Build a Cross-LLM Citation Audit in Six Steps: Tracking Brand Visibility Across ChatGPT, Gemini, Perplexity, and Claude

What a Cross-LLM Citation Audit Delivers

A cross-LLM citation audit is a structured measurement of how a brand appears across ChatGPT, Google Gemini, Perplexity, plus Claude when users ask the same buyer questions. The DSF Cross-LLM Audit Pipeline runs paired prompts across all four engines, captures cited URLs plus brand mentions, scores cross-engine divergence, maps gaps to architectural causes, then prioritizes the closeable fixes. The output is a quarterly scorecard that converts AI search visibility from a vanity metric into a measurable, comparable surface.

The discipline matters now because a 15,000-query Ahrefs analysis measured only 12 percent overlap in cited URLs across the major engines. A brand that ranks first inside one engine can be entirely absent from the other three, which makes any single-engine visibility report an 88 percent blind spot by construction.

What a Cross-LLM Citation Audit Actually Measures

A defensible audit measures three dimensions simultaneously, with each dimension producing a distinct decision input. The first dimension is the citation surface: the population of buyer queries the audit is designed to cover, the engines under measurement, plus the cadence at which the measurement repeats.

The second dimension is the citation output per engine: the URLs cited, the brand mentions inside the synthesized response, the ordinal position of the first brand mention, plus the response's framing of the brand. The third dimension is cross-engine divergence: the quantitative comparison of those outputs across the four engines, expressed as overlap percentage, citation share, plus position-weighted share.

A citation audit is not a screenshot tour. A team capturing a single ChatGPT screenshot for a board deck has performed a demonstration, not a measurement. The measurement requires the same query asked across all four engines in a structured run, the responses logged in a comparable schema, then the cross-engine comparison computed against a defined baseline. Anything short of that produces a story rather than a number.

The discipline borrows from established measurement traditions. Search-engine rank tracking compares positions across a fixed keyword set on a defined cadence. Brand-tracking surveys compare aided plus unaided recall across a stable panel quarter over quarter. A cross-LLM citation audit applies the same logic to generative answer engines, with the citation array replacing the rank position as the primary unit of measurement.

Why Single-Engine Audits Fail: The 12 Percent Overlap Problem

The single-engine audit is the dominant industry practice. A team selects ChatGPT because it has the most users, runs fifty buyer queries, screenshots the citations, calls the result an AI visibility report, then extrapolates that the brand's position on the other three engines is similar. The extrapolation is invalid.

The December 2025 Ahrefs analysis of 15,000 buyer-intent queries across ChatGPT, Gemini, Perplexity, plus Claude found that just 12 percent of cited URLs appeared on more than one engine for the same query. The remaining 88 percent of citations are platform-specific. A brand winning ChatGPT for "best CRM for boutique law firms" can be entirely absent from Gemini, Perplexity, plus Claude on the identical query.

The reason for the divergence is mechanical, not random. Each engine pulls from a different source mix. Citation-pattern analysis from the visibility platform Profound documented that ChatGPT sources are 47.9 percent Wikipedia and 11.3 percent Reddit, while Perplexity sources are 46.7 percent Reddit and 13.9 percent YouTube. Google AI Overviews cite 21.0 percent Reddit and 18.8 percent YouTube. The retrieval indexes differ. The source mix divergence compounds at the engine level because each engine's retrieval index was trained on a different corpus, runs a different reranker, plus applies different freshness logic before generating the synthesized answer.

The architectural cause traces deeper than retrieval mix. A 2025 Google DeepMind and Johns Hopkins University paper documented theoretical limits on how many distinct passages a single embedding model can rank as "most similar" to a given query, with the cap rising slowly as embedding dimensionality grows. Different engines use different embedding models, so the set of passages each engine considers retrieval-eligible diverges before any reranker runs. The 12 percent overlap is the downstream output of upstream embedding divergence.

The practical implication is direct. A brand that audits ChatGPT only and assumes the result generalizes is making a strategic decision on a 12 percent sample. The decision is no more reliable than running a national survey by polling only one ZIP code. The fix is structural: every audit measures all four engines or it does not count as an audit.

The DSF Cross-LLM Audit Pipeline at a Glance

The DSF Cross-LLM Audit Pipeline organizes a complete audit into six sequential stages, with each stage producing a defined artifact that feeds the next. The pipeline is engine-agnostic, prompt-format-agnostic, plus tooling-agnostic. It works whether the audit is built on a Python script calling four APIs, a paid platform subscription, or a hybrid configuration that combines API access for some engines with browser automation for the engines without an open API.

The six stages are: define the citation surface, run paired prompts across the four engines, capture cited URLs plus brand mentions in a structured schema, calculate divergence scores, map identified gaps to architectural causes, then prioritize the closeable fixes against a defined effort-impact matrix. The output of stage six is a quarterly scorecard that ranks fixes by expected lift per unit of engineering effort, with each line item traceable back to the specific gap that surfaced in stage four.

The discipline of the pipeline matters more than the tooling. Teams that follow a structured pipeline produce comparable scorecards quarter over quarter, with trend lines that show whether visibility is improving or eroding. Teams that improvise produce one-off reports that cannot be compared to the prior quarter, which means leadership cannot tell whether AEO investment is paying back.

Step 1: Define the Citation Surface

The citation surface is the population the audit measures. It contains three components. First, the buyer-query set: 50 to 200 questions a real customer would ask an AI engine during the purchase journey, weighted by funnel stage with at least 20 percent at awareness, 50 percent at consideration, plus 30 percent at decision. Second, the engine set: ChatGPT, Gemini, Perplexity, plus Claude as the four-engine baseline, with optional extension to Microsoft Copilot, Apple Intelligence, or Grok if the buyer persona uses them. Third, the cadence: monthly minimum, with quarterly deep-dives that expand the query set plus add long-tail variations.

Query selection is the highest-leverage step in the entire pipeline. The mistake to avoid is asking internal jargon questions. A team building a CRM platform might write "What is a unified customer data fabric for revenue operations" when the actual buyer asks "What CRM should a 50-person SaaS company use." The Ahrefs analysis above measured 12 percent overlap on buyer-intent queries, but jargon queries produce much higher overlap because they have fewer cited candidates.

The audit measures the wrong thing when the query set is built from internal language. The corrective: source queries from sales call transcripts, support ticket subject lines, plus the actual search queries that drove customers to the brand's site in the prior quarter.

The query set needs query reformulation awareness. AI engines do not necessarily search for the literal query text. They rewrite the query before retrieval, and the rewrite can change which sources surface. Including a small set of paraphrased variations for the top 20 queries reveals whether the brand's visibility is robust across rewrites or fragile under one specific phrasing.

The output of step one is a versioned query manifest. Versioning matters because the query set evolves: new buyer questions emerge, old questions become irrelevant, plus seasonal questions cycle in and out. A version number on the manifest lets the quarter-over-quarter comparison stay honest about what changed.

Step 2: Run Paired Prompts Across the Four Engines

Step two is the data-collection stage. The discipline is prompt equivalence: every query in the manifest is asked through an equivalent prompt format on every engine, with all four runs completed on the same calendar day. A run that captures ChatGPT on Tuesday plus Gemini on Friday is comparing two different days of citation reality, which is a confound.

The implementation path differs by engine. Anthropic publishes a Claude web search tool that returns citations as a structured field alongside the response, with each citation including the source URL plus the cited text excerpt. Perplexity's Sonar API includes a citations array in every response that lists the URLs the model used to ground the answer. Google's Gemini API grounding endpoint returns a groundingMetadata field containing groundingChunks (the source URLs) plus groundingSupports (the mapping from response text to source). The three APIs use different field names but expose equivalent data, which makes structured comparison straightforward.

ChatGPT is the harder engine to measure programmatically because the consumer ChatGPT search product does not currently expose a public API that returns citations in the same structured form. The available paths are browser automation through a tool like Playwright, the ChatGPT search action through approved enterprise tooling, or a manual capture process that screenshots the response then OCR-parses the citation block. Most production audits use a hybrid: API access for Anthropic, Perplexity, plus Gemini, browser automation for ChatGPT.

Prompt format equivalence does not mean identical strings. Each engine has a preferred input style, with ChatGPT favoring conversational phrasing, Gemini favoring concise direct questions, Perplexity favoring search-style keyword phrases, plus Claude favoring well-formed sentences. The audit can use engine-native phrasings provided every engine receives semantically equivalent queries. The equivalence test: a human reviewer reading both prompts must conclude they are asking the same question.

Step 3: Capture Cited URLs and Brand Mentions

Step three normalizes the heterogeneous engine outputs into a single logging schema that downstream analysis can compute against. The schema needs ten fields per query-engine pair, with each field producing a specific decision input later in the pipeline. Captured raw responses are stored verbatim in object storage for later reanalysis, with the normalized schema in a queryable database for divergence calculations.

The capture step is where most audits accumulate technical debt. A team that begins logging only the cited URLs will not be able to compute position-weighted share later. A team that omits the response_length field will not be able to compare verbose Gemini answers to concise Claude answers on equal footing. The schema is cheap to define correctly at the start plus expensive to backfill later, so the practical rule is to log everything from the first run.

Brand mention detection is a separate step from URL citation detection. The cited URLs are returned by each engine's API in a structured field, but brand mentions inside the response prose require a separate string-match plus disambiguation pass. A response that mentions "HubSpot" might be referencing the company or one of HubSpot's product lines, and the audit needs to disambiguate to produce a defensible mention count. The standard practice is a named-entity recognition pass against a brand alias list maintained per company.

Step 4: Calculate Divergence Scores

Step four converts the captured schema into three divergence metrics that together describe the brand's cross-engine visibility profile. The first metric is pairwise overlap: for each pair of engines, the fraction of cited URLs that appear in both engines' citation lists for the same query. The second metric is citation share: the percentage of total citations across the audit that point to the brand's domain. The third metric is position-weighted share, which weights the share by the ordinal position of the first brand mention so a brand mentioned in the first sentence counts more than a brand mentioned in the last paragraph.

The three metrics measure different things, and reporting all three matters. Pairwise overlap reveals platform-specific gaps. Citation share reveals the absolute size of the brand's footprint in each engine. Position-weighted share reveals whether the brand is the lead answer or the also-ran. A brand can have high citation share on one engine but low position-weighted share because it appears late in the response, which is a different fix than a brand with zero share on the engine entirely.

The citation divergence score is the headline single number derived from the three metrics. It is computed as 1 minus the average pairwise overlap across the four engines, which produces a value between 0 (perfect overlap, all engines cite the same sources) and 1 (perfect divergence, no shared citations across engines). The Ahrefs 12 percent overlap finding implies an aggregate citation divergence score of approximately 0.88 across the market, with category-specific scores varying around that anchor.

A second derived metric is brand mention density, calculated as the number of brand mentions per thousand response tokens, position-weighted. This metric isolates how prominently the brand features when it appears, separate from how often it appears. The two metrics together produce the brand's cross-engine visibility profile.

Step 5: Map Gaps to Causes

Step five is the diagnostic stage. Each gap surfaced in step four is labeled with one of four architectural causes, with each cause carrying its own remediation playbook. Skipping this stage and jumping directly to fixes is the most common mistake in cross-LLM auditing because the symptoms look similar across causes but the fixes diverge sharply.

The first cause is the crawl coverage gap. The brand's pages exist but the AI engine's crawler never reached them, or reached them and decided not to index them. Diagnostic check: fetch the brand's robots.txt, server logs, plus crawl coverage reports to confirm whether OAI-SearchBot, Google-Extended, PerplexityBot, plus ClaudeBot have visited the relevant pages. If a crawler never visited, the brand cannot be cited regardless of content quality.

The second cause is the embedding eligibility gap. The page exists, the crawler indexed it, but the page's passages do not embed close enough to the query embedding for the reranker to surface them. The Google DeepMind and Johns Hopkins paper cited above establishes that embedding models have a structural ceiling on how many distinct passages they can rank as most-similar to a given query, so even technically retrievable content can be embedding-ineligible relative to competitors with more semantically aligned passages.

The third cause is the schema markup gap. The page is crawled plus retrievable, but the engine cannot parse the structured information cleanly because the page lacks the appropriate Schema.org Article markup, FAQPage markup for question-style content, or Organization markup linking the brand identity to the entity graph. Diagnostic check: validate the page's JSON-LD against the Schema.org validator plus inspect what entities the engine extracts when the page is the citation source.

The fourth cause is the source mix divergence gap. The page is crawled, retrievable, plus well-marked, but the engine's source mix for the query category does not surface the page's domain type. If Perplexity sources 46.7 percent from Reddit plus the brand has zero Reddit footprint in the relevant category, the page cannot win that engine no matter how excellent the page is. The fix is a presence-on-source strategy rather than a content fix on the brand's own pages, which is a category of work many AEO programs do not realize they need.

Step 6: Prioritize Closeable Gaps

Step six is the prioritization stage. The audit will surface more gaps than the team can fix in a quarter, and ranking them by expected lift per unit of engineering effort separates the closeable from the theoretical. The pipeline uses an effort-impact 2x2 matrix with four quadrants: quick wins (low effort, high impact) ship first, strategic bets (high effort, high impact) get budgeted into the roadmap, cleanups (low effort, low impact) batch into housekeeping sprints, plus avoid (high effort, low impact) get documented but not funded.

Expected lift estimation uses the citation probability framework. Each gap, once closed, increases the brand's citation probability on the affected query set by a measurable factor. A schema markup fix on the top 20 buyer queries with a current 0.15 citation probability that moves to 0.40 represents a much larger expected lift than a freshness fix on the bottom 50 queries with 0.05 probability that moves to 0.08. The math is straightforward once the divergence scores from step four plus the cause labels from step five are joined.

The output of step six is a quarterly scorecard. The scorecard ranks the top 10 to 20 closeable gaps by expected lift, attaches an engineering estimate to each, plus traces the lineage back through the cause label, the divergence score, plus the original query that surfaced the gap. The scorecard is the input to the next quarter's AEO roadmap, with the prior quarter's scorecard items either marked as shipped (with measured lift confirmed in the next audit run) or carried forward with updated estimates.

Tools: DIY Scripts vs Paid Platforms vs Hybrid

A cross-LLM audit can be built three ways, with each path producing comparable results at different total-cost-of-ownership profiles. The DIY path uses a small Python codebase that calls the four engines' APIs (with browser automation filling the ChatGPT gap), stores results in PostgreSQL or DuckDB, then runs the divergence calculations in pandas or polars.

The paid platform path subscribes to one of the cross-LLM visibility platforms that have emerged over the past 24 months, with the platform handling capture, storage, plus scoring. The hybrid path uses paid platforms for the engines the platform supports well plus DIY scripts for the gaps, which is the configuration most production audits end up on.

The DIY path costs API tokens plus engineering time. A 200-query monthly audit across the four engines runs approximately 800 API calls per month, with total API spend under 100 dollars at current pricing. Engineering time to build the pipeline is two to four engineer-weeks of focused work, with maintenance running approximately one engineer-day per month for schema updates and cassette refreshes. The DIY path produces full control over the schema, the prompts, plus the export format.

The paid platform path costs platform fees plus reduced flexibility. Visibility platforms in the category typically price between 500 and 5,000 dollars per month for the cross-LLM tier, with enterprise pricing scaling by query volume plus engine breadth. The platform path is faster to value because there is no engineering build, plus the platform vendor handles engine API changes when they happen. The trade-off is schema rigidity, which makes integrating audit data into custom dashboards harder.

Common Audit Failures and How to Avoid Them

The failure modes are predictable, with each one tracing back to a specific stage of the pipeline that was either skipped or compressed under deadline pressure. The first failure is single-engine extrapolation, covered in section three. The second failure is query selection bias, where the audit measures internal jargon rather than actual buyer language. The third failure is stale prompt schedules, where the same query manifest runs for four consecutive quarters without refreshing for new buyer questions that emerged in the interval.

The fourth failure is counting raw mentions without position weighting. A brand mentioned twelve times across the audit but always as the third or fourth alternative looks strong in a raw count yet weak in actual visibility. Position weighting separates the two. The fifth failure is treating symptoms rather than causes, which is the failure step five exists to prevent. A team that fixes schema markup on every gap will overinvest in schema while ignoring the embedding eligibility, crawl coverage, plus source mix divergence causes that produced most of the gaps.

The sixth failure is producing a scorecard without prioritization. A list of 47 identified gaps with no effort-impact ranking gives the engineering team no way to start, which usually translates to the team picking the easiest gaps regardless of expected lift. The prioritization matrix from step six prevents the failure by forcing every gap onto the matrix before any fix work begins.

Audit Cadence: Weekly, Monthly, or Quarterly

The recommended cadence is monthly captures with quarterly deep-dive analysis. Monthly captures keep the trend line visible plus catch sudden divergence shifts within the same quarter they happen. Quarterly deep-dives expand the query set, refresh the manifest for new buyer questions, plus produce the scorecard handoff to engineering. Weekly captures are over-investment for most teams because the AI engines do not retrain weekly, and the noise-to-signal ratio at weekly granularity is high.

The cadence argument is strengthened by adoption trend data. Stanford HAI's AI Index Report documents accelerating adoption of generative AI for information search, with usage rates moving fast enough that quarterly-only measurement leaves teams looking at a stale visibility surface for most of each quarter. The Reuters Institute Digital News Report 2025 similarly documents AI chatbot usage establishing meaningful share of information-seeking behavior, with the share growing fast enough that even monthly capture cadence is the floor, not the ceiling.

Event-driven captures supplement the calendar cadence. A major engine update (Google's AI Overviews expansion, an OpenAI search feature change, a Perplexity model upgrade) warrants an off-cycle capture run within 48 hours of the change to measure whether the brand's position moved. The same applies to brand-side events: a major content launch, a rebrand, or a category expansion all warrant an immediate audit run rather than waiting for the next calendar capture.

FAQ — Cross-LLM Citation Audits

Next Steps — Cross-LLM Citation Audits

For teams that need cross-LLM citation auditing delivered as a managed service rather than built in-house, the Disruptive Strategy Consulting engagement covers the full pipeline build plus the first four quarterly scorecards.