How to Check What AI Says About Your Company at a Glance
| Step | Action | Expected Result |
|---|---|---|
| 1 | Write a reference fact sheet about your company | A list of facts to judge answers against |
| 2 | Build a fixed prompt set: brand, category, comparison, reputation | 15 to 30 prompts with fixed wording |
| 3 | Set up clean, recorded test conditions | Each run is logged with account state, mode and location |
| 4 | Run every prompt in each assistant, with and without web search where possible | Several answers per prompt per assistant |
| 5 | Log answers and cited sources in a table | One row per run, with claims and URLs |
| 6 | Check each claim against the fact sheet | Every claim marked correct, outdated, wrong or unsupported |
| 7 | Trace wrong claims to their sources | A list of pages to fix, correct or report |
| 8 | Re-check on a fixed cadence and after events | A dated history showing what changed |
| Time | Tools | Cost | Skill level |
|---|---|---|---|
| Half a day for the first audit; one to two hours per re-check | ChatGPT, Perplexity, Gemini, Google Search, a spreadsheet | Free tiers are enough to start; paid plans add models and usage | No technical skills; careful record-keeping matters most |
What You Need Before Starting
- A spreadsheet with the log columns shown in Step 5.
- Access to ChatGPT, Perplexity, Gemini and Google Search, ideally both signed in and signed out.
- Your company’s correct facts in one place (Step 1).
- Two to four named competitors and the categories buyers use to describe what you sell.
- A note of the country you are testing from. Results can differ by location, and AI Overviews availability varies by country and language according to Google’s help page.
What Is an AI Answer Audit?
An AI answer audit is a repeatable test in which the same fixed prompts about a company are run through AI assistants under recorded conditions, and every claim and cited source is logged and checked for accuracy. It is a sample, not a measurement of what every user sees. Answers change between runs, between accounts and over time, so the value is in the method and the dated record, not any single answer.
Step 1: Write a Reference Fact Sheet
Write down the facts an accurate answer about your company should contain before you run a single prompt. Without it, you end up judging answers by whether they sound right. Include the legal and trading name, founding year, headquarters, leadership, what you sell and to whom, pricing model if public, notable customers or coverage you can cite, and anything commonly confused (a similarly named company, a former name, a discontinued product).
Add the source for each fact: the page on your site, the filing or the article that states it. Those sources become the pages you check in Step 7.
Expected result: a one-page sheet of 15 to 30 facts, each with a source URL. How to verify: someone outside marketing confirms each fact is current. Common mistake: putting positioning statements (“the leading platform”) on the sheet. An assistant that does not repeat your slogan is not wrong.
Step 2: Build a Fixed Prompt Set
Create a list of prompts in four groups and freeze the wording. Changing wording between rounds makes results impossible to compare.
| Group | What it tests | Example prompts |
|---|---|---|
| Brand | Whether the assistant knows you and gets the basics right | “What is [Company]?” “Who founded [Company] and when?” “What does [Company] charge?” |
| Category | Whether you appear when buyers ask about the category, without naming you | “What are the best [category] providers for [use case]?” “Which [category] tools do [industry] companies use?” |
| Comparison | How you are described next to competitors | “[Company] vs [Competitor]: which is better for [use case]?” “Alternatives to [Competitor]” |
| Reputation | What risks and complaints the assistant surfaces | “Is [Company] legitimate?” “What are common complaints about [Company]?” “Has [Company] been in the news?” |
Write prompts the way a buyer would, not the way you would describe yourself. Avoid leading prompts such as “Why is [Company] the best?”; they test the assistant’s willingness to agree, not what it knows. Give each prompt an ID (B1, C1, and so on) for the log.
Expected result: 15 to 30 prompts with IDs (our suggested range, not a published standard). How to verify: every group has at least three prompts, and no prompt contains claims you want the assistant to repeat. Common mistake: only testing brand prompts. If you are absent from category answers, the brand prompts will not show it; why a company does not appear in ChatGPT covers the usual causes.
Step 3: Set Up Clean, Recorded Test Conditions
Decide the account state for each run and record it. Personalisation, memory and location can all change answers, so the same prompt can produce different results for you and a prospective customer.
- ChatGPT. OpenAI’s temporary chat help page says a temporary chat can be started as Unpersonalized, which does not use memory or custom instructions, or Personalized. OpenAI’s web search help page says people who are not signed in can also use web search, that ChatGPT may use an approximate location based on IP address, and that if memory is on, saved memories may be used when it rewrites a search query.
- Gemini. Google’s Gemini Apps help says a temporary chat does not give personalised responses and is available only to signed-in personal Google Accounts; some features can be used without signing in, depending on location and device. It also notes that in a temporary chat you cannot give feedback on a response, which matters in Step 7.
- Perplexity. Its help centre says it remembers context within a conversation, so start a new thread for each prompt.
- Google AI Overviews. Use a normal Google search. Google says AI Overviews appear only when its systems judge them helpful and often do not trigger, so record whether one appeared at all.
For each run, note the date, assistant, signed in or out, personalisation setting, model if shown, and country. A VPN or network location can change the location ChatGPT infers, according to OpenAI.
Expected result: a documented test setup you can repeat. How to verify: a colleague could reproduce the run from your notes. Common mistake: testing only in the founder’s long-used personal account, where memory and past conversations can shape answers.
Step 4: Run Each Prompt With and Without Web Search
Run every prompt in each assistant, in a fresh conversation, and repeat it. Answers drawn from live web results and answers drawn from the model alone can differ, and a wrong claim has different causes in each case.
- ChatGPT: OpenAI says ChatGPT may search the web automatically. To force a search, select View all tools, then Search, or type / in the composer and select Search. There is no documented switch that guarantees no search, so record whether each answer shows citations.
- Perplexity: searches the web for its answers and includes numbered citations, according to its help centre. Its Pro Search and Research modes search more extensively; note which mode you used.
- Gemini: record whether a Sources button appears. Google’s sources help page says not all responses include related links or sources.
- Google AI Overviews and AI Mode: Google’s documentation for site owners says the two may use different models and techniques, so the responses and links they show will vary. Treat them as separate rows.
Run each prompt at least three times per assistant and setup. Three is our working minimum, not a published standard; the point is to see whether an answer is stable or changes from run to run.
Expected result: several answers per prompt per assistant. How to verify: each run has a screenshot or copied text saved with a run ID. Common mistake: continuing in one long conversation, where earlier prompts colour later answers.
Step 5: Log Answers and Cited Sources in a Table
Record one row per run in a spreadsheet. The log is the deliverable: it is what you compare in later rounds and what you hand to whoever fixes the sources.
| Column | What to record |
|---|---|
| Run ID and date | For example 2026-10-06-ChatGPT-B1-r2 |
| Assistant and mode | Product, model if shown, search on or off (or citations present), Pro Search or Research |
| Account state and country | Signed out, temporary unpersonalised, signed in; location |
| Prompt ID | From Step 2 |
| Mentioned? | Yes or no; for category prompts, where in the list |
| Claims about you | Each factual statement, one per line |
| Accuracy | Filled in at Step 6 |
| Cited URLs | Every source link shown, in order |
| Notes | Competitors named, tone, anything odd |
Expected result: a complete log, with screenshots filed by run ID. How to verify: no row lacks a date, mode or prompt ID. Common mistake: summarising answers instead of recording the actual claims. A summary hides the specific wrong figure you need to trace.
Step 6: Check Each Claim Against Your Fact Sheet
Mark every claim with one label: correct, outdated, wrong, unsupported by the cited source, or confused with another entity. Open the cited page for each claim that names one. OpenAI’s help page advises users to open a cited source to check it supports the answer, and says search results and citations can be incomplete, outdated or incorrect. Google says AI Overviews “can and will make mistakes”.
Expected result: each claim labelled, with a short reason. How to verify: count the wrong and outdated claims by assistant and prompt group, and note which prompts produce them repeatedly. Common mistake: marking a claim correct because the cited page is about you. The page has to say the specific thing claimed; research summarised in how often AI search tools cite the wrong source shows attribution errors are common.
Step 7: Trace Wrong Claims to Their Sources
Find where each wrong claim came from, because that decides the fix. Work through these cases in order:
- The cited page states the wrong fact. Fix it at source. If it is your own site, update it. If it is a directory or profile you control, edit it. If it is a news article, request a correction; our guide on how to request a correction to a news article covers the process.
- The cited page does not say it. This is a misattribution. Report it through the product’s feedback controls. For AI Overviews, Google’s help page describes the thumbs-down icon followed by Report a problem; the feedback includes your query and its results.
- No citation, and the claim appears only without web search. It likely reflects older data the model learned from. Search the web for the exact wrong phrase to find pages still carrying it.
- The assistant confuses you with another organisation. Note which details are merged, and check whether your own pages state clearly who you are and are not.
Patterns in which sources assistants rely on are covered in which sources AI assistants actually cite. For your own site to appear as a source, OpenAI says it must allow OAI-SearchBot to crawl it; Google says a page must be indexed and eligible to show with a snippet, with no additional technical requirements for AI Overviews or AI Mode.
Expected result: a list of source pages, each with an owner and an action. How to verify: every wrong claim has a traced source or is marked untraceable. Common mistake: sending feedback without fixing the source. If the page still says it, the error can come back.
Step 8: Re-Check on a Fixed Cadence
Re-run the same prompt set under the same conditions on a schedule, and add an extra round after events that change the facts: a launch, a leadership change, a major article, a correction or a crisis. Monthly is a reasonable starting cadence for most companies; this is our suggestion, not a published benchmark. Compare against earlier rounds by prompt ID, not by impression. Fixes take time to show: Google says recrawling can take from several days to several months. Ongoing tracking is covered in watching what search and AI say over time.
Expected result: a dated series of rounds showing which claims changed. How to verify: each traced error has a later run showing it fixed, unchanged or new. Common mistake: editing the prompt set every round. Add new prompts as new IDs and keep the old ones.
How to Choose a Test Setup
| Situation | Recommended Method | Why |
|---|---|---|
| You want to see what a new prospect sees | Signed out, or ChatGPT temporary chat set to Unpersonalized; Gemini temporary chat | Removes memory and personalisation from your own account |
| You want to know where a claim comes from | Force web search and open every citation | Search answers show sources you can check |
| You suspect outdated training data | Compare answers with and without citations | Uncited claims that persist are less likely to come from a live page |
| You sell in several countries | Repeat key prompts from each market, recording location | ChatGPT uses approximate location; AI Overviews availability varies by country |
| You need a monthly view across many prompts | A monitoring tool or a scripted process, spot-checked by hand | Manual runs do not scale; hand checks catch tool errors |
| A single urgent error | Run the one prompt three times in each assistant now | Confirms whether it is persistent before you act |
Number of countries and territories on Google's availability list, read October 10, 2026
Source: Google Search Help, AI Overviews in Google Search, availability section.
249 countries and territories in total, counted from Google's list, which also names 98 supported languages. Being on the list does not mean an AI Overview appears for a given query; Google shows one only when its systems judge it helpful.
How to Verify It Worked
- You have a log with one row per run, covering every prompt, assistant and setup, each with a date and saved screenshot.
- Every claim is labelled, and every wrong claim has a traced source, an owner and an action, or is marked untraceable.
- Feedback submissions are recorded with date and run ID.
- A later round, run under the same conditions, shows whether each error is fixed, unchanged or new.
Common Problems and Fixes
| Problem | Likely Cause | Fix |
|---|---|---|
| The same prompt gives different answers | Normal variation between runs | Run three or more times and record the spread, not one answer |
| A colleague sees a different answer | Memory, personalisation, location or a different model | Compare test conditions; use unpersonalised or signed-out runs |
| ChatGPT does not cite anything | It answered without searching | Use View all tools, then Search, and run again |
| No Sources button in Gemini | Not all Gemini responses include sources | Record “no sources”; rephrase only as a separate prompt ID |
| No AI Overview appears | AI Overviews often do not trigger, and availability varies by country | Record its absence; it is a valid result |
| Cited page does not support the claim | Misattribution | Report through the product’s feedback controls; keep the screenshot |
| You are merged with a similarly named company | Entity confusion | Make your own pages state clearly who you are; trace the merged details |
| Fixed source, answer unchanged | Page not yet recrawled, or another copy still carries the error | Wait, re-check, and search for remaining copies |
What Not to Do
- Do not judge from one run. A single answer can be unrepresentative. Repeat and record the spread.
- Do not test only in a personalised account. It shows what you see, which may differ from what a buyer sees.
- Do not turn a small sample into a statistic. Thirty prompts run three times is a diagnostic, not a market share figure. Report it as “in our test on [date]”.
- Do not use leading prompts. They measure how agreeable the assistant is, not what it knows.
- Do not stop at feedback. If the source still states the error, fix the source.
- Do not expect a guaranteed result. No fix guarantees a mention or citation; OpenAI says placement in ChatGPT search results is not guaranteed, and Google says indexing and serving are not guaranteed. Ways to improve your odds are covered in how to get your company mentioned by ChatGPT.
Frequently Asked Questions
Do different people get different AI answers about the same company?
Yes. Answers vary between runs, and can vary with account, memory, personalisation and location. OpenAI says ChatGPT may use approximate location from an IP address and, if memory is on, saved memories when rewriting search queries.
Can I check what ChatGPT says without logging in?
Yes. OpenAI's help page says people who are not signed in can use ChatGPT web search. A temporary chat set to Unpersonalized is another way to remove memory and custom instructions from a signed-in test.
How many prompts should an AI answer audit include?
Fifteen to thirty prompts across brand, category, comparison and reputation groups is a practical size for a first audit. That range is our working suggestion, not a published standard.
Does Perplexity always search the web?
Perplexity's help centre describes it as searching the internet in real time for each answer and including numbered citations linking to the original sources.
Why does an AI answer cite a page that does not say what the answer claims?
AI search tools sometimes attribute claims to sources that do not support them. OpenAI says search results and citations can be incomplete, outdated or incorrect and advises opening the cited source to check it.
Can I make ChatGPT or Gemini correct a wrong answer about my company?
Not directly. You can fix the source pages the answer relies on, request corrections from publications, and report errors through each product's feedback controls. None of these guarantees a change, and it can take time to appear.
Does Google Search Console show AI Overviews traffic?
Google says sites appearing in AI Overviews and AI Mode are included in overall search traffic in the Search Console Performance report, under the Web search type. They are not reported as a separate line in that documentation.
How often should a company re-check what AI assistants say about it?
Monthly is a reasonable starting point, with extra checks after launches, leadership changes, major coverage, corrections or a crisis. Keep the prompt set and test conditions the same so rounds can be compared.
Sources and Methodology
This guide is based on the official help and documentation pages listed below, read on 6 October 2026. It describes a procedure; we did not run an audit for this guide and report no results from one. Product interfaces and labels change, so the menu names quoted are those in the help pages on that date. The prompt counts, the three-run minimum and the monthly cadence are our suggestions, not published standards.
- official OpenAI Help Center, Searching the web with ChatGPT.
- official OpenAI Help Center, Temporary chat in ChatGPT.
- official Perplexity Help Center, How does Perplexity work?, last modified 3 September 2026.
- official Gemini Apps Help, View related sources from Gemini Apps.
- official Gemini Apps Help, Use Gemini Apps (temporary chat and sign-in).
- official Google Search Help, AI Overviews in Google Search.
- official Google Search Central, AI features and your website.
Limitations. AI products change frequently and roll out features by plan and country. Microsoft Copilot, Claude and other assistants are not covered here but the same method applies.