LLM Citation Audit: How to Measure Whether AI Models Actually Cite You

What most guides get wrong

Most guides treat an LLM citation audit as a brand-monitoring exercise. Type your name, count mentions, compare to competitors, done. That approach has three flaws.

First, it audits vanity prompts. Nobody typing your brand name is undecided. The prompts that decide revenue are problem-shaped, not brand-shaped: "how long do I have to file a personal injury claim in Texas" or "is a HELOC tax deductible." Those are where citation gaps hide.

Second, it ignores the retrieval layer. AI Overviews and Perplexity pull from a set of candidate sources before they generate. If you are not in that candidate set, no amount of great content helps.

The audit has to inspect what the model retrieves, not just what it says. Third, it confuses being used with being cited. Your insight can shape an answer while another domain gets the link.

That is the Attribution Gap, and counting mentions will never surface it.

What Is an LLM Citation Audit, Really?

An LLM citation audit is a repeatable diagnostic process that measures how AI answer engines, including ChatGPT, Perplexity, Gemini, and Google AI Overviews, treat your content when generating answers to the questions your buyers actually ask. The word "audit" is deliberate. An audit implies a defined scope, a documented method, and outputs you can review and re-run.

A single screenshot is not an audit. A spreadsheet of one-time mentions is not an audit. What you want is a standing measurement system you can run this quarter, run again next quarter, and use to prove whether anything moved.

In practice, the audit measures four distinct things, and most people conflate them: 1. Presence. Does your domain appear in the retrieval set the model considers before it writes? 2. Extraction. When the model uses your page, does it quote a clean, self-contained claim, or does it struggle to pull anything usable? 3. Corroboration. Is your claim supported elsewhere on the web in a way that makes the model confident enough to repeat it? 4. Attribution. When your material shapes the answer, are you named or linked, or does the credit go somewhere else? These four layers behave differently by vertical. In YMYL categories like healthcare and finance, models rely heavily on corroboration and clear authorship, because the cost of citing a wrong or unverifiable source is high.

A page with a named clinician author and references tends to be treated differently from an anonymous blog. In lower-stakes categories, extraction and presence dominate. The purpose of the audit is to tell you which of these four layers is your bottleneck, because the fix for each is entirely different.

If your problem is presence, you have a discoverability and indexing issue. If it is extraction, you have a structure problem. If it is corroboration, you have an off-page and evidence problem.

If it is attribution, you have an entity and authorship problem. Chasing the wrong fix wastes a quarter.

• An audit is a repeatable process with defined scope, not a one-time screenshot.
• Measure four layers separately: presence, extraction, corroboration, attribution.
• Regulated verticals weight corroboration and authorship more heavily.
• Identify your bottleneck layer before choosing a fix.
• Design the audit to be re-run so you can measure movement over time.
• Different answer engines behave differently; audit at least three.

How Do You Audit the Prompts That Actually Matter?

The single biggest error in citation auditing is testing the wrong prompts. This is where the Retrieval Reality Test comes in, a method I built to force the audit onto the questions that decide purchases rather than the ones that flatter you. The test has three steps. Step one: build a decision-stage prompt set. Do not start with your brand.

Start with the buyer's problem at the moment of decision. For a personal injury firm, that is prompts like "what to do after a rear-end collision in Houston" or "how much is a whiplash claim worth." For a wealth manager, it is "should I roll my 401k into an IRA when I change jobs." These are the answers where a citation converts. Pull them from real search data, from your intake team's most common questions, and from the sales objections you hear repeatedly. Step two: run each prompt across surfaces. Take every prompt through ChatGPT, Perplexity, Gemini, and Google AI Overviews.

Perplexity is especially useful because it exposes its cited sources directly, which lets you inspect the retrieval layer without guessing. Record which domains get cited, in what order, and whether yours appears at all. Step three: classify the outcome per prompt. For each prompt, mark it as Cited, Used-but-uncredited, Absent, or Misattributed. That four-way classification is more useful than a yes/no, because it points straight to the bottleneck.

A wall of "Absent" means a presence problem. A wall of "Used-but-uncredited" means an attribution problem. What I've found is that the Retrieval Reality Test almost always reveals a painful gap between brand-prompt performance and decision-prompt performance.

A client can look strong when you type their name and nearly invisible when you type the question their prospect is actually asking. That gap is the real audit finding, and it is the one worth acting on. The hidden cost of skipping this test is quiet.

You never see the prospect who asked Perplexity a question, got a competitor's firm named as the answer, and booked with them instead. There is no analytics event for the citation you did not earn.

• Start with buyer-problem prompts, never brand-name prompts.
• Pull prompts from search data, intake questions, and sales objections.
• Run each prompt across ChatGPT, Perplexity, Gemini, and AI Overviews.
• Perplexity exposes its sources, so use it to inspect the retrieval layer.
• Classify each result: Cited, Used-uncredited, Absent, or Misattributed.
• Compare brand-prompt vs decision-prompt performance to expose the real gap.

What Makes AI Models Actually Cite a Source?

AI models do not cite the best content. They cite the most citeable content, and those are different things. The Citation Surface framework breaks down the four conditions a page has to meet, and it is the backbone of every audit I run. Condition one: Presence. The model has to be able to find you before it can cite you.

This is the retrieval layer. If your page is not indexed, not crawlable by the model's data pipeline, or buried well below the candidate threshold, nothing downstream matters. Audit this by checking whether your URLs appear as cited sources in Perplexity and whether the page is indexed and served for the target query in traditional search. Condition two: Extractability. The model has to be able to lift a clean, self-contained claim from your page.

Content that answers the question in a tight two-to-three sentence block near a clear heading is far easier to quote than a claim buried three paragraphs into a narrative. This is why answer-first structure matters so much. Audit extractability by asking: could a model copy one paragraph from this page and have a complete, accurate answer without the surrounding context? Condition three: Corroboration. In high-trust categories, the model looks for agreement across sources.

A claim that only your site makes, with no external support, is riskier for the model to repeat, especially in medical or financial contexts. Audit corroboration by checking whether your key claims are echoed and, ideally, cited by other credible domains, or whether you are alone. Condition four: Attribution clarity. The model has to know who is making the claim. Named authors with real credentials, clear organizational entity signals, and consistent identity across the web all raise attribution clarity.

Anonymous content gets absorbed into a generic answer with no credit. Audit this by checking your authorship markup, your entity consistency, and whether your organization has a well-defined presence the model can point to. The reason this framework earns its keep is that it turns a vague complaint ("we don't get cited") into a specific diagnosis.

Run each poorly-performing prompt against the four conditions and you will usually find one or two are failing. In my experience, extraction and attribution are the most common bottlenecks, because they are the two most publishers ignore while chasing raw content volume.

• Presence: your page must be in the retrieval candidate set.
• Extractability: claims must be quotable in a self-contained block.
• Corroboration: external agreement raises the model's confidence to repeat you.
• Attribution: named authors and consistent entity signals earn the credit.
• Score every failing prompt against all four conditions to find the bottleneck.
• Extraction and attribution are the most commonly neglected conditions.

Why Do AI Models Use Your Content Without Naming You?

The most frustrating audit finding is the Attribution Gap: your insight is clearly in the answer, but a competitor's name is on it. You provided the substance and someone else got the credit. Closing this gap is often where the largest gains hide, because the content already works.

Only the attribution is broken. Here is how the gap forms. When a model synthesizes an answer, it pulls claims from multiple sources and attributes selectively, favoring sources with the strongest identity signals.

If your page and a competitor's page make the same claim, but the competitor has a named author with verifiable credentials, a well-defined organizational entity, and corroborating references, the model tends to attribute the shared claim to them. Your content contributed. Their signals collected the credit.

To audit the Attribution Gap, do this for each Used-but-uncredited prompt from your Retrieval Reality Test: Compare authorship signals. Does the credited source name an author with real credentials? Do you? Anonymous corporate copy competes poorly against clearly-authored expert content in YMYL categories. Compare entity clarity. Is the credited organization clearly defined across the web, with consistent naming, a coherent knowledge-panel-style footprint, and internal linking that reinforces who they are?

Fragmented or inconsistent entity data weakens attribution. Compare evidence density. Does the credited page reference verifiable sources, cite regulations or standards, and show its work? In regulated verticals, evidence density is a strong attribution signal because it lowers the model's risk in repeating the claim. What I've found is that the Attribution Gap rarely closes by rewriting content.

It closes by strengthening the credibility architecture around the content: named authorship, consistent entity signals, and verifiable evidence. This is exactly the kind of work that also keeps content publishable in high-scrutiny environments, which is why I treat citation attribution and editorial review as the same problem viewed from two angles. The cost of an unaddressed Attribution Gap compounds.

Every uncredited use trains the surrounding ecosystem to associate your ideas with someone else's name. Over enough answers, the model's implicit sense of "who is the authority here" drifts away from you, and that drift is hard to reverse once it sets.

• The Attribution Gap is content that shapes answers without being named.
• Models attribute shared claims to sources with the strongest identity signals.
• Audit authorship: named credentialed authors beat anonymous corporate copy.
• Audit entity clarity: consistent, well-defined organizational signals matter.
• Audit evidence density: cited regulations and references lower model risk.
• Close the gap with credibility architecture, not content rewrites.

How Do You Build a Repeatable Citation Audit System?

A citation audit is only useful if you can run it again and compare. A one-time snapshot tells you where you stand today; a documented, repeatable system tells you whether your work moved anything. Here is how I structure it. Fix your prompt set. Lock in 25 to 50 decision-stage prompts from your Retrieval Reality Test.

This set becomes your standing benchmark. Resist the urge to swap prompts between runs, because changing the questions makes results incomparable. Add new prompts as a separate tracked cohort rather than replacing the core set. Define a scoring rubric. For each prompt, record the surface (ChatGPT, Perplexity, Gemini, AI Overviews), the outcome (Cited, Used-uncredited, Absent, Misattributed), and the competing domains that were cited.

Then, for any prompt where you underperform, score the four Citation Surface conditions as pass or fail. This produces a structured dataset, not a pile of screenshots. Capture the retrieval layer where you can. Perplexity's cited sources give you a rare direct look at what the model retrieved. Log those domains every run.

Over time you will see which competitors are entrenched in the retrieval set for your niche, which is more strategically useful than any single answer. Set a cadence. Quarterly is a sensible default for most verticals. LLM behavior shifts with model releases, and the ecosystem of competing content changes slowly, so quarterly captures real movement without drowning you in noise. Record the model versions and dates every time. Report movement, not absolutes. The output that matters is change over time: how many prompts moved from Absent to Cited, how many Used-uncredited results closed into Cited results, which competitors gained or lost retrieval presence.

This is the difference between an audit that informs a strategy and a screenshot that decorates a slide. Building this once is real work. Running it thereafter is straightforward, and that asymmetry is the point.

In my experience, the teams that treat citation auditing as a standing measurement discipline rather than an occasional curiosity are the ones that can actually show a board whether their AI-visibility investment is working. Everyone else is guessing with screenshots.

• Lock a fixed set of 25 to 50 decision-stage prompts as your benchmark.
• Score each prompt with a consistent rubric across surfaces and outcomes.
• Log Perplexity's cited sources to track the retrieval layer directly.
• Run quarterly and always record model versions and dates.
• Report movement over time, not one-off absolute snapshots.
• Add new prompts as separate cohorts to keep the core benchmark stable.

Why Is Citation Auditing Different in Legal, Healthcare, and Finance?

Citation auditing in regulated, high-trust verticals follows different rules, and applying a generic audit here produces misleading results. In legal, healthcare, and financial services, the content sits inside what search and AI systems treat as YMYL, or Your Money or Your Life, territory. Getting an answer wrong in these categories carries real consequences, so models behave more conservatively about who they cite.

The practical effect is that three of the four Citation Surface conditions carry more weight here. Corroboration matters more. A financial claim that no other credible source supports is a claim a model hesitates to repeat with attribution. If you state that a specific tax treatment applies and no established source echoes it, the model may soften the claim, hedge it, or omit your attribution entirely. Auditing corroboration in these verticals means checking whether your claims align with recognized authorities and standards, not just whether they read well. Authorship matters more. In healthcare especially, content attributed to a named clinician with verifiable credentials tends to be treated differently from anonymous copy.

When I audit a medical practice, one of the first things I check is whether the pages making clinical claims are clearly authored and reviewed by qualified people, because that authorship is doing measurable attribution work. Evidence density matters more. Citing the actual statute, the actual regulation, the actual clinical guideline, with correct references, raises the model's confidence that your page is a safe source to name. Vague content with no references competes poorly, even when it is technically accurate. There is also a compliance dimension unique to these verticals.

In legal marketing, jurisdictional advertising rules constrain what claims you can make. In healthcare, patient-facing claims face their own scrutiny. In finance, disclosure requirements shape what you can say.

This is why I insist on Reviewable Visibility: content built to be both citeable by AI and defensible under professional review. Content that games citation at the cost of compliance is a liability in exactly the industries where citations are worth the most. The swap test makes the difference obvious.

Generic advice like "add more content" applies anywhere. But "cite the controlling statute and have a named attorney review the claim for jurisdictional accuracy" only makes sense in a regulated context, and that specificity is precisely what earns citations in these fields.

• Regulated verticals sit in YMYL territory where models cite conservatively.
• Corroboration against recognized authorities carries extra weight.
• Named, credentialed authorship does measurable attribution work.
• Evidence density with correct references raises model confidence.
• Compliance constraints shape what claims you can safely make.
• Reviewable Visibility means citeable and professionally defensible at once.

What I Wish I Knew When I Started Auditing Citations

The lesson that took me longest to accept: citation is a credibility problem wearing a content costume. Early on, when a client was not getting cited, my instinct was to write more and write better. Sometimes that helped. Often it did nothing, because the content was already good enough to be used. It just was not good enough to be named. What changed my approach was watching the same claim get attributed to a competitor with weaker prose but a clearly-defined author, a consistent entity footprint, and references to the actual regulation. They were not winning on writing. They were winning on credibility architecture. Since then, I run every citation audit through the question: is this a content gap, or a credibility gap? In regulated verticals, it is a credibility gap far more often than people expect. That reframe is the difference between a quarter spent producing pages nobody credits and a quarter spent building the authorship, corroboration, and entity signals that actually collect the citation.

Your 30-Day Action Plan

1. Days 1-3 — Build your decision-stage prompt set of 25 to 50 buyer-problem questions using search data, intake questions, and sales objections.
2. Days 4-8 — Run every prompt through ChatGPT, Perplexity, Gemini, and AI Overviews. Record the surface, cited domains, and your outcome classification for each.
3. Days 9-14 — Score every underperforming prompt against the four Citation Surface conditions: presence, extractability, corroboration, attribution.
4. Days 15-21 — Fix extractability first: add answer-first quotable paragraphs under question-shaped headings for your top ten decision prompts.
5. Days 22-27 — Close attribution gaps: add named credentialed authorship, tighten entity consistency, and add verifiable references to key claims.
6. Days 28-30 — Document the whole audit as a repeatable system with a fixed rubric, and schedule the next quarterly run with dates and model versions logged.

Frequently asked questions

How is an LLM citation audit different from a brand mention audit?

A brand mention audit counts how often models say your name, usually in response to brand-shaped prompts. An LLM citation audit measures whether your content is retrieved, quoted, and attributed when models answer the problem-shaped questions your buyers actually type. The difference is significant. You can score well on mentions and still lose every high-intent decision answer to a competitor. The citation audit inspects the retrieval layer, the extractability of your claims, whether they are corroborated, and whether you are named versus merely used. In short, mention audits measure familiarity; citation audits measure whether you earn credit at the moment a prospect is deciding who to trust.

Which tools do I need to run an LLM citation audit?

You can run a solid audit with the answer engines themselves plus a spreadsheet. Use ChatGPT, Perplexity, Gemini, and Google AI Overviews to run your prompt set. Perplexity is particularly valuable because it exposes the sources it cites, giving you a direct view into the retrieval layer that most engines hide. Record results in a structured sheet with columns for the prompt, the surface, the outcome classification, and competing cited domains. Dedicated AI-visibility tracking tools exist and can automate collection at scale, but the method matters more than the tooling. A disciplined manual audit with a fixed prompt set and a clear scoring rubric will outperform an automated tool used without a coherent process behind it.

How often should I run a citation audit?

Quarterly works well for most verticals. LLM behavior shifts with model releases, and the competitive content ecosystem changes gradually, so a quarterly cadence captures genuine movement without generating noise you cannot act on. Always record the date and model versions for each run, because undated results are impossible to compare later. If you are in the middle of a significant content and credibility push, you might run a lighter interim check on your top ten prompts monthly to catch early movement. But the full documented audit, with the complete prompt set and scoring rubric, is best kept to a quarterly rhythm so that changes have time to be reflected in retrieval and attribution behavior.

Why does an AI model use my content but credit a competitor?

This is the Attribution Gap, and it is one of the most common findings in a citation audit. When a model synthesizes an answer from several sources making the same claim, it tends to attribute the shared claim to the source with the strongest identity signals: a named credentialed author, a consistent organizational entity, and verifiable references. Your content can contribute the substance while a competitor with better credibility architecture collects the credit. The fix is rarely rewriting the content, which is usually fine. It is strengthening authorship, entity consistency, and evidence density around the content. In regulated verticals especially, models favor sources they can trust to be accurate, and clear credibility signals are how you become that source.

Does content structure really affect whether I get cited?

Yes, structure is one of the most controllable levers you have. Models cite what is easy to extract cleanly. A claim written as a self-contained two-to-three sentence answer directly under a question-shaped heading is far easier for a model to quote than the same claim buried in the middle of a long narrative paragraph. In my experience, improving extractability by restructuring existing content into answer-first blocks tends to move prompts from Absent or Used-uncredited toward Cited faster than almost any other single change. It costs nothing in new content and works with material you already have. Structure will not overcome missing presence or weak credibility, but where those are in place, it is often the deciding factor.