You can run a credible manual AI visibility audit in an afternoon: about 30 prompts, three or four assistants, clean sessions and a consistent scoring sheet. It won't give you statistically precise share-of-voice numbers, but it will show where you're missing, what assistants get wrong about you, which competitors keep appearing and which sources shape the answers. This post gives you a timed plan, the spreadsheet columns, a scoring rubric and how to turn findings into actions.
What you'll have by the end of the afternoon
A manual audit produces four things you can act on the next morning:
- A mention rate for your brand across the prompts and assistants you tested
- A list of competitors that appear where you don't
- A list of factual errors and outdated claims about your brand
- A tally of the domains assistants cite in your category
| Time | Step |
|---|---|
| 0:00 to 0:30 | Set up the sheet and clean sessions |
| 0:30 to 1:15 | Write the prompt list |
| 1:15 to 2:45 | Run prompts and record answers |
| 2:45 to 3:30 | Score and analyze |
| 3:30 to 4:00 | Write up findings and actions |
Step 1: Set up (30 minutes)
Consistency matters more than tools, so fix your method before you run a single prompt.
Choose platforms. Pick three or four. A typical set is ChatGPT, Google (AI Overviews in regular Search, or AI Mode), Perplexity and one of Gemini, Claude or Copilot. See which AI platform to prioritize.
Use clean sessions. Personalization changes answers. In ChatGPT, use Temporary Chat; OpenAI's help documentation says Temporary Chats don't use or create memories. Elsewhere, use a logged-out or incognito session where the product allows it. Note your location, because local results differ.
Build the sheet with one row per prompt, per platform, per run:
| Column | What to record |
|---|---|
| Prompt ID and text | Exact wording used |
| Prompt type | Category, use case, comparison, alternatives, brand fact, validation |
| Platform | And mode, e.g. "ChatGPT, search on" |
| Run | 1, 2 or 3 |
| Brand mentioned | Yes or no |
| Position | 1st, 2nd, 3rd... in the list, or "body text" |
| Competitors mentioned | Names, in order |
| Accuracy | Correct, minor error, major error, n/a |
| Sentiment | Positive, neutral, negative |
| Sources cited | Domains, and your URLs if cited |
| Notes | Exact wrong claims, notable phrasing |
Step 2: Write the prompt list (45 minutes)
Write prompts the way buyers actually ask, weighted toward unbranded questions where the shortlist forms. Aim for about 30:
- Category (8): "What are the best [category] tools for [audience]?"
- Use case and constraint (8): "Which [category] works for a 10-person team on a budget?"
- Comparison (4): "[Your brand] vs [competitor]: which is better for [use case]?"
- Alternatives (3): "What are good alternatives to [competitor]?"
- Brand facts (5): "How much does [brand] cost?", "Does [brand] integrate with [tool]?", "Where is [brand] based?"
- Validation (2): "Is [brand] worth it?", "[Brand] reviews"
Pull wording from sales calls, support tickets and search queries rather than inventing it. For a longer-term set, see how to build a prompt set worth tracking.
Mark 8 to 10 prompts as "key." These get three runs each; the rest get one.
Step 3: Run the prompts (90 minutes)
Paste each prompt exactly as written, one per fresh conversation, and record the answer before moving on. Tips that keep the data clean:
- Don't ask follow-ups in the same conversation; they're a different test
- Turn on web search where it's a user choice, and note whether it was used
- Copy the full answer text into a second tab for reference; you'll want exact quotes for errors
- Record citations as shown, including when your own pages are cited
- For key prompts, run all three repetitions on each platform
At roughly a minute per answer, 30 prompts across three platforms plus repeats of key prompts fits in about 90 minutes.
Step 4: Score and analyze (45 minutes)
Turn the raw rows into a handful of numbers and lists. Use a simple scoring rule for each answer:
| Score | Meaning |
|---|---|
| 3 | Mentioned first or recommended outright |
| 2 | Mentioned in the top three |
| 1 | Mentioned lower down or in passing |
| 0 | Not mentioned |
| Flag | Any factual error, regardless of score |
Then calculate:
- Mention rate: answers mentioning you divided by all answers, overall and by prompt type and platform
- Average score on category and use-case prompts
- Competitor mention counts, to see who dominates the unbranded prompts. See competitor benchmarking in AI search.
- Cited-domain tally: count how often each domain is cited. This is a quick version of citation analysis.
- Consistency on key prompts: did you appear in all three runs, some, or none?
Treat differences of a few answers as noise. With three runs per key prompt, "appeared in 1 of 3" and "appeared in 2 of 3" aren't meaningfully different; "0 of 9 across three platforms" is. More on this in why AI answers change every time.
Step 5: Turn findings into actions (30 minutes)
Every finding should map to a likely cause and a next step. Common patterns:
| Finding | Likely cause | First action |
|---|---|---|
| Wrong price, feature or fact | Outdated pages or third-party sources | Fix your pages and the cited sources. See fixing AI hallucinations about your brand. |
| Missing on category prompts, present on branded ones | Weak third-party presence | Target the review sites and roundups that were cited |
| Present in ChatGPT, missing in Perplexity | Crawler access or index coverage | Check robots.txt, CDN rules and logs for that crawler |
| Competitor cited from a comparison page | You have no comparable page | Write an honest comparison page |
| Described with old positioning | Inconsistent descriptions across the web | Update profiles and your About page |
Fix errors first; they're the fastest wins and the most damaging if ignored.
Example: a B2B software audit
This example is illustrative; the numbers are made up.
A project management tool runs 30 prompts across ChatGPT, Perplexity and Google AI Mode, with 10 key prompts run three times each.
- Mention rate: about a third of answers overall, but under a fifth on category prompts
- Competitors: two competitors appear in most category answers
- Errors: three answers state a free plan that was discontinued last year
- Cited domains: two software review sites and one industry blog account for many category citations; the brand's own site is cited mainly for pricing
Actions: correct the discontinued free plan everywhere it appears, update review profiles, pitch the industry blog's roundup, and publish two use-case pages for the segments where competitors dominate.
Limits of a manual audit
A manual audit is a snapshot, not a measurement system. Its sample is small, it's hard to repeat identically, and it can't show trends. Use it to find priorities and make the case for investment, then move to repeatable tracking with a fixed prompt set, many runs and regular reporting. To set goals from the baseline, see setting AEO targets.
AEO-Rank automates this process, running your prompt set repeatedly across ChatGPT, Perplexity, Gemini, Claude and Grok.
The bottom line
An afternoon is enough to learn where your brand stands in AI answers: write about 30 buyer-style prompts, run them in clean sessions across three or four assistants, repeat the key ones, score every answer the same way, and tally competitors and cited sources. Fix factual errors first, tackle the biggest gaps on high-value prompts next, and then set up tracking so the next audit measures progress instead of starting over.
For the free tools that support this audit, see free AEO and GEO tools.
