Prompt Set Template: Categories, Fields and Labelled Examples

A prompt set is only useful if each answer is logged with enough context to repeat it. The prompt text alone is not enough. The same question can get a different answer depending on which engine answered, whether it searched the web, and when you asked.

Copy these columns into a spreadsheet. One row is one prompt run on one engine on one date.

The retrieval state column matters most. An answer with no web search comes from what the model learned before its training cutoff. An answer with search can draw on your current pages. They fail for different reasons, and they need different fixes.

All examples below are hypothetical. Say you run Ledgerly, a bookkeeping app for freelancers, and a sister business, a dental clinic in Austin.

Category prompts

The buyer names the type of product, not the brand. These show whether you are in the answer at all.

• "What is the best bookkeeping app for freelancers?"
• "Which accounting tools work well for a one-person consultancy?"

Comparison prompts

The buyer weighs named options. Include your brand against the competitors you lose deals to.

• "Ledgerly vs the leading bookkeeping app for freelancers: which is better for invoicing?"
• "What are alternatives to Ledgerly?"

Problem prompts

The buyer describes a pain and does not know the category yet.

• "How do I track expenses as a freelancer without a spreadsheet?"
• "How can I stop missing quarterly tax payments?"

Branded fact prompts

The buyer asks about you directly. These test accuracy, not presence.

• "How much does Ledgerly cost?"
• "Does Ledgerly integrate with Stripe?"
• "Who owns Ledgerly and where is it based?"

Local prompts

The buyer adds a place. Skip this category if you do not serve a location.

• "Best dentist in Austin for a same-day appointment"
• "Which Austin dental clinics take walk-ins on Saturdays?"

Category, comparison and problem prompts are unbranded. Branded fact prompts are branded. Report the two groups separately, as covered in splitting branded and unbranded prompts. A blended number hides whether you have a discovery problem or an accuracy problem.

1. Write your fact sheet first. List your price, plans, integrations, founding year, location and category in one place. Every branded answer gets checked against it. The brand fact sheet template covers the fields.
2. Draft prompts in buyer language. Pull wording from sales calls, support tickets and search console queries. Write how buyers ask, not how your marketing team talks.
3. Spread prompts across all five categories. If you serve more than one market, build a separate set per country and language, as in prompt sets for multiple countries.
4. Assign IDs and freeze the wording. Label the set v1. Any later wording change becomes v2.
5. Set up a clean session. As of October 2026, OpenAI says ChatGPT may use saved memories when it rewrites a search query. Log out or turn memory off so results do not reflect your own history.
6. Run each prompt on each engine separately. Google says AI Overviews and AI Mode may use different models and techniques. Treat them as two rows, not one.
7. Record the retrieval state honestly. If the answer shows citations, mark it on. If it shows none, mark unknown unless you switched search off yourself.
8. Copy facts word for word and grade them. Mark each claim correct, wrong or unverifiable against your fact sheet.
9. Save cited URLs. They show which pages, yours or third-party, the engine relies on.

If you want a quick first pass before building the full sheet, the free checker on this site runs up to three buyer questions and scores the result. It is a snapshot, so log it as one dated row, not a trend. For the full method of choosing prompts, see how to build a prompt set.

Run through this list before each round.

• Prompt set version noted, and wording unchanged since the last run
• Fact sheet current, including any price or plan change since last month
• Session logged out or memory off, with the same country and language as last time
• Each engine and surface logged as its own row
• Model or plan recorded where the app shows it
• Retrieval state recorded for every answer: on, off or unknown
• Run date filled in on every row
• Branded and unbranded prompts tagged so they can be reported apart
• Facts copied verbatim, not summarised
• Cited URLs saved in full
• Any engine that errored or refused marked as no answer, not as a clean result

The last item matters. A refusal or an error is missing data. Counting it as a correct answer makes your accuracy look better than it is.

Good results

• Branded fact prompts return your correct price, category and integrations.
• Unbranded prompts name you, and the cited URLs include your own pages or third-party pages that describe you correctly.
• Results are stable across runs with the same retrieval state.

Bad results, and the next move

• A wrong fact with search on. Find the cited page that states it. If it is your page, fix the page. If it is a third-party page, ask that site to correct it.
• A wrong fact with search off. This reflects training data from before the model's cutoff. Nothing you publish this month reaches it. Make sure the correct fact is stated plainly on a page crawlers can fetch, so answers with search draw on it.
• The wrong category. See what to do when AI puts you in the wrong category.
• Never named, on any engine, with search on. Check access before content. Confirm your robots.txt allows the search crawlers (OAI-SearchBot for ChatGPT search, PerplexityBot, Claude-SearchBot), that your firewall does not block them, and that key facts are in the HTML before JavaScript runs.
• Competitors named via the same third-party pages. Those pages are the route into the answer. Being listed on them is associated with being named, though nothing guarantees it.

Bad prompt sets

A set is broken if branded wording leaks into unbranded prompts, if wording drifts between runs, or if retrieval state and date are missing. You cannot tell a real change from noise in any of those cases.

Run the full set monthly, on roughly the same day, with the same session setup. Engines change models and defaults often, so a monthly rhythm catches drift without turning the log into a full-time job.

Add an extra run in three situations:

• After you fix something. Rerun the affected branded fact prompts once the corrected page is live and crawlable.
• After a model change you can see. When an app changes its default model, note it in the model column and rerun.
• Before a launch or price change. Capture a baseline so you can see how long the old fact lingers.

Revise prompt wording at most once a quarter, and bump the version number when you do. Comparisons across versions are not like for like. How large the set should be depends on your categories and markets, covered in how many prompts you need.

A prompt set measures answers to questions you chose. Engines do not search with those questions. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries. Google's Gemini API documentation says the model generates and runs its own search queries. Your prompt is an input to that process, not the query itself.

Three more limits to keep in mind:

• Retrieval is discretionary. Engines decide per query whether to search. Google AI Overviews do not appear on every query. Any average across rows mixes answers with and without search.
• Cited is not the same as used. An engine can draw on pages it does not cite.
• No rank exists to track. Google's own AI reporting gives impressions only. Log whether you were named, not a position.

The closest look at real queries comes from Bing Webmaster Tools, which shows grounding queries for Copilot. Comparing those with your prompts shows where your set misses, as explained in comparing grounding queries with your prompt set.

Should a prompt set include competitor names? Yes, in comparison prompts only. Name the competitors you actually lose deals to, and keep the wording neutral. Do not add competitor names to category or problem prompts, since that changes the question. Record which competitors each answer names, but avoid turning that into a rank. AI answers have no stable position to measure, and the same prompt can list brands in a different order on the next run.

How do I know if an answer used web search? Look for citations or source links in the answer. If they appear, mark retrieval on. If none appear in a consumer app, mark it unknown, because the engine may have searched without showing sources or answered from training data. Only mark retrieval off when you control it, for example when you run the prompt through an API with web search switched off.

Can I reuse the same template for every engine? Yes. Keep one template and add one row per engine per prompt. Do not merge engines into one score. As of October 2026, Google says AI Overviews and AI Mode may use different models and return different links, and other engines cite sources in different ways. Separate rows let you see which engine gets a fact wrong, which tells you where to look for the cause.

Do the example prompts need to be real buyer questions? They should be as close as you can get. Pull wording from sales calls, support tickets and search queries rather than inventing it. Even then, treat the set as a proxy. Engines rewrite prompts into their own search queries, so real buyer wording narrows the gap but does not close it. Bing Webmaster Tools grounding queries are a useful check on how close you are.