The buying questions worth tracking in AI answers

Someone in your company just sent you a screenshot. They asked ChatGPT to recommend software in your category, and your brand was not listed. Three of your closest competitors were.

Your immediate urge is to log into an SEO keyword tool, type in your core product category, and try to optimize your website for that phrase. That impulse will waste three months of work.

AI assistants do not process queries the way traditional search engines process search terms. Tracking how your brand appears across ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews requires a completely different approach to building and maintaining a question set.

Keywords belong in search engines. Questions belong in assistants.

When buyers use traditional search engines, they type shorthand. They query "best B2B billing software" or "enterprise subscription management." They do this because search engines trained them to strip away grammar and context.

When buyers query an AI assistant, they write full paragraphs. They explain their existing technology stack, their team size, their budget limits, and their technical constraints. A real buyer does not ask ChatGPT for "best CRM." They ask a multi-layered question: "We are a 40-person logistics company using PostgreSQL and HubSpot. We need a dispatch tool that handles custom API integrations without requiring a full-time engineer. What options should we evaluate?"

The prompt contains constraints, an environment description, and explicit capability requirements. If you build a tracking system around three-word head terms, you are measuring queries that actual buyers rarely ask an assistant.

AI recommendation engines evaluate whether your brand fits a specific context. They weigh trade-offs. If a prompt mentions strict data privacy requirements, the engine might pick a niche vendor that offers self-hosting over a market leader that only operates in a public cloud.

How to mine questions from actual buyer conversations

To build a prompt basket that reflects real buyer behavior, throw out traditional keyword discovery tools. Do not buy AI visibility software that auto-generates 500 prompt variations at the click of a button. Those tools produce generic questions that dilute your evaluation data with statistical noise.

Instead, gather your prompt basket directly from three sources inside your company.

First, review sales call transcripts from tools like Gong or Chorus. Look for the exact phrases prospects use when describing their problems during initial discovery calls. Pay special attention to how they describe their current tech stack, their operating constraints, and their mandatory features.

Second, read lost deal notes in your CRM. Identify the competitors you repeatedly lose deals to, along with the specific reason recorded by the sales rep. Prompts that directly compare your product against these specific rivals will be your most valuable tracking tools.

Third, look at your support tickets and onboarding logs. Find the technical questions prospect evaluation teams ask during proof-of-concept trials.

A strong prompt basket contains 15 to 30 highly specific prompts divided into three clear categories:

1. Contextual evaluation prompts: "What software should a 100-person healthcare company use for HIPAA-compliant internal messaging?"
2. Direct comparative prompts: "How does Vendor A compare to Vendor B for high-volume invoice processing?"
3. Alternative-seeking prompts: "What are the best open-source alternatives to Vendor C for a team that needs on-premise deployment?"

Why you must freeze your prompt basket

Once you assemble your basket of 20 or 30 questions, you must freeze it. Do not change the wording. Do not add five new questions next week because an executive had an idea in a meeting.

If you change your prompt basket every month, you are measuring changes in your questions, not changes in engine behavior. To track whether your brand visibility is improving or declining over time, your baseline input must remain identical.

Freezing the basket feels uncomfortable because it feels static. You will be tempted to refine a prompt when you notice an assistant misinterpreting a product name. Resist that urge. If you modify a prompt six weeks into tracking, you wipe out your historical comparison. Changing your prompt basket resets your measurement clock to zero.

This baseline stability matters because AI assistants are non-deterministic systems. If you ask ChatGPT the exact same question five minutes apart, it can return two different answers. It might pull a live web search on the first run and rely entirely on stored model weights on the second.

To handle this variance, Standing runs every question in your basket five times per question per engine across all five tracked platforms: ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews. We then calculate a score using a Wilson confidence interval. A measurement is never a single integer. It is a range, such as 34, plus or minus 6.

Competitors like Profound or Peec AI take a different approach. They run each prompt once through an API and display a single fixed score on a dashboard. That single run creates an illusion of certainty. If an assistant happened to cite your website on that single run, the competitor tool marks you as recommended. If the assistant omits you on the next run five seconds later, the tool misses it entirely. Running a prompt once is cheaper to compute, which is why competitors do it, but single-pass execution produces data that fluctuates wildly based on random generation noise.

What assistants actually read when answering

When a user asks a buying question, engines use two distinct methods to form an answer: internal model weights built during training, and live web search performed at the moment of the query.

Understanding this distinction requires looking at how AI companies crawl the web. Vendors run two entirely different categories of web crawlers, and confusing them is a critical mistake.

Training crawlers gather data to build future model weights. Examples include GPTBot, ClaudeBot, and Google-Extended. Blocking these crawlers in your robots.txt file prevents an AI company from using your content to train their next model update months or years from now. It has zero impact on whether an assistant cites your site during a live search today.

Answer crawlers perform live retrieval during an active user session. Examples include ChatGPT-User, Perplexity-User, Claude-SearchBot, OAI-SearchBot, Claude-User, and PerplexityBot. When an assistant searches the web to answer a comparison question, these are the agents fetching web pages. Blocking an answer crawler stops an assistant from reading your site right now.

In our crawler index update from August 8, 2026, we analyzed a panel of 54,082 domains, of which 33,670 returned a readable robots.txt file. Within those readable domains, 5,497 block at least one AI agent. More importantly, 2,576 block at least one answer crawler.

Looking at individual agents out of those 33,670 readable domains:

• GPTBot (training) is blocked by 5,080 domains.
• ClaudeBot (training) is blocked by 4,603 domains.
• Google-Extended (training) is blocked by 4,275 domains.
• PerplexityBot (answer crawler) is blocked by 2,147 domains.
• ChatGPT-User (answer crawler) is blocked by 2,074 domains.
• OAI-SearchBot (answer crawler) is blocked by 1,579 domains.
• Perplexity-User (answer crawler) is blocked by 1,402 domains.
• Claude-SearchBot (answer crawler) is blocked by 1,400 domains.
• Claude-User (answer crawler) is blocked by 1,384 domains.
• Googlebot is blocked by 610 domains.

In a separate census of 4,226 active Y Combinator companies conducted on August 7, 2026, 3,755 had readable robots.txt files, and 253 blocked at least one AI crawler.

Many companies accidentally block answer crawlers while trying to block training crawlers, or use security configurations that block automated fetchers entirely. If an assistant attempts to perform real-time research to answer a buyer prompt, but your server blocks ChatGPT-User or Perplexity-User, the assistant cannot read your pricing or documentation pages. It will cite a competitor whose site remains open to answer crawlers.

Do not spend money on agency services that promise to boost your AI citations through schema markup or llms.txt files. We have no evidence that adding schema or custom text files increases recommendation rates in large language models. Focus instead on ensuring answer crawlers can access your site, and build a stable basket of real buyer questions to measure how models evaluate you over time.