Should I write my own prompts or let a tool generate them?

Start with prompts you write from your buyers' own words. Use a tool to add coverage you missed. Then test the whole set against real grounding queries where you can see them.

The reasoning is simple. A prompt set is a sample of questions you ask AI engines to see whether they name you. The engines don't answer those questions directly. They turn them into their own search queries first. So both options are guesses at an input you can't observe.

The question is which guess is closer. A prompt in the words a real buyer used last week is closer than a template like "best [category] tools for [industry]". A generator fed only your homepage will mostly echo your own marketing language back to you. Your buyers rarely talk that way.

That doesn't make tools useless. They are fast, they catch question types you forgot, and they help when you're entering a market you don't know yet. Treat their output as a draft to edit, not a finished set.

The gap between what a user types and what the engine searches is documented by the engines themselves, as of October 2026.

• OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries that it sends to its search providers. It also says saved memories may shape that rewrite when memory is on.
• Google says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics.
• Google's Gemini API documentation says the model automatically generates one or multiple search queries and runs them.

Retrieval is also discretionary. ChatGPT decides per question whether to search at all, and AI Overviews often don't trigger. So a tracker can't see the real queries, and it often can't see whether search ran. For more on this, read whether AI searches for the exact question a user types.

What this means for you: no vendor can sell you the real prompts, and no hand-written set is the real prompts either. Your job is to make the proxy as close to your market as possible, then measure it consistently.

Your best prompts already exist in your own records. Pull them from places where buyers speak in their own words:

• Sales call notes and recorded discovery calls
• Support tickets and chat logs
• Questions asked in demos or on pricing calls
• Review sites and community threads where people compare options
• Internal site search logs

Say you run a dental clinic in Austin. A template might give you "best dentist in Austin". Your front desk hears "who does same-day crowns near Mueller and takes Delta Dental". The second question is the one a real patient asks, and it tests whether AI knows your specific facts.

Tool-generated prompts earn their place in three cases. You have no buyer records yet. You need variants across countries or languages. Or you want a check for blind spots in a hand-built list. Our prompt research method walks through combining both sources.

One caution: don't rewrite your pages to match every prompt variant. Google's guide on generative AI features says you don't need to worry about capturing every long-tail variation. Prompts are for measuring, not for keyword stuffing.

Bing Webmaster Tools has an AI Performance report covering Microsoft Copilot, Bing's AI summaries and some partner integrations. It shows grounding queries: the phrases the AI actually used when it retrieved content it later cited. No other engine exposes this.

Three limits to know:

• Microsoft says the grounding query data is a sample of overall citation activity.
• It covers Microsoft surfaces only, not ChatGPT or Google.
• As of our evidence review, there was no API for it, so you export it by hand. See whether Bing's AI Performance report has an API.

Even so, it's the best calibration you have. Export your grounding queries and compare them with your prompt set. Where they overlap, your set reflects real retrieval. Where Bing shows queries your set lacks, add prompts. The grounding queries vs prompt set playbook covers the comparison step by step.

The gap can be large. In Otterly's 2026 experiment, Microsoft's own report credited one site with 42,817 citations, while Otterly's prompt-based tracker logged 3,934. The time windows may not match, so it isn't an exact ratio. It does show that any prompt sample sees less than the engine's own count.

1. Collect 20 to 40 real buyer questions from your own records before you open any tool.
2. Rewrite each one only enough to stand alone as a question. Keep the buyer's wording.
3. Use a generator to suggest gaps, then delete anything that sounds like your marketing copy.
4. Cover different stages: discovery, comparison, pricing, and problems your product solves.
5. Pull grounding queries from Bing Webmaster Tools and add the patterns you missed.
6. Freeze the set. Change it on a schedule, not every week, so trends mean something.

For sizing, see how many prompts you need to track AI visibility. For structure, use how to build a prompt set.

If you want a quick first read, the free checker on this site runs up to three buyer questions. You can type your own or let it write them from your homepage. It's a snapshot, not tracking, but it shows whether your own wording and generated wording give different results.

Can any tool see the prompts people actually type into ChatGPT? No. No API exposes real user prompts or whether your page was retrieved for a real user's question on any consumer AI app. Tools that claim to show "real prompts" are showing their own estimates or panels. The closest first-party data is Bing Webmaster Tools' grounding queries, which show queries Microsoft's AI ran, not what users typed, and only as a sample.

Does it matter if my prompts are worded slightly differently each time? Yes, for measurement. Engines rewrite prompts and decide per question whether to search, so small wording changes can shift what gets retrieved. If you edit prompts often, you can't tell whether a change in results came from the engine or from you. Keep a fixed core set and add new prompts as a separate group.

Should I edit my website to match the prompts in my set? Not word for word. Google's guide on generative AI features says you don't need to write in a special way for AI or cover every long-tail variation. The original GEO paper found keyword stuffing lowered its own visibility metric. Use prompts to find what AI gets wrong or misses, then state those facts clearly on pages crawlers can fetch.

Are tool-generated prompts biased? Often toward your own language. A generator that reads your homepage tends to produce questions using your product terms and category framing. Buyers who don't know you yet rarely use those words. That can make results look better or worse than reality. Mixing in buyer phrasing and checking against grounding queries reduces the bias.