Run twenty buyer questions every week, in a few phrasings, across two AI engines and you get a map with three regions: questions where your brand is cited, questions where a rival is cited instead, and questions where nobody in your category is cited. The second region is the one teams act on first, and that is backwards, because a gap turns into a content brief only when the citation that would fill it could land on a page you own.
Most of them cannot. Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control: editorial coverage and community threads. The audit still earns its afternoon: G2 found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers now start research with an AI chatbot more often than with Google, and one in three buyers said an AI chatbot led them to choose a vendor they hadn't considered. The waste happens in the next step, when a spreadsheet of absences becomes a publishing calendar.
What does a citation gap map actually show?
Record three fields per question: who was named, whose page was used as a source, and which domain that source sits on. The third field decides whether any of this is your job.
The three regions behave differently enough that treating them alike is the core mistake. Where a rival is cited, start with where the citation lives, because a competitor quoted from their own documentation and a competitor named inside a review roundup are unrelated problems. Where nobody in the category is cited, the engine assembled the answer from generic material, and that is the only region with open ground in front of it.
Keep the rungs separate while you record. A mention is your name in prose, a citation is your page carrying the answer, and a recommendation is the engine telling the buyer to choose you. The IAB's 2026 measurement guidance separates presence, prominence, portrayal, and persuasion for the same reason, and groups those measures into four sets for brands. A gap on one rung does not imply a gap on the others, so one combined figure cannot tell you where to work.
How many runs before a gap counts as a gap?
Enough that the difference you are about to spend money on exceeds the noise between runs. SparkToro's research found AI engines are highly inconsistent when recommending brands, with the same prompt returning different vendor lists on repeat, a result Search Engine Land summarized as recommendation lists rarely repeating exactly. A single absence is a sample, not a finding.
Treat every cell as a rate with an interval around it. A 2026 variance-components study found run-to-run variance in AI visibility measurement, and work on quantifying uncertainty in AI visibility reaches the same conclusion by putting confidence intervals on visibility measures. Our own floor is several weeks of answers per question, across phrasings and engines.
The rule that falls out of this: rank the queue by the size of a gap relative to its own variance. Appearing in one run of ten while a rival appears in eight is worth a decision. Trading places between Tuesday and Thursday is worth another month of observation and nothing else. Category-level claims rest on corpora like Semrush's AI Visibility Index and its 126 million AI search prompts, a different instrument from a twenty-question panel.
Record the engine and the date with every run, or the map cannot tell you whether a gap closed.
Which questions belong in the set you measure?
The questions a buyer asks before choosing, in the words they use. The keywords you already rank for are a different list, built for a different machine. In our experience ten to thirty questions is where a weekly cadence survives contact with real work; past that, the running eats the analysis. Mix category questions, direct comparisons, and job-to-be-done phrasings that never name a vendor.
The three types fail differently, which is why the map is useful. Comparison questions tend to be answered from third-party sources. Job-to-be-done questions go unanswered by name most often, which is where an unowned region opens. Category questions are where prior associations dominate, and a 2026 analysis of brand dynamics in LLM recommendation systems found how thin that dominance is: well-known brands won every recommendation only when products were otherwise identical, and a competitor with less than a 0.1-star rating advantage erased the edge.
Size the set against how buyers actually decide. 6sense found buying group members already know 75% of the vendors on their shortlist, so the questions that matter are the ones that assemble that list. In Profound's 7.5 million-conversation en-US sample, the share of ChatGPT conversations classified commercial rose from 13.9% to 19.2% in a year, so their supply is growing faster than most panels are. Profound's "more than doubled" volume figure is an extrapolation from estimated weekly users.
How do you tell an addressable gap from one that is not?
Read the citations under the answer, then ask whether a page you could publish would have been eligible to appear there. That single test does more triage than any scoring model, and it is the reason a visibility gap is not by itself a content brief.
Four checks, in order, resolve most of the map. First, is the answer built on domains you cannot publish on? If a review platform and two forum threads carry it, your site was never in the running, the shape Profound's citation-source research describes.
Second, can the engine read your pages? Vercel's crawler research with MERJ documented that the AI crawlers it measured did not execute JavaScript, and Cloudflare's agent-readiness analysis of the 200,000 most visited domains found few sites have adopted the standards built for agents: 4% declare AI usage preferences and 3.9% serve Markdown on request. That is an engineering ticket, not a brief.
Third, does the engine resolve you as a distinct company? A brand split across three names makes it harder for readers and engines to attribute claims to the right company, so settle on one name and one description. Fourth, and only then: is there a claim you can prove that the current sources do not carry? That is the addressable gap, and it is a minority of the map.
Be skeptical of the tactic layer while you do this. C-SEO Bench (NeurIPS 2025) tested conversational-SEO methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. The strongest case left is for evidence: the GEO study published at KDD 2024 found that adding statistics, quotations, and citations raised citation visibility by roughly 30 to 40% on its benchmark, while keyword stuffing did nothing, though C-SEO Bench found those evidence methods rarely moved citation rank either. Some framing actively hurts, since scarcity and exclusivity framing measurably reduced recommendation rates in a controlled test on fictitious products.
How should the queue be ranked?
By addressability first, then by evidence you already hold, then by traffic value. Reversing that order is how teams publish four articles against a gap a review site owns.
Three inputs decide the top of the list. Whether the question is retrievable ground for you, which Seer's finding that 87% of SearchGPT citations matched Bing's top results makes concrete: if nothing of yours surfaces, the citation is decided among pages that did. Whether you hold provable material, since AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited: being retrieved is not the same as being worth quoting. And whether the gap survived the variance test.
Everything else belongs in one of two side lanes. Third-party displacement goes to whoever owns reviews and analyst relations, with the domains named from your own citation record. Fit gaps, where the rival is honestly the better answer, go to product and positioning.
Which analytics should decide what you publish next?
Only the ones that move when you change something you control. In Search Console, impressions, position, and clicks carry precise published definitions, and the performance report documents what it collects, which covers Google's surfaces and never separates an AI Overview appearance from a classic result. Position earns a vote only as evidence that a page is indexed and eligible in classic results.
Click metrics tempt hardest and hold up worst. Ahrefs' December 2025 re-run measured a 58% lower average clickthrough rate for the top-ranking page when an AI Overview appears, and Pew recorded clicks on a traditional result in 8% of visits with an AI summary against 15% without, an association in browsing data rather than a measured cause. A page can lose clicks in the quarter it wins the citation you queued it for.
Referral mix fails the same test. Similarweb tracked the share of ChatGPT referrals landing on homepages moving from roughly 26-29% to about 62-63% after the May 2026 update, a shift no publisher caused. Before a metric gets a vote, write down what would have to change for it to move; what survives enters the queue at the traffic-value tier above, never ahead of addressability.
How does Trovance turn a citation gap into a queue you can defend?
Trovance runs your tracked questions every week across AI engines and preserves each answer run as a snapshot with its citation record, so the map persists as a standing artifact across cycles. Share of answer across that record produces the three regions. Because every snapshot keeps its sources, you can see whether a competitor won on their own documentation or on a domain neither of you controls.
Variance is handled by repetition. Runs accumulate across cycles, so a gap is compared against its own history before it counts as a finding, and a question that flips week to week reads as instability. The Brain keeps the claims you are entitled to make with the proof behind each one, which is what turns an addressable gap into a specific brief.
Recommended actions carry the reason they exist, including the reason not to write. When the record shows an answer was assembled from third-party domains, the action names those domains instead of proposing a page, which keeps the queue honest about what publishing can reach. Drafts use your author voice guidance and research sources, and a person reviews and approves everything before it ships.
What Trovance will not promise is that a queued brief earns the citation. No system can, because engines are probabilistic and the sources they trust belong to other people. There is no universal visibility score here and no control over what a model says. The analysis cycle reruns the same questions after work ships, so the next decision is made against current evidence.
What should you do this week?
Start with twenty buyer questions, two or three phrasings each, run weekly on two engines. Record mention, citation, source domain, and the rival named, and resist acting until the sheet is full. The first pass is a base rate, and everything downstream depends on it being one.
Then sort into three lanes: answers resting on domains you cannot publish on, pages that existed but were unreadable, and questions where the rival is the honest answer. What survives is the queue, and it will be shorter than the audit implied.
Be plain about timing. Expect retrieval fixes to resolve faster than earned third-party presence, and expect the measurement noise above to stay. Neither interval is something anyone can quote you in advance. If you would rather not run the panel by hand each week, start a free Trovance trial and let the citation record build itself while you work the queue.
Measure the gap before you queue it
A visibility gap is not a content brief - the argument this method builds on.
How to track AI citations - the logging discipline behind a trustworthy map.
The prompt panel is measuring the wrong question - why question selection decides what you can see.
How to measure AI search visibility without one score - what to report when one composite hides the decision.
Decide what the gap justifies
Where AI citations come from - how much of the answer sits on domains you never own.
What gets cited by AI - the page traits that make a brief worth writing.
AI agents need evidence, not more content - why proof beats volume on an addressable gap.
How to build a competitor comparison page - the asset a rival-cited gap most often justifies.
