Skip to content
Chalk Labs / AI SEO
How-to guide

How to Audit Your Brand in AI Search

By Rahil Jain, founder of Chalk LabsUpdated How we research

The short answer

To check what ChatGPT says about your brand, write 20 to 30 realistic customer prompts, run each one in ChatGPT, Perplexity, Gemini and Google AI Mode, and record four things per answer: whether you are named, whether the facts are right, which sources are cited, and which competitors appear. Repeat every month, because one answer proves little.

  • A brand audit in AI search compares what assistants say about you with what is true and with what they say about competitors.
  • Test 20 to 30 prompts across four groups: brand, category, comparison and problem prompts. One question tells you almost nothing.
  • Score presence and accuracy, and list sources and competitors. The sources list tells you where to fix things.
  • Expect three to four hours for a first audit. Re-run it monthly and compare against the first sheet.

What is a brand audit in AI search?

A brand audit in AI search is a structured check of what AI assistants say when a buyer asks about you or your category. You record whether you appear, whether the facts are right, which pages the assistant cites, and which competitors show up instead.

The output is a gap list: a ranked set of problems such as a wrong price, a missing mention, or a competitor who owns a source you do not. That list is the useful part. A screenshot of one good answer is not.

This guide covers the manual method. It needs no paid tool, only a spreadsheet and three to four hours. If you want the long-running version of the same idea, see the guide to setting up an AI visibility tracking spreadsheet.

Do it step by step

0 of 8 done0%

  1. Write down three to five direct competitors and a one-page fact sheet about your own business: services, prices, locations, founding year, and what you do not do. This is the answer key you will score against.

    Tip: Include what you do not offer, because assistants sometimes credit businesses with services they lack.

  2. Write five to eight prompts each for brand, category, comparison and problem questions. Use the words customers use. Leave your name out of everything except the brand and comparison groups.

    Tip: The free AI prompt list builder can draft a first set from your category.

  3. Use a private window or a logged-out session where possible. Note the date, tool, country and whether you were signed in. Start a new chat for every prompt so earlier answers do not colour later ones.

  4. Run each prompt in ChatGPT, Perplexity, Gemini and Google AI Mode. Paste the full answer and the list of cited links into your sheet. Do not summarise yet.

    Tip: Run in batches by tool, not by prompt, so you stay in one interface and finish faster.

  5. Read each answer against your fact sheet. Give a 0 to 3 mark for presence and for accuracy. Copy each wrong sentence into a notes column word for word.

  6. List every domain cited. Count how often each appears across all answers. Do the same for every competitor named. The top five domains are your fix list for outside pages.

  7. Open each link that points to your own domain and each one in the top five. Mark any that return an error page or the wrong content. Plan a redirect or a correction for each.

  8. Write five to ten fixes in order: wrong facts first, then missing presence in your best-fit prompts, then source gaps. Put an owner and a date on each. Save the sheet as month zero.

    Tip: Re-run the same prompts in 30 days and compare against this sheet.

Your ticks are saved in this browser only.

Copy-ready templates

Starter prompt list (replace the brackets)

text
BRAND
What does [brand] do?
Who is [brand] best for?
How much does [brand] cost?
Is [brand] any good? What do customers say?
Where is [brand] based and what areas does it serve?

CATEGORY
Best [category] for [customer type] in [place]
Who offers [service] with [key feature]?
Top [category] providers, with pros and cons

COMPARISON
[brand] vs [competitor 1], which should I choose?
[brand] alternatives
[option A] or [option B] for a [customer type]?

PROBLEM
[Problem phrase a customer would say], who can help?
How do I [job the customer wants done] without [pain point]?

Add or remove lines, but keep the groups. Freeze the list after the first audit.

Audit sheet columns (paste into row 1 of a spreadsheet)

text
date,tool,country,signed_in,group,prompt,named_yes_no,presence_0_3,accuracy_0_3,wrong_sentences,cited_domains,own_domain_cited_yes_no,competitors_named,notes

One row per prompt per tool. Fourteen columns are enough.

Gap list format

markdown
| Rank | Gap | Evidence (prompt, tool) | Fix | Owner | Due |
|------|-----|--------------------------|-----|-------|-----|
| 1 | Wrong price on directory | Brand prompt 3, ChatGPT and Gemini | Update listing, add price page | | |
| 2 | Not named for category prompts | Category prompts 1 to 4, all tools | Publish a comparison page | | |
| 3 | Competitor owns top review site | 9 of 20 answers cite it | Request a listing | | |

Evidence must point to rows in your sheet, so every fix is traceable.

What do you need, and how long does it take?

You need three to four hours, a spreadsheet, a list of your three to five closest competitors, and free or paid access to at least ChatGPT, Perplexity, Gemini and Google AI Mode. Claude and Microsoft Copilot are worth adding if your buyers use them.

Use a private browser window where you can, so your own history shapes the answers less. Write down the date, the tool and the location for every run. Those three details make a later comparison fair.

Time budget for a first audit
TaskTimeOutput
Write 20 to 30 prompts40 minutesA prompt list in four groups
Run prompts in four tools90 minutesRaw answers pasted into the sheet
Score answers30 minutesTwo scores and two lists per answer
Check sources and facts30 minutesA list of wrong facts and missing sources
Write the gap list20 minutesFive to ten ranked fixes

Why does one question tell you so little?

One question tells you little because assistants decide for themselves whether to search and how many searches to run. OpenAI says its model can choose to search the web or not based on the prompt, and may keep searching in agentic mode 1. Anthropic says simple factual queries typically use 1 to 3 searches, while comparative or multi-entity research can use 10 or more 2. Google says Gemini can generate one or several search queries from a single prompt 3.

Answers also move for other reasons. A June 2026 Profound study of 753 prompts across 7 platforms named three sources of variation: random text generation, sensitivity to rewording, and platform drift after model updates 4. The same study found that a once-a-day reading gave a visibility score of 78.7%, against 80.4% for ten runs a day 4.

The practical lesson is to test many prompts once rather than one prompt many times. Profound concluded that which prompts you track matters more than how often you run them 4. Ahrefs makes a similar point: track a sample of prompts and see who gets recommended, because single snapshots are unreliable 5.

Which prompts should you test?

Test four groups of prompts: brand, category, comparison and problem. Aim for five to eight in each group, written the way a customer would type them, not the way your marketing team would.

Brand prompts check basic facts. Category prompts check whether you are named without being asked. Comparison prompts check how you are framed against rivals. Problem prompts catch buyers earlier, before they know your category name.

Leave your brand name out of the category and problem prompts. Comparison prompts may name you, but also include one or two without your name, such as 'Online bookkeeper or local bookkeeper for a three-person business?', so you can see whether you are picked unprompted.

Example prompt set for a fictional bookkeeping firm, Harbour Books
  • Brand: What does Harbour Books do?
  • Brand: Is Harbour Books good for small cafes?
  • Category: Best bookkeeping firms for small restaurants in Leeds
  • Category: Who offers fixed-fee bookkeeping for sole traders?
  • Comparison: Harbour Books vs Ledger Lane, which is better for a startup?
  • Comparison: Online bookkeeper or local bookkeeper for a three-person business?
  • Problem: My receipts are a mess before tax time, who can sort it?
  • Problem: How do I stop chasing invoices every month?

How do you score an answer?

Score presence and accuracy 0 to 3; record sources and competitors as lists. Keep each score simple, so two people would give the same mark.

Presence asks whether you are named and where. Accuracy asks whether every fact about you is right. Sources asks which URLs the assistant cites. Competitors asks who else is named. Add a free-text notes column for anything odd, such as an old address.

Seer Interactive argues that a bare visibility metric counts a positive and a negative mention the same 6. That is why accuracy and tone get their own columns here, rather than being folded into one number.

A simple scoring rule
Column0123
PresenceNot namedNamed in a list of manyNamed in the top threeNamed first or recommended
AccuracyMostly wrongSome key facts wrongMinor errorsAll facts right

What do the cited sources tell you?

The cited sources show where the assistant learned about you, and they are the fastest route to a fix. If a directory, review site or old press article is cited for a wrong fact, correcting that page often matters more than editing your own site.

Also check that every cited link works. In an Ahrefs study of URLs cited in answers that Ahrefs could match to its crawler database, ChatGPT cited pages that returned a 404 error 2.38% of the time, against 0.84% for the top 20 Google results 7. Ahrefs says some 404s may have other causes than made-up links, and that the true rate is probably higher, because it could only check the roughly 65% of cited URLs that were in its crawler database 7. Treat the rates as a warning, not a measurement of your own site. If assistants send people to a broken URL on your domain, add a redirect.

Finally, confirm the bots can read your site. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers 8. Run the AI crawler checker or follow the guide on checking whether AI bots can crawl your site.

What mistakes ruin an AI brand audit?

The most common mistake is testing only brand prompts. Assistants often recognise your name once you type it, though small or new brands may return little. The real question is whether they name you when a stranger asks for a recommendation.

The second mistake is keeping only the screenshots you like. Record every answer, including the bad ones, in the same sheet. The third is changing the prompt wording between audits, which makes month-to-month comparison meaningless. Freeze the list and add new prompts at the end.

A fourth mistake is treating the audit as a verdict. It is a snapshot with noise in it. Use it to find patterns, such as a wrong fact repeated in four answers, and ignore single odd results.

What do you do with the gap list?

Sort the gaps into three piles: wrong facts, missing presence and competitor-owned sources. Fix wrong facts first, because they cost trust. Our guide on fixing wrong information AI says about your brand walks through that.

For missing presence, publish a page that answers the prompt directly, and read how query fan-out breaks a question into searches so you can cover each one. For competitor-owned sources, find out whether you can be listed or reviewed there. Then re-run the same prompts in a month.

When is a manual audit not worth doing?

Skip the manual audit if your brand is brand new with no web footprint, because the answers will simply be empty. Build the basic pages and profiles first. Skip it too if you already track 100 or more prompts with a paid tool, since the sheet adds little.

For ongoing measurement, the guide to measuring AI share of voice covers the repeatable version.

Frequently asked questions

How do I check what ChatGPT says about my brand?

Open a new ChatGPT chat and ask what your brand does, who it suits and how it compares with named rivals. Then ask category questions without your name, such as the best provider in your city. Record the answers and cited links in a sheet, and repeat in Perplexity and Gemini.

How many prompts should I test?

Test 20 to 30 prompts for a first audit, spread across brand, category, comparison and problem questions. Profound's 2026 study found that the choice of prompts mattered more than how often each one was run, so variety beats repetition 4.

Do I need a paid tool?

No. A spreadsheet and three to four hours are enough for a first audit. Paid trackers help once you want daily runs across many prompts, but they cannot tell you which prompts matter. That judgement still comes from you.

Why do the answers change between runs?

Assistants generate text with some randomness, react to small wording changes, and change as models and the live web change 4. That is why you look for patterns across many prompts instead of trusting one answer.

How often should I repeat the audit?

Repeat monthly with the same frozen prompts and add new ones at the end. Profound's study compared one run a day with ten, not monthly with daily 4. Monthly is our practical suggestion for a small team, not a tested interval.

About the author

Rahil Jain

Rahil Jain is an AI SEO (answer engine and generative engine optimisation) consultant and co-founder of Chalk Labs.

LinkedIn · Method and credentials

Sources

  1. Web search (tools guide) · OpenAIOpened October 20, 2026. Supports: The model chooses whether to search based on the prompt, and in agentic mode can keep searching.
  2. Web search tool · AnthropicOpened October 20, 2026. Supports: Simple factual queries typically use 1 to 3 searches; comparative or multi-entity research can use 10 or more.
  3. Grounding with Google Search · Google AI for DevelopersOpened October 20, 2026. Supports: Gemini can generate one or multiple search queries from a prompt and ground its answer in the results.
  4. Is once a day enough? · ProfoundOpened October 20, 2026. Supports: 753 prompts across 7 platforms, 1 to 14 June 2026; visibility 78.7% once daily vs 80.4% at ten runs; sources of variation; prompt choice matters more than run frequency.
  5. How to Measure AI Search Visibility When Attribution Falls Short · AhrefsOpened October 20, 2026. Supports: Recommends tracking a sample of prompts; notes AI answers are volatile and single snapshots unreliable.
  6. AI Visibility Is Lying to You: There's No One-Size-Fits-All AI Search KPI · Seer InteractiveOpened October 20, 2026. Supports: A visibility metric counts a positive and a negative brand mention the same; proposes accuracy as a separate measure.
  7. New Study: How Often Do AI Assistants Hallucinate Links? (16 Million URLs Studied) · AhrefsOpened October 20, 2026. Supports: ChatGPT cited 404 URLs 2.38% of the time vs 0.84% for the top 20 Google results; some 404s may have other causes; the true rate is probably higher because only about 65% of cited URLs were in Ahrefs' crawler database.
  8. Overview of OpenAI Crawlers · OpenAIOpened October 20, 2026. Supports: Sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
How this page was made. Built on Rahil Jain's AI SEO method and drafted with AI research assistance. Every statistic links to the page we read it on (opened October 20, 2026 or nearby). Published October 20, 2026; last updated October 20, 2026. Chalk Labs sells AI SEO services, so we have a commercial interest in this topic. Spotted an error? Tell us and we will fix it.