Your First Manual Visibility Check

A first manual visibility check takes about an hour. Pick 10 real buyer prompts, run each one twice in two AI engines using clean sessions, and record mentions, recommendations and citations in a fixed sheet. It won't give you precise rates. It will show you the big patterns: whether you appear at all, which competitors get named instead, and which sources the engines lean on. Lesson 4 of GEO 101 walks through every step, the settings to write down, the mistakes that spoil a check, and what MentionX scan data says about why one engine and one run aren't enough.

In this lesson

• What this check is for (and what it isn't)
• Step 1: pick 10 prompts
• Step 2: pick two engines
• Step 3: set up clean sessions
• Step 4: run and record
• From MentionX data: why two engines and two runs
• Worked example: Ledgerline's first hour
• Common mistakes
• Reading 40 answers honestly
• Exercise
• Quiz
• Sources

Anything marked "Our view" is our opinion, not documented engine behaviour. Ledgerline is the fictional company used throughout this course.

What this check is for (and what it isn't)

The manual check answers one question: when buyers ask AI engines about my category, roughly what happens? You come out of it with:

• A rough mention, recommendation and citation rate per engine
• The competitors named most often in your place
• The 5 to 10 domains engines cite most
• Any factual errors about you

It isn't a baseline. Forty answers can't tell a real 5-point change from noise. Lesson 6 builds the proper baseline. Our view: the manual check is worth doing even if you'll use a tool later. Reading 40 answers yourself teaches you how engines talk about your category, and no dashboard teaches that.

Step 1: pick 10 prompts

Use the questions buyers really ask. The best sources are:

• Sales call notes and recordings: what prospects asked before they found you
• Support tickets and onboarding questions: the problems people are trying to solve
• Reviews of you and competitors: how customers describe the need in their own words
• Forums and communities: how people phrase questions when no vendor is listening

Write them as full questions, the way someone types into a chat box, not as keywords. "best AP automation" is a keyword. "What's the best AP automation software for a 300-person company on NetSuite?" is a prompt.

Mix the intents:

Leave branded prompts ("is Ledgerline any good?") out of this check. They almost always mention you, so they'll inflate your rate without telling you whether buyers find you. They come back in Lesson 6, scored separately for accuracy.

Step 2: pick two engines

Pick the two your buyers are most likely to use. For most B2B teams that's ChatGPT and one Google surface (AI Mode or AI Overviews). Perplexity is a good third if your audience is technical.

Google treats AI Overviews and AI Mode as separate features, and they can give different answers and links for the same query (Google Search Central). Treat them as two engines, not one.

In ChatGPT, note whether the answer used web search. Answers built from search show sources (OpenAI Help). Answers built only from the model's training don't. Both are worth seeing, but don't mix them up in your sheet.

Step 3: set up clean sessions

Answers can change with history, memory, location and account type. You can't remove every variable, but you can keep them fixed and write them down.

1. Use a fresh browser profile or a private window, with no extensions that alter pages.
2. Log out where the product allows it, or use a separate account with memory and personalisation turned off. Never test from your company account. Engines may have learned what you work on.
3. Fix your location and language to your target market. If you sell in the UK, don't check from a German connection without noting it.
4. Record the mode: search on or off, model name if shown, AI Overview or AI Mode.
5. Write the settings at the top of the sheet before you start. You'll forget otherwise.

Step 4: run and record

For each prompt, engine and run:

1. Start a new chat. Earlier context carries over within a conversation.
2. Paste the prompt exactly as written. Don't fix typos halfway through.
3. Save the full answer: paste the text into the sheet or a linked doc, or save a screenshot named after the row (P03-chatgpt-r2.png).
4. Score it using the Lesson 3 rules: mentioned, position, recommended, own cited, other domains cited, accurate.
5. Don't answer follow-up questions from the engine. Score the first response.

Minimum sheet columns:

The competitors column is easy to skip and expensive to miss. It's how you find out who's taking your place.

From MentionX data: why two engines and two runs

Method. We ran read-only aggregate queries over completed MentionX runs from 22 August to 29 October 2026 (UTC): 2,154 runs across 18 projects, mostly B2B software. Engines were called through their APIs with web search or grounding on, which can differ from the consumer apps. Brands are anonymised. Perplexity, Copilot and DeepSeek runs come from a single day (26 October), so treat those rows as a snapshot.

Engines disagree. The share of runs in which the tracked brand was mentioned:

The same projects saw mention rates from 24% to 47%, depending on the engine. The engines also differ in how much they cite: Perplexity almost always showed sources, while DeepSeek through the API almost never did.

Runs disagree too. In a separate cut, we tracked 28 prompt-and-engine pairs that were repeated 643 times between 4 September and 19 October 2026. In 11 of the 28 pairs, the brand appeared in some runs and not in others. In 19 of the 28, the number of tracked brands named changed between runs.

How to read it. One engine can't tell you what another one says about you, and one run can't tell you whether a mention is stable. Two engines and two runs is the smallest check that shows both kinds of variation.

Caveats. The data covers 18 projects, the engine mix is uneven, and API answers aren't identical to the apps your buyers use. Your numbers will differ. The data shows that the variation exists, not how large it will be for you.

Worked example: Ledgerline's first hour

Illustrative example. Ledgerline, its competitors and every number below are invented.

Priya, Ledgerline's marketing lead, took 10 prompts from the last quarter's sales call notes: four best-of, three problem and three alternatives. She ran them in ChatGPT (search on) and Google AI Mode from a fresh profile with a UK connection, twice each, on one afternoon.

What she noticed:

• Problem prompts were a blank. Ledgerline wasn't mentioned in any of the 12 answers about duplicate payments or approval delays. Tool A was named in 9.
• AI Mode cited Ledgerline's site more but named it less. Two of the three own-site citations came from a blog post on approval workflows used as a general source.
• The old price came from one review page. Both wrong answers cited the same 2024 review site page.
• Run 1 and run 2 disagreed on whether Ledgerline appeared for 3 of the 10 ChatGPT prompts.

Her three notes for the team were these. Problem prompts are the biggest gap. The approval-workflow post is being cited, but it isn't linked to the product. One review page is spreading an old price. None of these is a measured rate yet. All three are worth checking in the Lesson 6 baseline.

Common mistakes

• Testing from your work account. Memory and history may tilt answers toward what you work on.
• Rewording prompts between runs. Then you're testing two prompts once each, not one prompt twice.
• One long chat. Context from earlier prompts leaks into later answers.
• Recording only your own citations. The other cited domains are the most useful output of the check.
• Stopping after one run. You can't tell a stable mention from luck.
• Choosing prompts you know you win. It's comforting, and it's useless.
• Not saving the answer text. When something changes next month, you'll want to see what it said.

Reading 40 answers honestly

Look for big, repeated patterns, not percentages:

• Absent from a whole intent group (every problem prompt, every comparison)
• The same competitor named first in most answers
• The same domain cited again and again
• The same wrong fact in more than one answer

A difference of one or two answers out of 20 is noise at this size. Our view: write findings as "we appear to..." rather than "our mention rate is 45%". At 40 answers, the honest wording keeps people from treating a sketch as a measurement.

Save the sheet as Manual check v1. Your prompt list becomes the seed for Lesson 5, and the patterns become things to test in the baseline.

Exercise

1. Pick 10 prompts using the mix above, with at least five that include a buyer constraint.
2. Run them in two engines, twice each, in clean sessions. Write your settings first.
3. Score all 40 answers. Score run 2 without looking at run 1.
4. Write down your rough rates per engine, the top 5 cited domains, the competitor named most often, and up to three "we appear to..." notes.

Quiz

1. How many answers do 10 prompts × 2 engines × 2 runs produce? 40.

2. Why leave branded prompts out of this check? They almost always mention you, so they inflate the rate without showing whether buyers find you.

3. Why start a new chat for every prompt? Earlier context in the conversation can change later answers.

4. You're present in Perplexity but absent from ChatGPT search. What's a sensible first check? Whether robots.txt or bot protection blocks OAI-SearchBot.

5. In MentionX data, how many of 28 repeated prompt-and-engine pairs had the brand appear in some runs but not others? 11.

Next: Lesson 5: Building your first prompt set.

Sources

• Google Search Central, AI features and your website
• OpenAI Help Center, ChatGPT search
• OpenAI, Overview of OpenAI crawlers
• MentionX aggregate scan data, 22 August to 29 October 2026 (UTC), 2,154 engine runs, 18 projects, plus a variance cut of 643 runs from 4 September to 19 October 2026. Anonymised and read-only.