Prompt Research for AI Search Visibility

Classic keyword research is broken for AI search. Nobody types "best CRM small business 2026" into ChatGPT - they type "we're a 12-person agency, mostly project work, what CRM should we actually use". The unit of demand shifted from three-word keyphrases to full-sentence prompts, and the ranking surface shifted from ten blue links to a single generated paragraph with three to six citations. If you want to be one of those citations, you need a different research process.

Perves runs our AI-visibility desk at BGR Review. We track prompt inclusion across ChatGPT, Perplexity, Google AI Overviews (AIO), and Bing Copilot for clients in review management, local services, and SaaS. This is the operational playbook we use - the same one that lifted one client's citation share in AIO from 2 mentions per month to 47 in the last audit window.

What prompt research is (and how it differs from keyword research)

Keyword research asks: "what phrases get typed into Google, and how much volume do they have?" Prompt research asks four different questions.

1. What full-sentence questions do users pose to AI assistants? Prompts are 3.2x longer than keywords on average and carry explicit context ("we're a dental clinic", "in the UK", "under $50/month").
2. Which prompts trigger a generative answer at all? Only about 42% of informational queries trigger AIO in the US as of Q1 2026 - the rest still show classic SERPs.
3. Which sources does the model cite for each prompt? This is your competitive set, and it rarely matches the top-10 organic ranking.
4. What format does the cited paragraph take? Definition-first, list, comparison table, or step sequence - each prompt has a preferred format the model reuses.

The output of prompt research is not a keyword list. It is a prompt-to-answer map: for each target prompt, the current citation set, the preferred answer format, and the specific gap your page can fill.

Where to mine real prompts (not guessed ones)

The biggest mistake teams make is brainstorming prompts in a meeting room. Real user prompts have context and constraints that no product marketer writes naturally. Six sources give you real ones.

1. ChatGPT shared-conversation URLs. Every "share link" a user creates is indexed. Search Google for site:chat.openai.com/share "your topic" to see real conversations users had about your niche.
2. Perplexity Discover feed and public threads. Perplexity publishes millions of user threads at perplexity.ai/search - filter by category, sort by recent, extract the questions.
3. Reddit comment mining. Comments phrased as questions in relevant subreddits are the closest analogue to AI prompts - full context, real constraints. Use PullPush or Reddit's own search with "how do I", "should I", "what's the best".
4. Google's "People Also Ask" expansion. Click through 3 or 4 levels of PAA on your seed queries. The deeper questions match prompt phrasing more closely than the top level.
5. Your own support inbox and sales chat logs. Every question a prospect asks a human is a prompt they would ask an AI first if they could. Export 6 months of chat transcripts.
6. Purpose-built prompt-tracking tools. AlsoAsked, Otterly.AI, Peec AI, Profound, and Bluefish publish live prompt datasets scraped from AI surfaces.

Target volume for a first audit: 400 to 600 real prompts per topic cluster. Fewer than that and your clustering is noisy; more than that and you paralyse the writing pipeline.

Clustering prompts into answerable intents

Raw prompts are messy. "Best CRM for a small agency", "which CRM should our 8-person team use", and "recommend a CRM for freelancers who bill hourly" are the same intent from three angles. Cluster before you write.

The lightweight process: embed each prompt with a small OpenAI or Cohere model, run k-means with k=20 to 40, then label each cluster manually. Every cluster becomes a candidate page or a section within a page. Prompts that do not cluster (long tail) become FAQ entries on the closest cluster page.

The heavier process: build a two-tier hierarchy - top-level intent (compare, define, troubleshoot, choose, learn how) and topic (CRM, invoicing, project management). Each intent-topic cell is one page. This mirrors how the underlying models actually retrieve.

Auditing current citation share

You cannot optimise what you do not measure. For each priority prompt, capture the current AI-generated answer and its citations. Two ways to do it at scale.

1. Manual sampling. Run each prompt in ChatGPT (browsing on), Perplexity, and Google AI mode. Screenshot the answer and note the cited domains. Slow but honest. Budget 3 to 5 minutes per prompt.
2. Automated tracking. Tools like Peec AI, Otterly, Profound, and Athena HQ run daily prompt panels through the model APIs and log citations. Cost: $99 to $499/month for 200 to 2,000 prompts.

The metric that matters is share of citations: of the total citations across your tracked prompts, what percentage point to your domain? A healthy target for a mid-authority site (DR 40 to 60) is 8 to 15% citation share in its home category within 6 months of focused work.

What actually makes a page citable

We reverse-engineered 2,140 cited passages across AIO, Perplexity, and ChatGPT browsing over Q4 2025. Five patterns show up disproportionately.

1. A definitive first paragraph of 40 to 60 words. Not a hook. Not a story. A direct answer. AI models overwhelmingly cite the first substantive paragraph.
2. Original data or a unique number. "3.4x more citations" is more citable than "significantly more citations". Numbers from your own dataset beat rewrites of a competitor's.
3. Named entities. Products, versions, dates, jurisdictions, dollar amounts. Vague pages get skipped in favour of specific ones.
4. Short, self-contained sentences. Under 24 words each. The model quotes at the sentence level; long sentences with dependent clauses get truncated or dropped.
5. An H2 that restates the prompt. If a user asks "how do AI overviews choose sources", an H2 reading "How AI overviews choose sources" (not "The selection logic behind generative citations") gets pulled 2.7x more often.

The prompt-to-page mapping template

For every priority prompt, build one row with these seven fields. This is the artifact your writers work from.

1. Target prompt (verbatim).
2. Intent bucket. Compare, define, choose, troubleshoot, learn how.
3. Current citation set. Domains cited today, ranked by frequency across ChatGPT / Perplexity / AIO.
4. Preferred format. Paragraph, list, table, step sequence.
5. Assigned URL on your site. New or existing.
6. Gap statement. One sentence describing what the current citations miss and what your answer will add.
7. Unique proof. The specific number, quote, dataset, or example your page will contribute.

Ship the sheet before you write. If you cannot fill the gap and unique-proof columns, you are not going to earn the citation - either the topic is not for you, or you need to run the underlying research first.

Publishing cadence and internal linking

AI models weight recency for time-sensitive topics (news, product versions, policy changes) and authority for evergreen topics (definitions, how-to). A hybrid cadence works best.

For evergreen clusters, ship one deep pillar page (2,000 to 3,500 words) plus 8 to 12 supporting pages (700 to 1,200 words) over 60 days, then refresh every 90 days. For time-sensitive clusters, ship a shorter, faster cadence (400 to 800 word updates every 2 to 3 weeks) so the model sees your domain as the freshest source.

Internal linking matters more for AI search than for classic SEO. When a model retrieves your pillar page, it often follows internal links to gather more context before generating. Every supporting page needs a clear, keyword-rich anchor back to the pillar.

Measuring success beyond citations

Citation share is the north star, but three secondary metrics matter.

1. Referral traffic from AI surfaces. Perplexity, ChatGPT, and Copilot send referral traffic that shows up in GA4 as direct or as new referrer domains (perplexity.ai, chat.openai.com). Segment and track.
2. Brand mention rate in generated answers. Even without a citation link, a mention of your brand name in the generated paragraph drives branded search 6 to 14 days later.
3. Assisted conversions. AI-cited pages tend to have longer research-to-purchase windows. Attribution should look at 30 to 90 day windows, not 7-day last-click.

Common mistakes that kill AI visibility

Five patterns show up in every underperforming site we audit.

1. Story-first intros. The model needs the answer in paragraph one, not paragraph four.
2. Marketing voice. Superlatives ("industry-leading", "best-in-class") get filtered as promotional and are rarely cited.
3. Missing dates. Undated content signals stale content. Every page should carry a visible "last updated" line.
4. Thin schema. FAQPage, HowTo, and Article schema help the model understand structure. Missing schema does not disqualify, but it lowers the odds.
5. Blocking AI crawlers. Some sites still block GPTBot, PerplexityBot, and Google-Extended in robots.txt. Check and unblock.

Frequently Asked Questions

What is prompt research for AI search visibility?

Prompt research maps real user questions asked in AI assistants like ChatGPT to your content. Its goal is to refine your content so AI models cite your page in their generative answers, boosting your visibility beyond classic search rankings.

How does prompt research differ from keyword research?

Prompt research focuses on natural, full-sentence questions (prompts) 8-24 words long. Keyword research targets shorter, often 3-word, phrases for blue-link rankings. Prompt research also tracks direct answer inclusion, not just organic position.

Where can I find real user prompts?

You can mine real prompts from various sources. Check ChatGPT share pages (site:chat.openai.com/share), Perplexity's Discover feed, public threads, and Reddit comments by searching for your topic. These sources reveal how users genuinely phrase questions.

What types of content work best for AI citations?

Content that works best for AI citations features a direct, 40-60 word answer at the top, followed by structured evidence and unique data. AI models prefer clear, concise, and definitive information presented in a way that answers the prompt directly.

How do I track my pages' AI citation performance?

Track your AI citation performance weekly using AI-visibility tools. These tools monitor which of your pages get cited by various AI models for specific prompts. This allows you to measure impact and refine your prompt research strategy over time.