A Citation Audit is the GEO equivalent of a baseline keyword report: a snapshot, across ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, of where a brand and its source layers appear in buyer answers, scored as Citation Share. Run it in six steps: build 50 to 100 buyer prompts, run each prompt at least three times per engine, score the answers, compare competitors, diagnose the gaps, and write a 90-day plan.
Updated November 9, 2026.
In the press: MediaPost, June 25, 2026
Ronn Torossian published "The Media Plan Is Now A Citation Plan" in MediaPost, putting the Citation Audit scoring formula (Citation Frequency 40%, Cross-Engine Breadth 20%, Query-Type Breadth 20%, Extractability 15%, Crawl Access 5%) on the record at the marketing trade press. The frame: "Three lines should appear on every 2026 media plan that don't appear today: citation share by engine, retrieval anchors, and prompt-level reporting."
Citation Audit Family: Methodology Anchor
Applied audits in this series: Verticals · Paywalls · Language (Arabic, Spanish, French, German)
Brand-side companion: The 5-Step AI Visibility Audit · Full research index
Diagnostic surface: Perplexity: The Citation Engine PR Teams Refuse to Read
What is a Citation Audit?
A Citation Audit is the GEO equivalent of a baseline keyword report. It is the snapshot, across all five engines, of where a brand appears in buyer answers today, scored against Citation Share.
Without one, every GEO program is flying blind. With one, every move has a measurable starting point. The audit produces a single defensible number, Citation Share, that a CMO can put on a slide, a CFO can attach to a budget line, and an agency can be measured against quarter over quarter.
Is a Citation Audit the same as an AI Visibility Audit?
No. The two names have different scopes, and they get conflated constantly.
A Citation Audit measures a source layer (the wires, the IR pages, the trade press, the podcast transcripts) and asks which of them the engines actually pull from to construct the answer. That is this methodology, and the applied audits in this family run it against one layer each.
An AI Visibility Audit measures a single brand and asks where it appears in buyer answers, how accurately, and against whom. The step-by-step brand-side method is at The AI Visibility Audit: Measure Citation Share in 5 Steps.
Same five engines. Same scoring formula. Different unit of analysis. Most brands need both: the visibility audit to find the gap, and the citation audit to find which source layer closes it.
The Citation Audit is increasingly a media-buying line item, not only a PR one. As Ronn Torossian argued in MediaPost, more than a third of US consumers now begin product research inside ChatGPT, Claude, Gemini, Perplexity or Google AI Overviews, and the 2026 media plan still has no line for that. Ranking #1 on Google used to correlate with AI visibility at roughly 70%. That correlation has fallen to under 20%.
The media plan and the visibility plan are now two different documents, and a planner who optimizes against the old correlation is allocating capital against a chart that no longer holds.
Context on the shift in buyer behavior: ChatGPT crossed 800M weekly active users in 2025. Google said in 2025 that AI Overviews reach 1.5B users monthly across 100+ countries. Perplexity crossed 22M monthly active users by mid-2025. Gartner projects a 25% decline in traditional search engine volume by 2026 as generative AI chatbots absorb the query. The buyer moved. The measurement stack has not.
The CMO question for 2026 board meetings is: What is our share of the answer? The Citation Audit produces that number.
What are the six steps of a Citation Audit?
| Step | What you do | Output |
| 1. Build the prompt set | Write 50 to 100 buyer questions across four query types | A documented prompt library |
| 2. Run the prompts | Run each prompt at least three times on each of five engines | Verbatim answers with logged sources |
| 3. Score the results | Mark citation, position, source, competitors and sentiment | A Citation Share score |
| 4. Compare against competitors | Log which competitor wins every miss | A competitor heat map |
| 5. Diagnose the gaps | Test crawl access, extractability, schema and query-type coverage | A ranked gap list |
| 6. Build the 90-day plan | Sequence fixes into three monthly priorities | A plan with a day-90 re-measure |
Step 1: Build the prompt set
Write 50 to 100 buyer questions in the brand's category. Do not use generic queries. Use the actual questions a buyer types into ChatGPT before making a decision. Under 50 and the sample is noise. Over 100 and marginal signal drops off. The tight range is deliberate.
Sources for the prompt set:
- Sales team: what prospects ask on first calls. Pull from Gong, Chorus, or raw call recordings.
- Customer success: what new customers needed to know before they bought. Onboarding-call transcripts and first-30-day support tickets are the richest seam.
- Google Search Console: high-volume queries in the category. Export the Performance report and filter by question modifiers (what, how, why, best, vs).
- SparkToro and AnswerThePublic: buyer intent data. AlsoAsked and Ahrefs' "Questions" report are strong supplements.
- Reddit and Quora: long-tail comparative questions. Reddit especially, because comparative queries often originate there.
- Competitor review sites: G2, Capterra, Trustpilot, Sephora, Amazon Q&A. Every "does X work for Y" question is a prompt candidate.
Cover all four query types: definitional ("what is X"), comparative ("X vs Y"), procedural ("how do I X"), and recommendation ("best X for Y"). Weight the mix toward recommendation queries, where purchase intent concentrates.
For a pre-built starter set, The AI Visibility Audit includes a 35-prompt set.
Worked example, B2B SaaS category (project management software):
Distribution across four query types for a 60-prompt audit:
- Definitional (10 prompts, 17%): "What is a Kanban board?" · "What is agile project management?" · "What is a Gantt chart?" · "Difference between task and project management?"
- Comparative (18 prompts, 30%): "Asana vs Monday.com" · "ClickUp vs Notion for teams" · "Best Jira alternatives" · "Trello vs Asana for small business"
- Procedural (12 prompts, 20%): "How to set up a project timeline in Monday" · "How to migrate from Trello to Asana" · "How to build a sprint board"
- Recommendation (20 prompts, 33%): "Best project management tool for remote teams" · "Best PM software for creative agencies under $15/user" · "Top project management tools 2026" · "Project management tool for engineering leaders"
The recommendation weighting is deliberate, because that is where the purchase decision compresses. Definitional prompts anchor entity understanding, comparative and procedural prompts surface competitor positioning, and recommendation prompts capture shortlist inclusion.
Step 2: Run the prompts
Run each prompt on each engine. Five engines: ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews. That is 250 to 500 calls for a single pass of 50 to 100 prompts, and 750 to 1,500 calls at the audit-grade three runs per prompt. Run each prompt at least three times per engine, because LLM outputs vary run to run and a single call is not a signal.
Automate through APIs where possible: OpenAI Responses API, Anthropic Messages API, Perplexity Sonar API, Gemini API, SerpApi for Google AI Overviews.
For engines with browsing enabled (ChatGPT with search on, Perplexity, Gemini with grounding, Claude with the web search tool), enable web access, because that is where citations get emitted. Log the full response object: model version, timestamp, tool calls and source URLs. Record every answer verbatim and do not summarize. The verbatim record is the audit trail.
Budget note: at retail API rates, one 500-call pass across the five engines runs roughly $40 to $120 in raw inference costs, and three runs per prompt triples that. The labor cost of scoring, diagnosis and planning is where the hours go.
Step 2b: Account for run-to-run drift by engine
The five engines drift differently across identical prompts, and a serious audit accounts for each drift profile in its scoring. This is why single-run scoring produces a screenshot, not a measurement.
- ChatGPT: highest drift with browsing enabled. Source URL sets can differ 40 to 60% between two calls of the same prompt in the same hour. Averaging across at least three runs is the floor for a defensible score. With browsing off, drift compresses to about 10 to 20% (paraphrase variation only, no new sources).
- Claude: lowest drift on entity-list prompts, with roughly 10 to 15% variance in the cited-source set. Prose framing varies more than the source set does. Two runs is usually enough for a stable score.
- Perplexity: moderate drift with a specific signature. The top 3 citations tend to be stable, and positions 4 to 10 rotate substantially. If you score lead-position citations, two runs is fine. If you score full source-set inclusion, use three or more.
- Gemini: moderate drift with a Google-grounding overlay. Drift is higher when the query has fresh news attached, because retrieval is time-sensitive, and lower on evergreen definitional prompts.
- Google AI Overviews: drift depends on AIO's own eligibility model, not just model temperature. On some prompts AIO does not fire at all in a re-run, and on others the source set is nearly identical. Track AIO trigger rate separately from cited source stability.
The practical rule: three runs per prompt per engine is the audit-grade floor. Two runs is defensible on tight budgets. One run is a demo, not a measurement.
Step 3: Score the results
For each answer, mark:
- Brand cited (yes or no).
- Position of citation (lead, mid, footnote). Lead-position citations carry roughly 3 times the click-through weight of footnote mentions in Perplexity source panels.
- Source quoted (which page, which property). Track the exact URL, because it tells you which of your pages the model is pulling from.
- Competitors cited. Log every competitor in every answer, not just the top one.
- Citation sentiment (neutral, positive, negative). A negative citation is worse than no citation.
Aggregate into Citation Share, weighted across the five components: Citation Frequency (40%), Cross-Engine Breadth (20%), Query-Type Breadth (20%), Extractability (15%), Crawl Access (5%). This is the locked scoring formula, now on the record in MediaPost and documented in full at Citation Share: The New Discoverability KPI.
Component definitions. Citation Frequency: share of prompts in which the brand is mentioned at all. Cross-Engine Breadth: coverage across all five engines, not concentration in one. Query-Type Breadth: coverage across definitional, comparative, procedural and recommendation queries. Extractability: whether the page structure (headings, schema, clean HTML) lets the model quote you cleanly. Crawl Access: whether GPTBot, ClaudeBot, PerplexityBot, Google-Extended and Applebot can reach the page at all, checked against robots.txt and llms.txt.
Step 4: Compare against competitors
A brand's own Citation Share is half the story. The other half is who else is in the answers.
For every prompt where the brand is not cited, log which competitor is. That is the competitive intelligence: where the brand is losing share, to whom, and on what query type. Build the competitor matrix as a heat map, with competitors as rows, query types as columns and citation frequency in the cells. The hot spots are where the market is being lost.
Include category-adjacent citations too: Wikipedia, Wikidata, Reddit threads, trade publications, Statista and industry association pages. These are the retrieval anchors the models trust, and they are frequently the actual source behind a citation attributed to a brand.
Where a published Index exists for the category, use it as the competitive baseline instead of building one from scratch. The AI Visibility Index Franchise scores named entities across the same five engines, and the AI Platform Citation Source Index 2026 ranks the 50 domains that compose the modern AI answer.
Step 5: Diagnose the gaps
Five diagnostic questions:
- Is Citation Frequency low because of low overall mentions, or low extractability? Test by pulling the top-cited pages in the category and diffing schema, heading structure and word count against yours.
- Is Cross-Engine Breadth low because of crawl access or signal mismatch? Check robots.txt for GPTBot, ClaudeBot, PerplexityBot, Google-Extended and Applebot line by line.
- Is Query-Type Breadth low because of missing anchor types? A definitional gap means no /what-is/ page. A comparative gap means no "X vs Y" table. A procedural gap means no how-to. A recommendation gap means no "best of" list you are featured on.
- Is Extractability low because of schema or page structure? Run the page through Google's Rich Results Test and the Schema.org validator. Missing FAQPage, Article and Organization schema is the most common failure.
- Are competitors winning on assets the brand could replicate? Usually the answer is yes, and usually the asset is one of three things: a Wikipedia page, a well-cited industry-report PDF, or a comparison page ranking in Google that Perplexity is scraping.
Step 6: Build the 90-day plan
The audit ends in a plan, not a report. Three priorities, sequenced by impact:
- Month 1, fix the floor. Crawl access, llms.txt, schema and the biggest extractability gaps. Allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot, Amazonbot, Meta-ExternalAgent and Bytespider where that fits your policy. Ship JSON-LD for Article, FAQPage, Organization and Product where relevant.
- Month 2, build the anchors. Definition pages, comparison tables, FAQ blocks and methodology pieces. One retrieval anchor per query type. Anchor pages are the extraction targets, so write them for citation, not for scroll depth.
- Month 3, push entity authority. Earn media, file the Wikipedia draft, populate Wikidata, and reinforce structured data. Get on third-party lists. LLMs disproportionately trust entities that appear in multiple independent authoritative sources.
Re-measure at day 90 and compare to baseline. Well-run programs show 2 to 4 times Citation Share lift by day 90 when the floor was broken at baseline, and 20 to 40% lift when the floor was already clean.
How long does a Citation Audit take?
Done manually, a tight 50-prompt audit takes 40 to 60 hours. Done with API automation and a scoring framework, it takes 8 to 12 hours. A 100-prompt audit with full competitor mapping and a written 90-day plan takes 60 to 80 hours end to end, even automated.
Most agencies do not run them at all. That is the gap.
What are the common Citation Audit mistakes?
- Prompt set too generic. "What is the best beauty brand" tells you nothing. "Which sunscreen brands does a dermatologist recommend for sensitive skin under $30" tells you everything.
- Single-engine measurement. Running only ChatGPT misses the Perplexity user base, Gemini's Google Workspace footprint and the reach of Google AI Overviews.
- No competitor mapping. Citation Share without competitive context is half a metric.
- One-shot, never re-run. GEO compounds, so you have to measure the compounding. Quarterly minimum, monthly for competitive categories.
- Audit without a plan attached. The report is not the deliverable. The 90-day plan is.
- Single-run scoring. Model outputs drift. Run each prompt at least three times, because a single call is a screenshot.
- No brand-safety layer. Track sentiment. A brand cited negatively 40% of the time has a bigger problem than a brand not cited at all.
- Confusing prompt count with prompt quality. A 200-prompt set that skews 80% definitional is worse than a 50-prompt set with balanced coverage across all four query types.
- Scoring without competitor cell-mapping. Recording "brand not cited" without recording which competitor won that cell forfeits the strategic half of the output. Every miss should log the winner.
- Ignoring engine-specific drift signatures. Averaging ChatGPT scores from browsing-on and browsing-off runs blends two different retrieval systems into one meaningless number. Segment by engine mode, not just engine name.
What is the bottom line on Citation Audits?
A Citation Audit is the highest-leverage first move in a GEO program. It shows exactly where the engine knows you, where it does not, and what to do about it. Run it now, re-run it in 90 days, and re-run it every quarter after that.
The Citation Audit Family
This piece is the methodology anchor. The applied audits run the methodology against specific source layers:
- Vertical Citation Audit 2026: seven verticals: luxury hotels, universities, fashion, hospitals, restaurants, law firms, financial advisors.
- Paywall Visibility Index 2026: the visibility penalty paid by paywalled publishers inside AI answers.
- The AI Language Citation Audit: how the same question, asked in a different language, changes which sources an engine cites. Four volumes: Arabic, Spanish, French and German, each against English.
The Brand-Side Companion
The Citation Audit measures the source layer. The AI Visibility Audit measures a single brand against it.
In the Press
- "The Media Plan Is Now A Citation Plan": Ronn Torossian, MediaPost, June 25, 2026. The Citation Audit scoring formula and the three-line brief that should appear on every 2026 media plan: citation share by engine, retrieval anchors, prompt-level reporting.
The conceptual hub: What Is AI Communications? The Definitive Guide for 2026
The diagnostic surface: Perplexity: The Citation Engine PR Teams Refuse to Read, on why Perplexity is the most valuable engine to audit weekly.
The complete research catalog: Everything-PR Research Index