GEO Lab #03 · S3-CHATGPT-M1-R2 · Original response

This page presents one selected response from GEO Lab #03. It adds a smaller reading and citation entry point; it does not replace the full archive or collect a new answer.

The OpenAI arm uses the Responses API with web search and does not represent consumer ChatGPT Search. Grok records identify their actual consumer-interface mode and available model label.

Recorded conditions

Record ID
S3-CHATGPT-M1-R2
Platform
OpenAI API + web search
Question / condition
M1
Repeat
2
Model label
gpt-5.6-sol
Visible mode
Not independently recorded
Reasoning
medium
Submitted / first recorded time
2026-11-02T09:26:37.188796+00:00
Completeness
Recorded body complete
Search
Actual search activity observed; exact controls and evidence remain in the raw record.

Original prompt

I'm a solo SaaS founder with no GEO budget. How can I reliably measure whether ChatGPT and Grok recommend my product to potential buyers?

The agreed prompt is reproduced exactly. Any recorded difference in the actual submitted text or search UI is retained in the raw record and full appendix.

Original answer

You need to measure **two different things**:

1. **AI visibility:** Does your product appear in relevant ChatGPT/Grok answers?
2. **Business impact:** Are those answers producing visits, signups, or sales?

You can measure both manually for free, but don’t treat a few prompts from your own account as representative. Both services can personalize answers, and both may search the web depending on the query. ChatGPT can use memories, location, and rewritten search queries; Grok can use X activity and interests for personalization. ([help.openai.com](https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt))

## 1. Build a fixed “buyer prompt panel”

Create 15–30 prompts reflecting how buyers actually describe their problem—not prompts containing your brand.

Cover four categories:

| Intent | Example |
|---|---|
| Category | “What are the best invoicing tools for solo consultants?” |
| Problem | “How can I automate overdue invoice follow-ups?” |
| Requirements | “Recommend an invoicing SaaS under $20/month with Stripe integration.” |
| Alternatives | “What are simpler alternatives to QuickBooks for freelancers?” |

Include buyer context such as company size, geography, budget, required integrations, and technical level.

Your primary metric should be **unaided recommendation**: whether your product appears when you did not mention it.

Separately test **aided knowledge** with prompts such as:

- “What is `[Product]`?”
- “Compare `[Product]` with `[Competitor]`.”
- “Who should and shouldn’t use `[Product]`?”

Aided tests diagnose what the model knows; they do not measure discovery.

## 2. Run a controlled audit

For each prompt:

- Start a fresh conversation.
- Use identical wording on both platforms.
- Don’t correct or guide the answer.
- Save the complete response and cited sources.
- Repeat each prompt three times on different days.
- Record date, model/mode, whether web search occurred, location, and personalization state.

For ChatGPT, use **Temporary Chat with personalization disabled** where available. Also test two separate tracks:

- **Default behavior**
- **Explicit web search:** append “Search the web and cite current sources.”

This distinction matters because answers without search rely more heavily on learned model knowledge, while searched answers draw from current web sources. OpenAI also cautions that citations and search results can be incomplete or incorrect, so inspect the actual cited pages. ([help.openai.com](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt))

For Grok, disable personalization where possible or use a clean testing account. Grok may search both public X posts and the broader web, and its answers can be personalized using account activity and interests. ([help.x.com](https://help.x.com/en/using-x/about-grok))

## 3. Score every answer consistently

Use a simple 0–3 score:

- **0:** Not mentioned
- **1:** Mentioned, but not recommended
- **2:** Positively recommended or shortlisted
- **3:** Top recommendation

Add separate columns for:

- Position in the list
- Description accurate? Yes/no
- Correct target customer? Yes/no
- Website linked or cited? Yes/no
- Third-party source cited?
- Competitors recommended
- Any serious factual error

Then calculate:

```text
Mention rate = Runs mentioning you / Total runs

Recommendation rate = Runs scoring 2 or 3 / Total runs

Top-pick rate = Runs scoring 3 / Total runs

Citation rate = Runs citing your site / Searched runs

Consistency rate = Prompts where you appear in at least 2 of 3 runs / Total prompts
```

Also calculate **competitor share of voice**:

```text
Your recommendation appearances /
All recommendation appearances across you and tracked competitors
```

The consistency metric is especially important. Appearing once in nine runs is materially different from appearing eight times, even if someone posts the one successful screenshot.

## 4. Use this spreadsheet structure

One row per run:

```text
Date
Platform
Model/mode
Web search used?
Personalization off?
Prompt ID
Exact prompt
Your score: 0–3
Your position
Accurate description?
Your site cited?
Other source cited?
Competitors named
Answer URL/screenshot
Notes
```

Maintain a separate prompt sheet with:

```text
Prompt ID
Intent category
Buyer persona
Buying stage
Business importance: 1–5
```

Weight your overall score by business importance. A recommendation for a high-converting “alternative to X” query matters more than one for a broad informational query.

## 5. Track actual AI-generated pipeline

Visibility testing tells you what *can* happen. Attribution tells you whether buyers are experiencing it.

### In analytics and server logs

Look for referral traffic from domains associated with ChatGPT and Grok. However, treat referrer data as a lower bound: apps, redirects, and privacy controls may prevent the original source from being passed.

Track:

- AI-referred sessions
- Signup rate
- Trial-to-paid rate
- Revenue
- Landing pages receiving AI traffic

### On signup

Add a required or optional question:

> How did you first hear about us?

Options:

- ChatGPT
- Grok
- Another AI assistant
- Google/search engine
- Social media
- Recommendation
- Other

If they select an AI assistant, ask:

> What were you asking it?

That answer is often more valuable than analytics because it gives you the real buyer prompt.

### During customer conversations

Ask:

> What other products did the AI recommend, and why did you choose us?

This reveals whether you were merely mentioned or actually framed as the best fit.

## 6. Use a sustainable solo-founder cadence

A reasonable free process:

### Weekly — 20 minutes

- Test your five highest-intent prompts once on each platform.
- Review new AI-attributed signups.
- Investigate any incorrect descriptions.

### Monthly — 60–90 minutes

- Run the full 15–30 prompt panel.
- Repeat important prompts three times.
- Calculate mention, recommendation, top-pick, citation, and consistency rates.
- Compare against the previous month.

Avoid testing every day and reacting to individual changes. Models, search results, source selection, and wording vary; evaluate rolling monthly trends instead.

## Your minimal dashboard

Track only these six numbers:

1. High-intent recommendation rate
2. Top-three placement rate
3. Consistency rate
4. Share of voice versus three competitors
5. AI-attributed signups
6. AI-attributed revenue

Your most defensible statement will then be something like:

> “Across 20 fixed buyer prompts and 120 controlled runs this month, we were recommended in 18% of ChatGPT answers and 11% of Grok answers. Five signups and one paid conversion self-attributed to those assistants.”

That is far more meaningful than claiming you “rank in ChatGPT.” There is no single universal ranking—only recommendation rates under documented prompts, conditions, and dates.

Recorded sources

These are the source roles recorded in the public dataset. A retrieval or source-panel item is not automatically an answer-body citation. A URL supplied in the prompt and repeated by an answer does not prove content access. No destination has been repaired or independently validated here.

Recorded answer-body links (3)

Retrieval-only / source-panel inventory (15)

Evidence and qualifications

This is the existing redacted public record. It has not been rewritten for this page. Earlier privacy edits and collection qualifications continue to apply. Editorial text is English; a few original captures retain non-English UI badges or source-panel date labels. An answer’s empirical claims are platform output, not verified findings of this archive.

Body review

{
"reviewed_at": "2026-11-02T09:27:44.309263+00:00",
"reviewer": "assistant",
"complete_body_read": true,
"coherent_ending": true,
"annotation_and_explicit_url_coverage_checked": true,
"external_claims_verified": false,
"independent_human_double_coding": false
}

Existing public-text privacy policy

{
"body_changed": true,
"policy": "Only actual URL tracking, signed/session URL parameters and necessary identifying information are removed; original wording, including analytics parameter examples outside URLs, remains.",
"annotation_positions": "Original position indexes omitted because URL/privacy edits can change character locations."
}

Prompt SHA-256: 709fd6fc82fe3f87b623fb3dd8b2adecdd6655c3508115f8b40c573d2397776d
Answer SHA-256: 39473f30f5802b57323c30457f5d41e204da355d2672925cfc201cbb24e1049d