI engines retrieve when answering your buyers' questions, and it is about what the answer says. The two lists overlap much less than people expect: a high-authority domain that no engine retrieves for your category is irrelevant to the citation gap, and a modest community thread that engines keep citing is near the top of it.
How often should a citation gap analysis be repeated?
Quarterly is enough for most categories, with an extra pass after any major model update or after you publish something intended to be cited. Running it more frequently tends to surface noise rather than change, because engines re-weight sources on their own schedule.
Do I need a tool to do this?
No. The method is a spreadsheet exercise: prompts in one column, cited domains in another, counted and weighted. Tooling helps with the repetitive parts — running the prompt set on a schedule and diffing citation sets between runs — but the analysis itself is judgement, and the first pass is worth doing by hand so you see what the sources actually say.
Which engines should be included?
Include the ones your buyers use, which for B2B software usually means starting with ChatGPT and Google's AI surfaces and adding others as your own data justifies. Coverage breadth matters less than consistency: the same prompts, in the same engines, compared across runs. A narrow set measured repeatedly beats a wide set measured once.
See what AI currently says about your brand
Perciva runs buyer-intent prompts across ChatGPT and Gemini and shows you the answers verbatim — including which competitor gets recommended instead of you.
Get a free snapshot