Using ChatGPT to Audit Your Own GEO

You can use ChatGPT to audit parts of your own GEO: ask unbranded buyer questions to see if you're named, ask branded questions to check accuracy, and paste your pages in for a critique of clarity and missing answers. Do it in a clean session, with search on, running each prompt more than once. It's a quick, free check of one engine, not a replacement for a multi-engine baseline.

Key takeaways

• Start unbranded; mentioning your brand first biases the answer.
• Check every fact about you against your own source of truth.
• Treat page critiques as suggestions to test, not rules.

Step 1: How do you set up the session?

• New chat, logged out or a fresh account, memory off if available.
• Use search for questions that need current information, and note when it's used.
• Keep a sheet open to record answers and cited URLs.

Step 2: Which unbranded prompts show whether you're found?

Use your buyers' language, without your name:

• "What are the best [category] tools for [audience/constraint]?"
• "What's a cheaper alternative to [competitor]?"
• "[Competitor A] vs [Competitor B] for [use case]: which should I choose?"
• "Which [category] tools integrate with [platform]?"

Record: are you named, in what position, recommended, and which sources are cited. Run each twice.

Step 3: Which branded prompts check accuracy?

• "What is [brand] and who is it for?"
• "How much does [brand] cost?"
• "What are the pros and cons of [brand]?"
• "Does [brand] support [feature/market]?"

Compare each claim with your source of truth. Log errors with the cited source, if any.

Example row only.

Step 4: How do you use ChatGPT to critique a page?

Paste the page text and ask:

• "List the questions a buyer would have that this page doesn't answer."
• "Which sentences are vague or unverifiable? Rewrite them with placeholders for real facts."
• "Summarise this page in two sentences. Is the main answer clear?"

If the summary misses your main point, the page probably buries it. Don't let it invent facts or statistics for you; insert real data yourself.

Step 5: Can it help analyse your scan data?

Yes. Paste or upload your scan sheet and ask for mention rates by intent, the most-cited domains, and patterns. Check the arithmetic. Language models can make calculation mistakes, so verify key numbers with a formula.

What are the limits?

• One engine. Perplexity, Gemini and Google's AI features can answer differently (Google).
• Variation. Answers change between runs.
• Opinions, not ranking factors. Its critique isn't how any engine actually selects sources.
• Access isn't tested. Check robots.txt separately; ChatGPT search relies on OAI-SearchBot (OpenAI).

What should you do next?

Turn the errors and gaps into a short action list, then run a proper baseline across engines (GEO 101, Lesson 6).

Checklist

• [ ] Clean session, search noted
• [ ] 10 unbranded prompts, run twice
• [ ] 5 branded accuracy prompts
• [ ] Errors logged with sources
• [ ] 3 key pages critiqued
• [ ] Robots.txt checked separately
• [ ] Next step: multi-engine baseline

Which prompts work best for a page critique?

Paste a page's text and ask focused questions rather than "is this good?". Examples you can reuse:

• "Summarise this page in two sentences. What question does it answer best?"
• "List the factual claims on this page. Which lack a source or a number?"
• "What questions would a buyer still have after reading this?"
• "Which sentences could be quoted on their own as a clear answer?"
• "Compare this page with [competitor URL text]. What does each cover that the other doesn't?"

Treat the output as a second opinion, not a verdict. The model doesn't know your buyers or your product better than you do.

How do you check brand accuracy systematically?

Build a short fact sheet first: pricing, plans, key features, founding facts, locations and who the product is for, each with its official source URL. Then ask a branded question per fact and compare the answer line by line.

Leave the cells blank in your own template; fill them in at each check. When answers are wrong, ask "what sources did you use?" in a search-enabled session and follow the links.

How do you avoid biased results?

• Use a fresh, logged-out or temporary session; memory and past chats can shape answers.
• Don't mention your brand in unbranded prompts.
• Run each prompt at least twice; answers vary between runs.
• Note whether search was used, because answers with and without web results can differ.
• Ask the same questions in at least one other engine before drawing conclusions.

What can you automate safely?

Using ChatGPT to clean and categorise a pasted table of results, draft a summary of findings, or suggest page headings is low risk because you can check the output. Using it to generate statistics, studies or quotes for your pages is high risk; models can produce plausible but false details. Verify every fact against a primary source before publishing.

How should you record the audit?

Save prompts, dates, session settings and full answers in a sheet with the same columns you use for other engines. Keep the fact sheet versioned so you can see when an error first appeared and when it was fixed. Repeat monthly, or after any major page or pricing change.

How do you turn findings into tasks?

Each finding should produce one task with an owner: "Update pricing on partner page X", "Add a comparison table to page Y", "Request correction from publisher Z". Put them in your normal project tool, link back to the logged answer and recheck the same prompt after the change has had time to be recrawled.

Sources

• OpenAI, Overview of OpenAI crawlers
• Google Search Central, AI features and your website