If someone on your team has asked whether your brand shows up in ChatGPT, Gemini and Copilot, you've probably already typed a few questions in and looked at what came back. That's a fair start, but one answer in one tab doesn't tell you much, and it's hard to hand to a leadership team. An AI presence audit is a better way to benchmark brand across AI engines, because it replaces a hunch with a repeatable record. You pick a fixed set of buyer questions, run them the same way in each product, write down what you see, and finish with a short list of places where your brand is missing. This guide walks you through the whole thing in eight steps, and by the end you'll have a completed evidence sheet and a clear picture of where to look first.
This is a one-time snapshot with a start and an end. It isn't the same as setting up ongoing tracking, which is a separate job you can decide on after you've seen the results.
You'll need a spreadsheet, your brand names and domains, a short competitor list, and a few hours over a couple of days. The method here is an ai visibility audit methodology we're proposing as a manageable design, not a published standard, so feel free to scale it up or down as long as you say what you did.
1. Set the audit boundary and comparison conditions
Start by writing down the rules of the audit before you open any AI product. Put these in a short protocol note at the top of your sheet:
- The audit date and the time window you'll run it in
- The locale and language you'll use
- Your brand name, product names, and any aliases people might use
- Your owned domains
- A short list of competitors
- How you'll handle an error, a blank response, a refusal, or a missing source panel
It also helps to decide what you're trying to see: the normal signed-in experience a typical person has, or a more controlled, low-personalization one. Either is fine. Pick one, use it across all three products where you can, and write down anywhere you couldn't match the settings.
Run the comparisons close together in time, since these products and their answers change.
You're done with this step when your sheet has a protocol note that another person could follow, plus a date, time and locale you'll record for every run.
Common mistake: treating "ChatGPT," "Gemini," and "Copilot" as fixed, identical engines. They're changing products with different search, personalization, account, and source-display behavior, so your notes about which surface and settings you used matter as much as the answers.
2. Select buyer prompts before you query anything
The prompts are the most important choice in the audit, so make them before you see any answers. Build the main set from real buyer questions, not from questions that tell the engine which brand to recommend.
One simple template is four topic clusters crossed with three buyer intents, which are awareness, consideration and decision. That gives you 12 prompts. Adapt the clusters to your real products and audience, because generic examples copied from an article won't match how your buyers ask. If you want help finding the questions, our guide on finding the prompts to track goes through where real questions come from, and mapping prompts to the buyer journey shows how to sort them by stage.
Keep every discovery prompt free of your own brand name. If the prompt names you, the engine is more likely to name you back, and you'll be measuring your own question instead of the engine's habits. Then add a small second group of three branded diagnostic prompts, such as asking what your brand does, how it compares with a named alternative, or whether it suits a particular use case. These are for inspecting how you're described and whether the facts are right. They stay out of the mention rate for the unbranded set, and our piece on branded, unbranded and competitor prompts explains why.
Once the list feels right, freeze it. Give each prompt a stable ID, tag it with its topic and intent, and mark it as discovery or diagnostic. From then on, use the exact same wording in every engine and every repeat.
DeepSmith's Discover Prompts can generate candidate questions from your product, persona and buyer-stage context. Treat that as a starting list and keep only the questions that really sound like your buyers. You still need to define and freeze the test set for this one-time audit yourself.
You're done when the sheet has the final prompt list, the IDs and tags, and a clear line between discovery prompts and branded diagnostics.
Common mistake: mixing the branded diagnostic questions into the discovery results, and then saying the brand appears naturally at the measured rate. It doesn't, because you asked about it.
3. Prepare a clean, documented session in each product
Every run should start in a fresh conversation. Don't carry on with follow-up questions, because the earlier turns in a chat can shape the later answers. Keep account, locale and personalization settings as consistent as you can, and write down any exception.
Each product handles this a little differently, so here's what to check in each one.
ChatGPT. Search can turn on automatically when a question would benefit from current information, and you can also start it by hand. The documented route on the web is to open a conversation, choose View all tools, then Search, and then send your prompt. The details are in OpenAI's help page on searching the web with ChatGPT. Decide whether your audit is measuring a search-enabled answer. If it is, use that same control every time and note that you did. If you use Temporary Chat, pick Unpersonalized before you send the first message, since that choice can't be made afterward. It keeps memory, custom instructions and plugins out of that chat.
Gemini. On a computer, the documented path is to go to Gemini Apps and select Temporary chat next to New chat. Temporary chats don't show up in your activity or recent chats and don't give personalized responses. They need a signed-in personal Google Account, so they aren't available for work or school accounts. Gemini also picks a model automatically unless you choose one from the menu, so write down any model label you can see.
Microsoft Copilot. Be clear in your notes that you're auditing the consumer Copilot experience, not Microsoft 365 Copilot signed in with a work or school account. For a signed-in personal account, Microsoft's Copilot privacy controls include personalization and memory settings. If you want a low-personalization baseline, turn off the relevant personalization and Microsoft usage-data options where you find them, and write down what you chose. Turning off the use of your conversations for AI training is a separate control and doesn't switch off personalization.
If your goal is to look like a typical user, don't quietly turn every personalization feature off. Choose the scenario, record it, and keep it consistent. If you want both views, run them as separately labeled groups instead of blending them into one rate.
You're done when the first row for each engine has the session and settings record, and you know every prompt will be run in a fresh conversation.
Pro tip: on ChatGPT, Temporary Chat and Unpersonalized are two separate choices. A temporary chat can still be personalized if that's the option selected.
4. Run identical prompts and repeat them independently
Now you can start collecting answers. Send each frozen prompt to each engine in its own fresh conversation, and then repeat it, again in a new conversation each time. In the design we're using, that's three runs per prompt per engine.
The numbers work out like this. Twelve discovery prompts across three engines with three runs each gives you 108 planned answers. Three branded diagnostic prompts across the same engines and runs gives you another 27. That's 135 planned answers in total, before any retries for errors. This is a scope we chose because it's manageable, not an industry benchmark. A smaller audit is fine, but call it a lighter, noisier snapshot.
Record a timestamp with every run. For ChatGPT, note whether you selected Search yourself or it turned on by itself. For Gemini and Copilot, note whether the answer showed source links or other grounding cues. Keep the full answer text, or a faithful copy of it. Log errors and inaccessible runs separately, retry them only under the rule you wrote in step 1, and keep the original error row so the retry is visible.
You're done when every planned prompt, engine and repeat has either an answer or a clearly labeled failure. A missing cell should never quietly count as "the brand wasn't mentioned."
Why repeats matter: one answer is an observation, not a stable ranking. SparkToro had 600 volunteers run 12 prompts across three tools, for 2,961 runs in all, in its research on how inconsistent AI recommendations can be. A separate study on measuring visibility in AI search reports that answers vary across runs, prompts and days. In that study's campaigns, the source sets from consecutive days overlapped by 34 to 42 percent and the brand sets by 45 to 59 percent, with a wide spread. Those are findings from particular studies, not rates to expect for these three products, and they don't prove that three runs is the right number for everyone.
5. Record mentions, citations and framing in one evidence sheet
Use one row for each prompt, engine and run. Keep the raw observations in their own columns and your coded judgments in separate ones, so a colleague can check your calls later. Here's a layout that works.
| Field | What to enter |
|---|---|
| Audit ID and run timestamp | A unique audit ID, local date and time, and timezone if practical |
| Engine and surface | ChatGPT, Gemini, or consumer Microsoft Copilot, and whether it was the app or the web |
| Account and session condition | Signed-in status, temporary or new chat, personalization setting, locale, and any visible model or mode |
| Prompt ID and exact wording | The frozen wording, topic, buyer intent, and discovery or diagnostic label |
| Run status | Completed, error, refused, unavailable, or another outcome you defined |
| Full answer | The full response text or a faithful copy |
| Brand mention | Yes or no, the exact name or alias, and a short verbatim excerpt |
| Competitor mentions | Names and a supporting excerpt, without guessing from ambiguous terms |
| Framing | Positive, neutral, negative, mixed, or unclear, using the same codebook throughout |
| Claim accuracy | Accurate, partly accurate, unsupported or incorrect, or not verified, plus the evidence you used |
| Sources presented | Each visible URL, its title if shown, its domain, and whether it is owned, a competitor, or other |
| Citation classification | Presented as a cited source, a related link, or a post-answer verification result |
| Owned citation | Yes or no, and the exact owned URL or URLs |
| Reviewer notes | Ambiguity, aliases, broken links, an unexpected mode, or coding doubts |
For framing, copy the relevant text from the answer and then apply your own consistent code. It's your judgment, not a score the engine gave you.
Use a strict rule for citations. Count only a source the product presents as a citation or source for that response, and keep other relevant links in a separate column. The products show sources differently, so the details matter.
- In ChatGPT, a response that used Search can include citations, and the Sources panel can show cited sources and other relevant links. Search results and citations can be incomplete, outdated or wrong, so check the links that matter.
- In Gemini, the Sources panel can mix direct sources with related content. Gemini's double-check feature runs a separate Google Search after the answer, and its links aren't necessarily what Gemini used to write it. Google explains this on its page about viewing related sources and double-checking responses. Put double-check results in a separate verification column and never add them to your citation count.
- In Copilot, record the hyperlinks shown with the answer as source citations when they're there. Microsoft describes web grounding for certain information-seeking conversations, not every conversation, so a missing citation doesn't always mean something went wrong.
Keep a pre-approved alias list for your brand, since name matching can miss abbreviations and generic words in a brand name can cause false positives. Sort out unclear cases by hand.
You're done when each completed answer has its raw text and separate, consistently coded fields for mention, owned citation, competitors and framing.
Common mistake: counting every related link, or a post-answer fact-check link, as a citation the engine used. It makes your owned-citation rate look better than it is.
6. Calculate your rates and show the denominator
With the sheet filled in, you can work out the numbers. The discovery set is your main comparison, and the branded diagnostics stay separate.
For each engine and each prompt, there are two rates.
- Mention rate: completed discovery responses that name the brand, divided by completed discovery responses, times 100.
- Owned-citation rate: completed discovery responses that cite at least one owned page, divided by completed discovery responses, times 100.
Count a response once, even if your name appears in it five times or it links to three of your pages. The individual mentions and URLs stay in the raw columns. For each engine's overall rate, add up the completed discovery responses across all the frozen prompts and repeats and use the same definitions. The guide to measuring your citation rate shows the math from another angle. If you're unsure whether mentions or citations matter more for your goals, this comparison can help you decide.
Report the counts next to every percentage, like "14 of 36 completed answers," and not just the percentage. Leave errors and unavailable responses out of the completed-answer denominator, and report how many there were. An error is not a zero-mention answer.
If you want share of voice, decide first whether each brand counts once per answer or every time it's named, and use that rule for all three engines. Share of voice compares your mentions with the total brand mentions in the set, which is a different thing from the share of answers that mention you. Our walkthrough on calculating share of voice across AI answers covers the counting choices. A simple extra check is to record, for each answer, whether your brand, a competitor, both, or neither appeared. Don't turn the order of a generated list into a "rank," because the order isn't stable.
You're done when, for every engine, the sheet shows the numerator, denominator, rate and failure count for both mentions and owned citations, and the prompt-level rows are still there to explain the totals.
DeepSmith AI Visibility separates mention rate from citation rate and reports prompt and platform visibility, and it can show the answer text and the pages cited for a prompt. That can help you gather evidence and compare competitors. It doesn't run this exact one-time protocol across the three consumer interfaces for you, so the hand-built sheet stays the record of this audit.
7. Compare engines and find the gaps
To get a fair comparison, look at the engines prompt by prompt before you look at the overall rates. Build a simple matrix with prompts down the side and engines across the top. In each cell, mark whether the engine mentioned your brand, cited an owned page, cited a competitor page, and described you accurately.
Then filter for patterns like these:
- Your brand is absent from all three engines on an important unbranded prompt.
- You're mentioned but never cited on an important prompt.
- A competitor is cited while you're absent or uncited.
- The description of your brand is conflicting, outdated or wrong.
- A citation link is irrelevant, broken, or doesn't support the claim in the answer.
- A difference shows up in only one engine, or only in one topic or intent group.
To decide what to deal with first, use a plain rubric you write yourself. Four things are worth weighing: how much the prompt matters to buyers, how often the gap appeared across your sampled answers, how serious any inaccurate framing is, and whether you already have a useful page you could improve. Call this your team's prioritization, not a statistically validated score. A one-off difference is a hypothesis to look into, not proof of how an engine always behaves, and the research on day-to-day noise in AI visibility scores shows why.
When you're deciding who to compare yourself against, our guide to choosing which competitors to track in AI search can help you decide whom to put on the competitor list. DeepSmith's competitor tools show which competitors are cited for your prompts and on exactly which pages, which makes it easier to look into a gap. You still need to check the sampled answers and links yourself, and one competitor winning a single prompt doesn't mean they have a market-wide advantage.
You're done when you have a short list of prompt-level gaps in priority order, each with a supporting answer excerpt, the source URLs saved in the working sheet, and a proposed next action and owner.
Pro tip: if a gap is about wrong facts and not about absence, put it near the top of the list, because a wrong description is easy to point to and act on.
8. Close the snapshot and keep its limits visible
The last step is to stop, because this ai visibility audit methodology is meant to have an end. Save the final prompt list, the session conditions, the completed-answer counts, the raw answers, your coding rules and the summary table together as the audit record. Put the audit date and the exact prompts on the front.
Then write a short conclusion that stays inside what you saw. Say what showed up in this sample, what didn't, where the evidence is uncertain, and which gaps are worth attention. The wording matters here. Don't write "the brand never appears" because you were absent from these answers. Don't claim one product permanently beats another. Write "not observed in this audit sample" or "no owned citation appeared in these completed responses." A no-mention answer, a no-citation answer, a non-web answer and a failed run are four different things, and your summary should keep them apart.
Common mistake: presenting the audit as a permanent ranking. It's a picture of one set of prompts, at one time, under the conditions you wrote down.
What to do next
Your gap list is the starting point for what to write, fix or reposition. A prompt where you're mentioned but never cited usually points to a page that needs to be easier to cite, while a prompt where you're absent and a competitor is cited points to a topic you may need to cover.
If, after the audit, you decide you want a tool to do this kind of checking for you, the best AI visibility tools roundup compares options, and the enterprise shortlist covers larger teams. When you're ready to see your own prompts, answers and cited pages in one place, you can start a DeepSmith free trial and use the real data to compare against what you found by hand.
Related guides
- Buying an AEO Platform: Contract Terms, Onboarding and Multi-Year Commitments: Review the contract terms that matter when buying an AEO platform: renewal and exit clauses, data ownership, onboarding commitments and multi-year pricing, so you can negotiate with a.
- A Phased Rollout Plan for Enterprise GEO Implementation: Lay out a phased enterprise GEO rollout: pilot scope, success gates, team onboarding, expansion order across business units, and the checkpoints that decide when to move to the.
- How to Track Many Products Across AI Assistants and Report Up: Organize prompt sets per product, track each across the major assistants, and roll results into one leadership view that shows which products gain or lose presence without hiding.
- API Answers vs Scraped Answers: Which Is More Accurate for AI Tracking?: Compare API responses with answers captured from the live product on the points that affect tracking: citations shown, web search use, and fidelity to what users see.
- Prompt Volume: Can Anyone Know How Often People Ask AI a Question?: Learn why no one has true prompt volume for AI engines, what estimates exist and how to prioritize without it.


