When teams first measure how they show up in AI answers, they almost always start by counting mentions: how many times does ChatGPT name us, how often does Perplexity cite us. That number is worth having, but on its own it is a vanity metric. It tells you about the ground you already hold and nothing about the ground you are losing. The metric that actually drives work is the citation gap, the set of specific questions where a competitor gets cited and you do not. Tracking mentions and tracking gaps are different jobs, and the second one is where the value is. Here is how to run both.
What a citation gap actually is
A citation gap is not “we do not get mentioned enough”. It is precise: for a specific prompt, on a specific engine, a competitor is cited as a source and you are absent. Each gap is a single, concrete, fixable thing. That precision is what makes gap analysis so much more useful than a mentions count. A mentions number going down tells you something is wrong but not what to do; a list of gaps is a worklist.
Think of it as the difference between “our market share fell” and “we lost these five accounts to this competitor”. The second is actionable because it names the specific losses. Gap analysis does the same for AI visibility: it turns a vague worry into a named list of prompts to go win.
Two things to track, and why they differ
Track mentions to understand your current footprint: where you already appear, how you are described, and whether that footprint is growing or shrinking over time. Track gaps to understand your opportunity: the specific questions you are losing and to whom. They answer different questions and they drive different actions. Mentions tell you whether your existing content is holding its ground; gaps tell you where to build next.
A healthy program watches both. If mentions are flat but gaps are growing, competitors are expanding into questions you have not covered. If mentions are falling on prompts you used to win, that is a defensive problem, closer to the diagnostics in why is ChatGPT citing my competitor instead of me. The two metrics together tell a story neither tells alone.
Why gaps are per-engine, not global
The single most important rule of gap analysis is that gaps are per engine. A prompt where you are proudly cited in ChatGPT can be a total gap in Perplexity or Gemini, because each engine weights and sources differently. Profound’s analysis of citation patterns across AI platforms documents how much the cited set varies from one platform to another (Profound), and Ahrefs’ study of the most-cited domains in ChatGPT shows a very particular source profile, with Reddit and Wikipedia dominating across 9.6 million queries (Ahrefs). Different engines, different winners.
Practically, this means you never average across engines. An overall “we are cited 30 percent of the time” hides that you are at 60 percent in one engine and near zero in another, and the near-zero engine is exactly where your attention should go. Always break gaps down by engine.
Step one: define the prompt set that matters
Everything starts with the right prompts. A gap analysis is only as good as the questions you feed it, so build a set of the prompts your actual buyers ask, weighted toward high-intent, bottom-of-funnel questions. Include the obvious category questions (“what is the best tool for X”), the comparison questions (“X versus Y”), and the problem questions your product solves. Keep the set tight and meaningful rather than sprawling, because every prompt you add multiplies the tracking cost, a dynamic explained in why AI visibility tools cost so much.
Step two: capture mentions across engines
For each prompt, run it on each engine you care about and record who gets cited, including yourself and your competitors, and how each is described. Because AI answers are non-deterministic, a single run is unreliable, so sample each prompt a few times and aggregate. What you want out of this step is, for every prompt-and-engine pair, a list of the sources cited and where you sit in it, if at all.
This capture is the raw material. Do it consistently, on a schedule, so you can see change over time rather than a single snapshot. A one-off capture tells you today’s state; a repeated one tells you the trend, which is what you actually manage.
One practical caution at this step: be disciplined about phrasing your prompts the way real users do, not the way you wish they would. If you seed the capture with prompts that name your product or use your internal jargon, you will flatter yourself with mentions that no real buyer would ever trigger. The whole value of gap analysis depends on the prompt set reflecting genuine buyer language, so write the prompts as questions a prospect who has never heard of you would type. That discipline is what keeps the exercise honest and the gaps real.
Step three: map the gaps
Now turn the capture into a map. Lay prompts down the rows and engines across the columns, and mark each cell: are you cited, is a competitor cited, both, or neither. The cells where a competitor is cited and you are not are your citation gaps. The cells where nobody in your set is cited are open territory, a different and often easier opportunity. This grid is the single most useful artifact in the whole exercise, because it makes the gaps literally visible.
Step four: diagnose each gap’s cause
A gap is not a diagnosis, it is a symptom. For each meaningful gap, work out why you are absent. The common causes are: the engine cannot crawl your relevant page, so confirm the crawler is allowed, since OpenAI documents that its OAI-SearchBot must be able to reach a page for it to surface (OpenAI); you have no content that directly answers that specific question; your content exists but is not authoritative or specific enough to be chosen; or you are simply outcompeted by a stronger source. Ahrefs, studying tens of thousands of brands, found that AI visibility tracks with authority and relevance signals rather than being random (Ahrefs), so a gap often traces back to one of those signals being weaker than a competitor’s.
Gap type, cause, and action
This table maps the common gap patterns to what to do about them.
| Gap pattern | Likely cause | Action |
|---|---|---|
| Competitor cited, you have no page on the topic | Coverage gap | Create a direct, specific page for that question |
| You have a page but it is never cited | Authority or clarity gap | Sharpen the answer, build authority signals |
| Cited in one engine, absent in another | Per-engine weighting | Study what that engine favours and adapt |
| Nobody in your set is cited | Open territory | Publish first and own the question |
| Page exists but engine cannot read it | Crawl access | Unblock the crawler, fix indexability |
Step five: prioritize by value, not volume
You will find more gaps than you can close, so do not work them in the order you found them. Score each gap by the business value of the question behind it. A gap on a high-intent, ready-to-buy question is worth many gaps on low-intent informational ones. It is far better to close the five gaps that sit on your most valuable buying questions than to chip away at fifty trivial ones. Volume is a trap; value is the guide. This is the same share-of-voice logic explored in how to increase ChatGPT share of voice, applied at the level of individual questions.
Step six: close the gap and re-measure
Closing a gap means becoming the source that engine will choose for that prompt: a crawlable, authoritative, current page that answers the specific question more directly than the competitor who currently holds the slot. Then re-run the capture and watch the cell flip from gap to mention. This close-and-re-measure loop is the whole point; it turns gap analysis from a report into a growth engine. Without the re-measure, you never know whether the work landed.
Track mentions and sentiment, not just presence
One refinement worth adding: when you capture mentions, record not just whether you appear but how you are described. Being mentioned inaccurately or unfavourably is its own problem, distinct from being absent. A brand can have strong presence and poor sentiment, which no simple mentions count would reveal. Tracking the framing alongside the presence gives you a fuller picture and sometimes surfaces work more urgent than closing an absence, a nuance covered in are ChatGPT citations worth tracking.
A worked example
Say you sell CRM software. You define 40 buyer prompts and capture across ChatGPT, Perplexity, and Gemini, sampled three times each. The grid shows you are cited well in ChatGPT for “best CRM for startups” but absent in Perplexity for the same prompt, and absent everywhere for “CRM with the best API”. You diagnose: the Perplexity gap is an authority issue, since your page exists but a rival’s is more referenced; the API gap is a coverage issue, since you have no page that directly answers it. You prioritize the API gap because it is a high-intent question your product wins on, publish a specific, authoritative page, and re-measure two weeks later to find you are now cited in two of three engines for it. That is gap analysis doing its job: a named loss became a named win.
Notice how much better that outcome is than what a mentions-only view would have produced. A mentions dashboard would have shown a modest overall number and left you guessing where to invest. The grid told you exactly which question to fix and on which engine, so a single, well-chosen page moved a metric that matters instead of a dozen scattered edits moving nothing. The leverage comes entirely from having mapped the gaps rather than just counted the mentions, which is why the mapping step is worth the effort even when it feels tedious.
Common mistakes
The biggest mistake is tracking only mentions and never gaps, so you feel informed but never know what to fix. The second is averaging across engines and missing that your worst engine is where the opportunity is. The third is working gaps by volume instead of value, closing many cheap ones while the valuable ones stay open. The fourth is capturing once instead of on a schedule, so you see a snapshot but never a trend and can never prove a fix worked. Related single-engine competitor tracking is covered in track competitor citations in ChatGPT.
The bottom line
Mentions tell you where you are; gaps tell you where to go. Track both, but let the gaps drive the work: define the prompts that matter, capture across engines, map where competitors win and you are absent, diagnose each gap, prioritize by value, close, and re-measure. Keep it per engine, because a win in one is not a win in another. Done this way, AI visibility stops being a number you watch and becomes a list of specific questions you go out and win.
Frequently asked questions
What is the difference between an LLM mention and a citation gap?
A mention is any time an AI answer names or cites you. A citation gap is a specific question where a competitor is cited and you are not. Mentions measure your current footprint; gaps measure your missed opportunities, which is the more actionable of the two because each gap is a concrete thing to go fix.
Why track gaps per engine instead of overall?
Because citation behaviour differs sharply between engines. A prompt where you are cited in ChatGPT can be a gap in Perplexity or Gemini, since each engine weights sources differently. An overall average hides these, so you track per engine to know exactly where to act.
Do I need a paid tool to find citation gaps?
Not to start. You can run your key prompts by hand across the engines and note where competitors appear and you do not. A tool scales this to many prompts, engines, and repeated samples over time, which is why they cost what they do, but a manual pass on your top prompts is a valid, free way to begin.
How do I prioritize which gaps to close first?
By business value, not by how many gaps exist. A gap on a high-intent, bottom-of-funnel question your buyers actually ask is worth far more than gaps on a dozen low-intent informational prompts. Score gaps by the value of the question, then work top down.
