AI Citation Gap Analysis: Finding the Questions You Never Answered

You check whether AI assistants cite your site. They do — sometimes, for some queries, occasionally in a buried source list. But you cannot tell which questions they would have cited you for if you had answered them, because you never wrote the answer. That blind spot is the gap analysis, and it is the most underused technique in this field.

The gap is invisible from inside the site

A conventional content audit asks what you have and whether it is good. A gap analysis asks: what does the market ask that you have not answered at all?

The distinction matters because the two produce different work. An audit tells you your pricing page is thin. A gap analysis tells you nobody on your site explains what happens to a failed payment, so every assistant asked that question assembled an answer from forums and competitor pages — and mentioned a competitor instead of you. One produces a rewrite; the other produces a new page.

Most teams run audits and skip gap analysis, which is why assistant visibility plateaus. They keep improving answers to questions they already know about, while the questions that generate traffic go unanswered.

Building the question universe

The universe has to come from somewhere other than your imagination. Four sources, in order of quality.

Sales and discovery call recordings. The richest source and the most neglected. Every recorded qualification call contains the questions buyers ask before they even know what your product is called. Transcribe them, extract the questions, strip the conversational filler. What survives is a list of real questions with real commercial weight attached.

Support and ticket history. Tickets are questions that arrived attached to a person. Subject lines and first-line descriptions are almost pure question inventory, and they are usually already grouped by topic.

Search-console-style query data. Pull the queries your site already receives and read them as questions rather than keyword strings. A query like "pricing page export limits" is a question a person typed. Your existing data tells you which questions your domain is already adjacent to — and the gap between adjacent and answered is often small.

Community forums, industry groups, and review platforms. This is where you find the questions nobody thought to ask you. Someone else's customers are asking them publicly, and the answers circulating in those threads are frequently wrong.

Deduplicate aggressively. The same question phrased six ways is one question with six surfaces, and treating them as six will distort your prioritisation.

Classifying and checking coverage

Once you have the universe, classify each question on two axes: can we answer it, and does answering it help.

Answerability has three states worth separating. Some questions you can answer authoritatively — you know the answer and it is publishable. Some you cannot answer without exposing something you should not, such as internal pricing logic or a commitment you cannot make; these need a boundary statement rather than a page. Some you cannot answer because you do not know, and these deserve a research task. Conflating the second and third categories is where teams waste effort.

Commercial value also has a spectrum. A question asked by someone already in your funnel is worth more than one asked with no intent to buy. A question whose answer contains your distinctive vocabulary is worth more than one any competitor could answer identically. And a question left unanswered is actively costing you, because the assistant will answer it with someone else's content.

Coverage, page by page

This is the step most people get wrong. They read their homepage and ask "do I answer this?" The real question is not whether the site answers it somewhere, but whether it answers it in a form that can be extracted.

For each question, locate the candidate answer and judge extractability. A page that discusses the answer across four paragraphs of context, with the actual answer in the middle, is not extractable. A page that states it in one self-contained sentence is. A page that answers it in an image, a PDF, a video, or a form field is not, in practice, even though it is technically published.

Mark each question with one of three states: answered and extractable, answered but buried, or not answered. The middle category has the highest return, because the content already exists and needs restructuring rather than research.

Prioritising by commercial intent

You will have far more questions than you can answer. Prioritisation is the difference between a gap analysis that produces a roadmap and one that produces a list.

A workable scoring approach uses four factors, weighted by what your business rewards: buyer proximity (does the asker look like a customer, a researcher, or someone irrelevant?), answer durability (will this still be correct in a year, or is it tied to a feature that changes quarterly?), distinctiveness (can you answer it better than a generic page could?), and replacement cost (if you skip it, who fills the gap, and how badly does that alternative represent you?).

The first two matter most. High-proximity questions should be answered even when they are undurable, because the traffic and the positioning compound; low-proximity ones should be skipped even when they are easy.

The spreadsheet and prompt-panel workflow

This is deliberately unglamorous, and it is the whole method. Three sheets, one prompt panel.

Sheet one, the universe. Columns: question, source, verbatim quote, buyer proximity, answerability state, distinctiveness note, durability note, coverage page, coverage state, score, owner, status.

Sheet two, the coverage map. One row per page, with the questions it currently addresses and whether each answer is liftable as a single sentence. This stops you from creating a second page for a question you already cover badly.

Sheet three, the publish queue. The prioritised list, with a target format for each item — new page, restructure of an existing page, added section, or FAQ entry.

The prompt panel. Keep four reusable prompts, one per classification job:

1. Classify: here is a question and a candidate passage — is this a self-contained answer that would still make sense quoted alone? Quote the liftable sentence or say none.
2. Rewrite: here is a question and a passage containing the answer — rewrite as a single self-contained sentence, then a three-sentence version, preserving every specific detail and adding no new claims.
3. Gap-check: here is a list of questions and a summary of your site — which questions are not addressed at all, and which are addressed only in a form that could not be quoted?
4. Verify: here is a drafted answer — list every factual claim you cannot support from a source document, and every place the wording implies a commitment or a guarantee.

Prompt four is the safety check. It catches the sentence that slipped from "typically responds within one business day" to "instantly", the most common way answer optimisation damages trust.

Re-running the analysis after each publishing cycle

The gap analysis is not a project. It is a recurring loop with a short cadence: collect new questions, update the universe, re-score the publish queue, publish the top items, re-check coverage on the pages you touched.

The signal that makes this improve over time is assistant behaviour itself. Once you have published, ask the assistants your highest-value questions and record what comes back — whether you are cited, what is said about you, which competitor appears instead. That observation belongs in the universe sheet as a new column, because it turns the analysis from a static document into a feedback loop showing where the real gaps are.

Two habits keep the loop honest. Re-run the whole analysis at a fixed interval even when you are busy, because the question universe moves as your market and product change. And keep the failed predictions: when you thought a question was covered and it was not, record why, because the reason is usually a structural blind spot that will repeat.

Summary

To find the questions you never answered, build the universe from sales calls, support tickets, your own query data, and community forums; classify each question by answerability and by whether you can answer it better than a generic page could; check coverage page by page with extractability as the standard rather than existence; prioritise by buyer proximity and durability first; and run the whole thing on a spreadsheet with a four-prompt panel, including the prompt that strips out unsupported claims. Then re-run it on a cadence, and let what assistants actually say close the loop.

Frequently Asked Questions

How do I find questions my site has never answered? Build a question universe from four sources: recorded sales calls, support ticket subjects, the queries your site already receives, and public forums and review platforms where your category gets discussed. Deduplicate aggressively, then check each question against your published pages. The ones with no candidate answer are your gaps.

What is the difference between a content audit and a citation gap analysis? An audit evaluates what you have published and asks whether it is good. A gap analysis asks what the market asks that you have not addressed, and where your existing content fails to be quotable. The output differs: an audit produces rewrites of known pages, while a gap analysis produces a prioritised queue of new pages and newly extractable answers.

How many questions should a citation gap analysis produce? Enough to be a genuine inventory and few enough to be actionable — for most teams, a few hundred deduplicated questions is realistic, and the discipline is in the prioritisation rather than the collection. If the list runs to thousands, group by intent family first and analyse the families that carry commercial weight, because a long unranked list produces paralysis rather than a plan.