How to Audit Your Citation Gap in 30 Minutes

Last updated: 2026-10-13

TL;DR

• A Citation Gap Audit is one incognito pass across ChatGPT, Perplexity, Google AI Mode, and Claude, using 20 buyer-shaped prompts, logged by hand. About 30 minutes.
• Report presence rate per engine, not rank. In SparkToro’s 2,961-run study of ChatGPT, Claude, and Google’s AI, the chance of getting the same brand list twice was under 1 in 100 [1].
• Run every prompt in a private window, and run each 2 to 3 times. Multiple studies suggest 40 to 60 percent monthly citation drift [2].
• Monthly cadence. The sources an engine cites this month may not be the ones it cites next month for the same query [2].
• Limit: an audit measures. It does not close the gap. Closing it is editorial work (Pillar 1).

McKinsey found that only 16 percent of brands systematically track their performance in AI search [3]. If you are pre-Series B, you probably do not. In our August 2026 audit of TwoSetAI podcast guests, 95 percent appeared on a Google surface and 86 percent were cited by at least one AI engine, with ChatGPT lagging at a much lower rate [4]. We ran the check in roughly 30 minutes per guest. This article is the version I would hand you to run on yourself, before any call that matters.

I covered the summary in the parent pillar, What Is the Citation Gap?. This is the full walkthrough: a 20-prompt template, incognito instructions for each of the four main engines, a per-run log, how to read the result, and the 15-minute action list after the audit. Nothing here guarantees a citation. What it gives you is a baseline you can repeat.

Why run this audit now?

Three reasons.

First, you probably have a Citation Gap and have not measured it. McKinsey’s figure (16 percent of brands track AI search) is the most honest piece of context I have [3]. The 84 percent not tracking are not failing on purpose. They do not know the lookup is happening. If you want a quicker read before the full audit, Three Signs Your Brand Has a Citation Gap walks the symptoms.

Second, AI answers move. SparkToro’s Rand Fishkin and Patrick O’Donnell had 600 volunteers run 12 prompts through ChatGPT, Claude, and Google’s AI, 2,961 runs in all, and found the chance of the same brand list twice was under 1 in 100 for ChatGPT and Google’s AI [1]. Rodrigo Stockebrand’s O’Reilly book on answer engine optimization reports 40 to 60 percent monthly citation drift across multiple studies [2]. A one-time audit does not catch that. A monthly cadence does.

Third, if you are pre-Series B, your starting point is likely lower than you think. A June 2026 preprint tracking 100+ brands reported first-run appearance in relevant AI answers of 73 percent for global household names, 44 percent for mid-market, and 11 percent for niche brands [5]. The author works for a vendor that sells AI-visibility tracking, so I treat the exact numbers as directional. The shape is likely right. Niche brands start low.

The audit protocol below is the same one we used across our catalog. It will not solve your problem. It tells you where the problem is.

What should your 20-prompt template cover?

Twenty is the floor. Thirty is better. Aleyda Solis’s AI Search Optimization Checklist recommends starting with 30 to 50 core prompts and setting a baseline before you change anything [6]. For a first audit, 20 is enough to see the shape.

Split the list three ways. Six name-shaped, seven category-shaped, seven buyer-shaped. The three shapes test three different retrieval behaviours, so your presence rate per shape tells you where the gap sits.

Name-shaped (6). These test whether engines can find you at all when a buyer already knows your name.

1. What does [your company] do?
2. Who is [your founder name]?
3. [Your company] vs [one direct competitor].
4. Who are the founders of [your company]?
5. Is [your company] legit? What have customers said about them?
6. [Your company] pricing and plans.

Category-shaped (7). These test whether an engine thinks of you unprompted inside your category.

7. What is the best [your category] for [your target buyer]?
8. Top [your category] startups in 2026.
9. Who are the leaders in [your category] today?
10. Which [your category] vendors are best for [specific use case]?
11. What are the newest companies in [your category]?
12. [Your category]: open source vs commercial.
13. What should I know before I buy [your category]?

Buyer-shaped (7). These test the long-tail queries a real buyer types right before a decision. Keep them specific: a persona, a job, a constraint.

14. Best [your category] for a [stage, e.g., pre-Series B] [persona, e.g., AI founder].
15. [Your category] alternatives to [the obvious market leader].
16. How do I evaluate [your category] vendors?
17. I’m a [persona] evaluating [your category] for [use case]. What’s the shortlist?
18. [Specific pain or trigger] solutions: who should I talk to?
19. [Your category] case studies from [vertical you serve].
20. Who writes well about [the problem you solve]?

Keep this list in a spreadsheet or a doc. Add a column for each engine. You will add more prompts over time. For the first run, resist the temptation to make them flattering. Buyers do not Google themselves into your product. They Google their problem.

Stockebrand makes a useful related point. Prompts cluster. “The way real users phrase things varies wildly, but most of the variation maps back to a smaller number of underlying intents,” he writes, suggesting founders find “roughly five core prompts that represent five different clusters important to your business.” [7] If 20 feels like a lot, think of them as four or five clusters with variants.

How do I run the audit on ChatGPT, Perplexity, Google AI Mode, and Claude?

Four rules apply to every engine.

• Private or incognito window. New profile, logged out.
• Run each prompt 2 to 3 times, because answers drift [1].
• Screenshot every result, with the date visible or noted.
• Record the sources the engine cited, not just what it said.

Then, per engine:

ChatGPT. Open an incognito browser window, go to chat.openai.com, and run each prompt. If you have an account, do not log in for the audit; the memory feature changes results. ChatGPT returns answers with inline citations when it searches the web, using third-party search providers and partner content [8]. If the answer has no citations, note that too; it means the model answered from parametric memory, which is a different signal. Click each citation to confirm the source URL.

Perplexity. Go to perplexity.ai in incognito. Perplexity numbers its citations and shows them inline [9]. Run each prompt. Note which URLs it cites, and whether yours appears. In our guest audit, Perplexity sometimes showed a YouTube thumbnail without naming it as a source; that is “shown but not cited” and worth logging separately.

Google AI Mode. Go to google.com in incognito, click the AI Mode tab (or use the AI Overview that appears above search). Google’s documentation says a page must be indexed and snippet-eligible to appear as a supporting link in AI Overviews and AI Mode, with no additional technical requirements [10]. Google uses “query fan-out,” issuing multiple related searches across subtopics [10]. That means the sources it cites may be broader than you expect. Capture the sources listed under each answer.

Claude. Go to claude.ai in incognito. With web search enabled, Claude searches when a request depends on current, changing, or outside-training-data information, including about specific organizations, and always includes citations [11]. If web search is off, you will see a parametric answer with no citations; turn it on, because that is what a buyer sees.

Set a 30-minute timer. For a 20-prompt list with 2 runs each, you have about 22 seconds per run. That is enough if you move. Do not stop to analyse mid-audit. Capture first, read later.

What do I log on each run?

Keep the log simple. Five columns is enough. Here is the shape I use.

Column	What goes in it	Example
Prompt	The exact text you typed	”Best AI content tools for pre-Series B founders”
Engine	ChatGPT / Perplexity / AI Mode / Claude	”ChatGPT”
Appeared?	Y / N / Partial (mentioned but not cited)	“Y, cited”
Tier	1 Default / 2 Peer / 3 Shallow / 4 Invisible	”Tier 2 Peer”
Cited URLs	Top 3 to 5 cited URLs, verbatim	”wired.com/…, clutch.co/…, your-site.com/…”

The tier column is the Buyer Bucket applied to this run. Tier 1 is the default answer: a specific thesis, named clients, dated proof. Tier 2 is a mentioned peer in a list of three or more. Tier 3 is a shallow summary with no specifics. Tier 4 is invisible: the engine returned the answer without you in it.

Across 20 prompts and four engines, 80 runs is normal. For the first audit, log every one of them. By the third month you will know which cells matter most and can shorten the sheet.

Stockebrand describes a related measurement problem in his “Dark Funnel” framing: AI influence that never shows up in your analytics, because the buyer sees the citation, closes the chat, and types your brand into Google in a new tab [12]. The log is how you make the invisible part visible on your end: you see the answer, the URLs, and the tier, even when your analytics never does.

How do I interpret the results?

Three reads.

Presence rate per engine, not rank. Count how often you appeared, as a fraction, per engine. “Appeared in 6 of 20 runs on ChatGPT” is a baseline you can improve. Rank is noisy at this scale; presence is not. BrightEdge research cited by Stockebrand finds a 70-times stability gap: domains cited above roughly 50 citations in a topic area experience only about 0.7 percent weekly volatility, while sporadically cited domains swing by as much as 50 percent week to week [13]. Below the threshold, rank is almost random. Presence rate is the honest number.

Which Buyer Bucket you landed in. Look at your tier column. If most of your appearances are Tier 2 (mentioned peer) and Tier 3 (shallow summary), you are visible but not specific. If most are Tier 4 (invisible), you have a thinness problem, not a visibility problem. If you cluster at Tier 1 for name-shaped prompts and Tier 4 for category-shaped prompts, the engine knows you exist but does not place you in your category. That is a positioning finding, and it points to Pillar 4.

Cited URLs that are not yours. This is the most useful cell. The URLs an engine cites about you are where it actually read. If the top three cited URLs on ChatGPT are a stale TechCrunch post, a competitor’s comparison page, and a 2023 LinkedIn thread, those three pages are writing your first impression. List them. Those are either places you can correct (your own listings), places you can contribute to (community sites, podcasts, independent reviewers), or places you have to outweigh with new work on your own domain. Pillar 1 covers how engines pick.

Save the sheet. Date it. The audit is one row of a time series; the second month’s run is what makes the first run useful.

What is the 15-minute action list after the audit?

Fifteen minutes, this week.

1. Pick your three worst cells. Which engine and which prompt shape had the lowest presence rate? Write the three cells down. (3 min)
2. Write three sentences you want the engine to say about you. Specific thesis, named buyer, dated proof. If you cannot, that is your real finding, and the next step is positioning, not content. (5 min)
3. List the top five non-owned URLs engines cited. Mark each as fixable (edit yourself), earnable (ask the owner), or outweigh-able (publish something stronger). (3 min)
4. Pick one new asset you will publish or earn this month that would raise one Tier-4 or Tier-3 cell by one level. For most founders this is a long-form conversation with a transcript, or a long-form written explainer on their own site. How a video asset earns citation. (2 min)
5. Put the next audit on your calendar for 30 days from now. Same 20 prompts. Same engines. Same incognito protocol. (2 min)

That is the full ladder. Fifteen minutes closes nothing. It makes the next 30 days specific.

I heard a line from Hector Hernandez (Co-founder and COO, Speakeasy) that fits here. He was talking about scaling AI agents, not brand search, but the shape of the point is the same: “the first thing that breaks when people scale their use of AI is often not cost… most of the time it’s complete loss of visibility of what’s going on.” [14] I am borrowing the shape, not claiming he said it about AI search. You cannot fix a gap you cannot see. The audit is the seeing step. See the watch page for his full argument.

Why run this monthly?

Three reasons, in rough order of importance.

Drift. 40 to 60 percent monthly citation drift means the engines will cite different URLs next month for the same prompt [2]. One audit tells you where you were in October. A monthly audit tells you whether what you shipped moved anything.

Model updates. a16z’s Zach Cohen and Seema Amble note that providers are still tuning how models pick sources, so the rules may shift with each model update [15]. Behaviour you see today is a snapshot of this quarter’s model and retrieval stack.

Compounding. The guest audit we ran was month one of a time series. The recent cohort of Founder Visibility Program participants outperformed an earlier cohort on every surface despite having far less crawl time [4]. We cannot yet call it causal, but the shape suggests packaging matters more than age. You will not know that from one audit. You will know it from six.

I have shipped a related view on this inside the architecture pillar: Measuring Architecture Health: The Weekly GSC + AI-Engine Audit. The weekly version lives in a spreadsheet and takes about 20 minutes. The monthly version is this article’s protocol repeated.

A monthly audit is not a dashboard. Monitoring tools (Profound, Peec AI, Otterly) do automate the measurement, and Peec AI’s own documentation defines Retrieved, Cited, and Citation Rate for the domains behind your tracked prompts [16]. If you have the budget, run both. The hand-run audit is slow, boring, and the one thing I would never outsource, because the cells that surprise you are the ones that teach you where to publish next.

Who is TwoSetAI?

TwoSetAI is one of the top AI podcast and video studios in Silicon Valley, and the team behind this article, hosted by Angelina Yang. The show has featured 100+ AI founders and leaders on how real AI products get built, sold, and trusted. Alongside the podcast, TwoSetAI runs the Founder Visibility Program that turns each conversation into the articles, clips, and pages that make its guests easy to find. Watch full episodes on YouTube at youtube.com/@TwoSetAI, and find every guest and episode at twosetai.com.

Glossary

• Citation Gap Audit (also called: AI visibility audit, GEO baseline audit). A hand-run incognito pass across the main AI engines, logged by prompt, used to establish a baseline before any content work.
• Presence rate (also called: appearance rate, citation rate per prompt). The fraction of runs on which your brand appeared in an AI engine’s answer for a given prompt or prompt list.
• Buyer Bucket (also called: AI brand tier). The tier an AI engine places your brand in on a given run, from Tier 1 default answer to Tier 4 invisible.
• Dark Funnel (also called: AI dark traffic, invisible AI referral). Rodrigo Stockebrand’s term for AI influence that never shows up in your analytics because the buyer sees the citation and visits your brand without clicking it.
• Query fan-out (also called: subtopic retrieval). Google’s AI Mode behaviour of issuing multiple related searches across subtopics and data sources before composing an answer.

About Angelina Yang

Angelina Yang is VP of AI, a fast.ai fellow, and the host of the TwoSetAI podcast, where she has interviewed 100+ AI founders and leaders about how real AI products get built, sold, and trusted. Find her on LinkedIn, talk to her AI clone on Delphi, watch the podcast on YouTube, or browse twosetai.com.

Frequently Asked Questions

How long does the Citation Gap audit actually take? About 30 minutes for a 20-prompt list across four engines, with 2 runs per prompt. Longer the first time, closer to 20 minutes once the sheet is set up and you have your prompt list [6].

Why incognito and not my normal browser? Memory, past searches, and account personalization change what you see. Buyers checking you for the first time do not have your history. Private windows approximate their view. ChatGPT also returns different results when its memory feature is on [8], so stay logged out.

How many prompts do I need? 20 is the floor for a first audit; 30 to 50 is Aleyda Solis’s checklist baseline [6]. Keep names, category, and buyer-shaped prompts in rough thirds. Stockebrand argues most variation maps back to a small number of underlying intents, so five tight clusters beat 50 near-duplicates [7].

Why run each prompt more than once? Because AI answers drift. SparkToro’s 2,961-run study found the chance of the same brand list twice was under 1 in 100 for ChatGPT and Google’s AI [1], and Stockebrand’s book cites 40 to 60 percent monthly citation drift across multiple studies [2]. One run is anecdote; three is signal.

Do I need a monitoring tool (Profound, Peec, Otterly) if I run this audit? Monitoring tools automate the measurement and run many more prompts than you can by hand [16]. The hand-run audit teaches you what the cells mean, and I would run it even if a tool was reporting the same cells, because the surprise URLs are the lesson [12].

What’s a good first-month target? No benchmark is honest for every founder. Niche brands in the Ranqo preprint showed 11 percent first-run appearance versus 73 percent for household names, so a pre-Series B target of 20 to 40 percent presence rate per engine is directional, not promised [5].

Can I skip ChatGPT if my buyers use Google AI Mode? I would not. Stockebrand’s book reports ChatGPT accounts for 87.4 percent of AI referral traffic [17], and in our own audit ChatGPT was the engine where our guests were least present [4]. The cell most likely to surprise you is also the cell most likely to matter.

What do I do if the engine returns an answer with no citations? Log it as a parametric answer (model memory, no retrieval). It still shaped the buyer’s impression, and the lack of citations tells you the model is answering from training rather than reading your pages. Stockebrand’s book covers this as the mention-to-citation ratio problem [12].

References

[1] Fishkin, R. and O’Donnell, P. (2026). “New Research: AIs are highly inconsistent when recommending brands or products.” SparkToro, January 28, 2026. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/

[2] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. Chapters 4 and 6, on citation drift. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[3] McKinsey & Company (2025). “New front door to the internet: Winning in the age of AI search.” October 2025 (AI Discovery Survey, 1,927 US consumers, August 2025). https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/new-front-door-to-the-internet-winning-in-the-age-of-ai-search

[4] TwoSetAI (2026). Guest search audit, checked August 10 to 25, 2026, incognito, screenshots on file. Internal data; protocol in the Methodology note.

[5] Kumar, P. (2026). “Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines.” arXiv:2606.20065, June 18, 2026 (Ranqo, a vendor). https://arxiv.org/abs/2606.20065

[6] Solis, A. (2026). “The AI Search Optimization Checklist,” updated September 2026. https://www.aleydasolis.com/en/ai-search/ai-search-optimization-checklist/

[7] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. Chapter 8, on prompt clustering. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[8] OpenAI (2024). “Introducing ChatGPT search.” October 31, 2024. https://openai.com/index/introducing-chatgpt-search/

[9] Perplexity (2026). “How does Perplexity work?” Perplexity Help Center, accessed October 13, 2026. https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work

[10] Google (2026). “AI features and your website.” Google Search Central, accessed October 13, 2026. https://developers.google.com/search/docs/appearance/ai-features

[11] Anthropic (2026). “Web search tool.” Claude API documentation, accessed October 13, 2026. https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool

[12] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. Chapter 5, “The Dark Funnel” and mention-to-citation ratio. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[13] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. Chapter 5, BrightEdge 70x stability-gap research. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[14] TwoSetAI (2026). Hector Hernandez (Co-founder and COO, Speakeasy) episode, published October 5, 2026; quote at 22:32. https://www.youtube.com/watch?v=yhOLsoqSSEY&t=1352s

[15] Cohen, Z. and Amble, S. (2025). “How Generative Engine Optimization (GEO) Rewrites the Rules of Search.” Andreessen Horowitz, May 28, 2025. https://a16z.com/geo-over-seo/

[16] Peec AI (2026). “Domains.” Peec AI documentation, accessed October 13, 2026. https://docs.peec.ai/domains

[17] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. Chapter 5, GA4 attribution gap and ChatGPT referral share. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

Methodology note. Last verified 2026-10-13. Engine behavior described above comes from each vendor’s own documentation, dated above. The TwoSetAI audit [4] covered a cohort of past guests, not the full catalog: about seven queries per guest (one Google name search, and two each in Google AI Mode, Perplexity, and ChatGPT), run in incognito windows between August 10 and 25, 2026. A guest counted as present on a surface if any query for it surfaced them. “Citation Gap Audit,” “Buyer Bucket,” “Presence rate” (as used here) are TwoSetAI terms; “Dark Funnel” and “query fan-out” are not. The Stockebrand book is paywalled on O’Reilly Learning; chapter numbers are provided for verification. Nothing here guarantees a citation, a ranking, or a business outcome.