How to Run an AI Citation Audit for Your Brand

Ask ChatGPT for the best products in your category and one of two things happens: your brand is in the answer, or a competitor's is. Most merchants have never actually checked. They assume AI visibility is either fine or hopeless, and both assumptions are usually wrong. An AI citation audit replaces the assumption with a number - a repeatable measurement of whether assistants name you, cite you, or skip you - plus a diagnosis of why. This is the exact process we run. It takes an afternoon the first time and about an hour on every re-run.

What is an AI citation audit?

AI citation audit

A structured test of how AI assistants treat your brand when shoppers ask buying questions in your category. You run a fixed panel of questions across the major assistants and score each answer on a three-level scale - named in the recommendation, cited as a source, or absent - producing a score you can track and a gap log you can act on.

The output is two artifacts. First, a percentage score: how much of the available citation surface you actually hold across your question panel. Second, a gap log: exactly which questions you lose, on which assistants, to which competitors. The score tells you whether you have a problem. The log tells you what kind.

The audit is the measurement half of Commerce GEO. Everything else in GEO - structured data, content shape, third-party corroboration - is treatment. Treatment without measurement is guesswork: you can't tell whether the work moved anything, and you can't prioritize because you don't know where the losses are. The audit comes first, then the fixes, then the audit again.

How do I check if ChatGPT recommends my brand?

The fast manual check: open a fresh ChatGPT conversation and ask the question your customer would ask - not "tell me about [your brand]" but "what's the best [category] for [use case]?" Then ask two or three variants. Watch two things: whether your brand appears in the answer body, and whether your site shows up among the cited sources when the assistant browses. Then repeat the same questions in another fresh chat, because answers vary run to run.

If you want the fast version of the whole audit, our free AI Citation Check runs a starter panel for your brand across the major assistants and shows where you're named, cited, or absent - in minutes instead of an afternoon. The manual audit in the rest of this guide is the deep version: it tells you why you're losing and what to fix first.

Build a panel of 15-30 buying questions

Buying-question panel

A fixed, versioned list of 15-30 questions phrased the way real shoppers ask them, mostly without your brand name in the question. Locking the wording is what makes results comparable between audit runs.

The audit is only as good as its questions. You want the questions shoppers actually ask, spread across five types:

• Head terms. "What's the best [category] under [price]?" - the broadest, most contested queries. Expect to lose most of these at first.
• Comparisons. "[Competitor] vs [competitor]" and "best alternatives to [category leader]" - where assistants weigh brands directly against each other.
• Use cases. "What should I buy for [specific job or person]?" - high intent, less contested, usually the first questions you can win.
• Spec and compatibility. "Does [product type] work with [thing]?" - answered from structured product data, which is why enriched catalogs dominate here.
• Trust. "Is [your brand] legit?" - the one branded question every panel needs, because assistants answer it whether you like the answer or not.

Mine the questions from places where shoppers already talk: on-site search logs, support tickets, paid-search query reports, and what prospects ask on sales calls. Phrase each one the way a person types, not the way a marketer writes. Then lock the wording. A panel you keep rephrasing can't show you a trend.

Run the panel across four assistants

Run every question on ChatGPT, Claude, Gemini, and Perplexity. Four surfaces, not one, because each fails differently - and the pattern of where you're absent is itself diagnostic. A brand missing everywhere has a content problem; a brand missing on only one surface has a retrieval problem on that surface.

1. 1

   Use clean sessions

   Fresh conversation per question, memory and personalization off where the assistant allows it. You're measuring the default answer, not your own echo.

2. 2

   Ask verbatim

   Same wording on every assistant, every run. Paraphrasing between runs invalidates the comparison.

3. 3

   Run each question more than once

   Assistant answers are not deterministic. Score the pattern across runs, not a single lucky or unlucky response.

4. 4

   Log everything

   One spreadsheet row per question per assistant per run: brands named, sources cited, and the verbatim recommendation sentence. The verbatim text is what you'll study later.

Score every answer: named, cited, or absent

Keep the rubric brutally simple. Three levels, three point values:

Sum the points across every question and assistant, divide by the maximum possible, and you have your citation score - a single percentage you can put on a dashboard and re-measure monthly. Most brands score far lower than they expect on the first run. That's normal; the first audit exists to set the baseline, not to flatter you.

While you score, keep a second tally: which competitors get named, and how often. The three or four names that keep appearing are your real competitive set in the AI channel - often different from your competitive set in paid search - and their content is your diagnostic material for the next step.

Why does AI recommend my competitor?

Because their content is easier for a model to retrieve, trust, and quote. In nearly every audit we run, the gap comes down to three properties - readability, corroboration, and quotability - and almost never to brand size or ad spend. Assistants don't know your revenue. They know what they can fetch, verify, and lift into an answer.

Readability

Can AI systems fetch and parse your pages at all? Blocked crawlers, JavaScript-only content, and prose-buried specs all make you invisible at the retrieval stage - you can't be cited if you can't be read. The fixes are mechanical: crawler access, structured data for AI shopping, and a machine-readable site map of what matters via llms.txt.

Corroboration

Models weight claims that are confirmed off your domain. A claim that lives only on your site is an assertion; the same claim echoed in reviews, best-of lists, and forum threads is a fact the model will repeat. Brands that win the memory-based answers - the ones ChatGPT gives without browsing - almost always have the deepest third-party footprint.

Quotability

When an assistant does read your page, is there a sentence it can lift? Clear claims with numbers, spec tables, and answer-shaped headings get quoted. Ten paragraphs of brand storytelling do not. The winning competitor's page usually contains the exact sentence the assistant used - go look, it's instructive.

The full treatment playbook for each lever is in getting your products recommended by AI. The audit's job is to tell you which lever to pull first.

Fix the highest-leverage gap first

Your gap log will show a pattern. Match it to the fix instead of trying to do everything at once:

• Absent everywhere, on every assistant. A retrieval problem. Start with readability: crawler access, structured data, parseable product pages.
• Cited but rarely named. Assistants find you but won't recommend you. A quotability problem: your pages lack the clear, liftable claims a model can put in an answer.
• Named on Perplexity, absent from ChatGPT's no-browse answers. A corroboration problem. Retrieval-based surfaces see you; model memory doesn't. Third-party mentions fix this, slowly.
• Losing only spec and compatibility questions. A product-data problem - the answer lives in fields you haven't populated. That's Catalog Enrichment for AI work, not content work.

Then pick one. The pattern that accounts for the most lost points gets the fix; everything else waits. A single fix shipped completely and re-measured beats five fixes started. This discipline is the difference between an audit that changes your score and an audit that becomes a slide.

Re-audit on a cadence

One audit is a snapshot; the loop is the asset. Run the scoring pass monthly and the full diagnosis quarterly. Keep the panel wording locked between runs - if you need new questions, version the panel and track the versions separately so the trendline stays honest.

Set expectations by surface. Retrieval-based answers - Perplexity, ChatGPT when it browses - can move within weeks of a readability fix. Model-memory answers move on the timescale of months, because they depend on corroboration accumulating across the web. The normal progression for a losing question is absent, then cited, then named. Cited is not the goal, but it is progress: the assistant now reads you and just doesn't trust you enough yet.