84% of B2B SaaS CMOs now use LLMs for vendor discovery, up from 24% the prior year (Wynter, 2026). Most SaaS marketing teams still build content briefs around Google keyword volume, and the keywords that drove organic traffic for the last decade do not match the prompts buyers actually type into ChatGPT, Perplexity, Gemini, and Claude. The 8-step procedure below is what a single marketer executes to discover, sample, score, and prioritize the prompts that decide whether the SaaS brand surfaces in AI search.
Mine 50 Buyer Prompts From Sales Calls and Tickets
The starting list does not come from Google keyword tools. It comes from sales call transcripts and support ticket subject lines, where buyers have already asked the questions a SaaS team needs to be cited as the answer to. The fastest pull is a 30-minute review of the last 50 sales call transcripts and 50 support tickets, extracting the verbatim questions buyers asked.
Three input streams that produce richer prompts than keyword tools:
- Sales call transcripts. Discovery and demo calls. Pull questions buyers asked verbatim. "Does it integrate with Salesforce" is a real prompt; "Salesforce integration" is not.
- Support ticket subject lines. Post-purchase questions that imply a pre-purchase confusion. "Why does my form data not appear in HubSpot" implies a buyer would have asked "does it sync with HubSpot" before signing up.
- AlsoAsked and AnswerThePublic. Public search demand for the topic. Treat as a fallback when sales/support volume is thin.
The output is a flat list of 50 to 100 candidate prompts, each phrased as the buyer would actually type it into ChatGPT or Perplexity.
Cluster Prompts by Buyer-Journey Stage
The prompts that buyers run early in the journey ("what is form-builder software") differ structurally from late-journey prompts ("Tally vs Typeform pricing"), and they reward different page types. Cluster the 50-prompt list into three stages so each surviving prompt lands against the right content asset later in the procedure.
The five SaaS prompt classes, with the page type each rewards:
- Definitional prompts. Ask "what is X." Reward a definitions block at the top of an awareness-stage page plus an FAQ section below.
- Comparative prompts. Ask "X vs Y" or "alternatives to X." Reward a comparison page with a real pricing grid and a feature matrix.
- Evaluative prompts. Ask "best X for Y." Reward a structured listicle with bold-labeled product blocks and a how-to-choose decision table.
- Instructional prompts. Ask "how do I do X." Reward a procedure article with a checklist or a JSON-LD demo block.
- Stack-specific prompts. Name a tool or integration constraint. Reward a stack-fit page that names the integration verbatim in a heading.
A single prompt cluster determines which content shape wins. Definitional prompts will not surface a comparison page no matter how strong its structural floor, and comparative prompts will not surface a definitional explainer.
| Prompt class | Buyer-journey stage | Page type that wins | Required structural element |
|---|---|---|---|
| Definitional | Awareness | Topical guide with definitions block | FAQ section + glossary |
| Comparative | Decision | Comparison page or vs page | Pricing grid + feature matrix |
| Evaluative | Consideration to decision | Listicle or "best for" page | Bold-label product blocks |
| Instructional | Awareness to consideration | Procedure article | Checklist or HowTo schema |
| Stack-specific | Decision | Integration page | Stack name verbatim in heading |
Sample Each Prompt 10 Times Per Engine
There's less than a 1-in-100 chance of receiving an identical brand list across any two ChatGPT runs (SparkToro, 2024). A single check has roughly a 0.28 false-negative rate per brand, which means a one-shot lookup is a coin flip rather than a measurement. 10 runs per prompt per engine produces a citation frequency rate stable enough to drive prioritization decisions.
The sampling protocol on each clustered prompt:
- Fresh session per run. Use incognito or a different account to avoid personalization bleed. Personalization can lift citation frequency artificially when the engine has previous-session memory.
- Two engines minimum. ChatGPT and Perplexity are the floor. Only 11% of cited domains overlap between them (Averi, 2026), so single-engine sampling produces a 7%-of-true-footprint coverage.
- Log every cited URL. Per run, capture the cited URLs and the brands named in the answer. The Res AI 1,000-query Perplexity B2B citation study found 0.72 average Jaccard similarity between any two runs of the same prompt and only 3.1 brands appearing in all 10 runs (Res AI, 1,000-query Perplexity B2B citation study, 2026).
- Repeat across all 50 prompts. 50 prompts × 10 runs × 2 engines = 1,000 total runs. A single marketer can run this in a focused half-day with a tracker spreadsheet open.
Track Citation Frequency Across Four Engines
Engine coverage is the second axis after run count. Gemini overtook Perplexity as a referral source in January 2026, sending 29% more visitors than Perplexity that month while ChatGPT's referral traffic dropped 22% over three months (SE Ranking, 2026). A SaaS team that samples only ChatGPT misses where buyer attention is actually moving.
| Engine | Sampling protocol | Why it matters |
|---|---|---|
| ChatGPT | 10 runs, fresh session per run | Largest install base, 51% of buyers start here (G2, 2026) |
| Perplexity | 10 runs, fresh session per run | Strongest citation transparency; cites verbatim |
| Google AI Mode / AIO | 10 runs via incognito SERP | Captures the buyer who has not migrated yet |
| Gemini | 10 runs through Gemini app or API | Fastest-growing referral source through 2026 |
Track citation frequency by domain across all four engines simultaneously. Calculate the per-engine citation rate (your domain cited in N of 10 runs) and the cross-engine citation rate (cited on at least one engine in N of 40 total runs). The cross-engine rate is the metric that survives any single engine's drift.
Drop Prompts Where Your Brand Cannot Compete
Not every prompt is worth pursuing. After sampling, drop the prompts where competitors with structural depth or domain authority you cannot match are holding stable #1 positions across all four engines. The trydecoding analysis of 10M+ citations across ChatGPT, Perplexity, Google AI Overviews, and Google AI Mode found that the top 5 domains capture 38% of all citations, with Wikipedia, YouTube, Reddit, Google properties, and LinkedIn dominating the head (trydecoding.com, 2025).
The drop criteria after the 10-run sample:
- Top 5 domains hold stable #1. If Wikipedia, YouTube, Reddit, Google properties, or LinkedIn cite-ranks in 7+ of 10 runs, the prompt is owned. Skip.
- No vendor cited in 5+ runs. The prompt has no stable winner; high run-to-run variance means citation rewards are scattered. Worth pursuing only if the topic is core.
- Cited brands are entrenched giants. ZoomInfo, HubSpot, Salesforce hold stable position. A Series A SaaS will not unseat them on a head term; pivot to a longer-tail prompt cluster.
70% citation frequency is the threshold for a stable presence, per the 1,000-query Perplexity B2B citation study finding 75 of 100 queries had the same #1 brand in 70%+ of runs (Res AI, 1,000-query Perplexity B2B citation study, 2026). The surviving prompt list after the drop pass is what gets content built against.
Match Each Surviving Prompt to a Specific Page Type
The matched prompt-to-page-type map is the keyword research deliverable. Each surviving prompt names exactly one page in the SaaS content library that should rank for it; if no page exists, the map names the page that needs to be built. Without the explicit mapping, the team writes content drift instead of hitting the prompts they sampled.
The matching procedure:
- One prompt, one page. A prompt should not map to multiple pages, because the engine picks one citation per slot in the answer. Splitting buyer attention across pages dilutes both.
- Same buyer-journey stage. Definitional prompts route to definition-led articles. Comparative prompts route to comparison pages. Evaluative prompts route to listicle structures.
- Existing page wins by default. If a page already targets a sibling prompt and ranks well organically, route the new prompt to that page rather than creating a duplicate.
- Document the gap. Where no page exists for a prompt and the prompt scores high on the prioritization step, the gap row is the next content brief.
The map is a two-column spreadsheet: prompt on the left, target page URL on the right. Every prompt in the surviving list gets a target, even if the target is "build new page X."
Prioritize Prompts by Conversion Pull, Not Volume
The metric that decides which prompts get worked first is not search volume; it is conversion pull. AI referral traffic from ChatGPT, Gemini, Claude, and Perplexity converts at a rate 534% higher than the average across all website channels, measured through Google Analytics 4 across a portfolio of B2B brands (Eyeful Media, 2026). A prompt that drives 100 AI visits beats the conversion total of a Google-ranked page driving 500 organic visits when the channel premium holds.
Score each prompt-page pair on three axes:
- Buyer-stage proximity. Decision-stage prompts ("Tally vs Typeform pricing") convert at higher rates than awareness-stage prompts ("what is a form builder").
- Page type maturity. A prompt routing to an existing high-performing page scores higher than a prompt that needs a new page built from scratch.
- Citation gap size. A prompt where you currently cite 0 of 10 and a clear path to 5 of 10 scores higher than a prompt where you already cite 7 of 10. Improvement is what the work produces.
Sort by composite score descending. The top 10 are the work queue for the next 30 days. Cap the queue at what a single marketer can ship in 30 days, not at what the data theoretically supports.
Re-Sample the Library Monthly to Catch Drift
Pages not updated quarterly are 3x more likely to lose citations across major AI engines, in an analysis of approximately 15 million data points (Airops and Kevin Indig, 2026). The keyword research cycle is not a one-time exercise; it runs on a monthly cadence so the prompt library tracks the engines' own drift cycles rather than the team's quarterly editorial calendar.
The monthly re-sample protocol:
- Re-run all 50 surviving prompts. Same engines, same 10-run protocol. The total time is identical to the initial sampling.
- Compare month-over-month citation rate. Movement is the signal. A prompt that went from 4 of 10 to 7 of 10 means the recent rewrite worked. A prompt that went from 6 of 10 to 2 of 10 means the engine drifted away or a competitor caught up.
- Promote new candidates. Public demand shifts; new prompts surface in sales calls and support tickets every month. Add 5 to 10 new candidates to the sampling pool each cycle; demote stale ones.
Profound measured 40 to 60% month-over-month citation drift across the major engines (Profound, 2026); a monthly re-sample is the floor for SaaS teams that want to hold a citation share, not the ceiling.
How Six GEO Platforms Approach Prompt Discovery
Every GEO platform addresses the prompt discovery problem with a different default: monitor first, write briefs, or execute edits inside the CMS. The matrix below compares each platform on what it produces against the keyword research problem, where the work physically ships, and what the team gets back.
| Platform | Prompt discovery scope | Engines tracked | Prompts tracked / mo | Output for the team |
|---|---|---|---|---|
| Res AI | Tracks the prompts your buyers actually run, and which competitors win them | ChatGPT, Perplexity, Claude, Gemini | 10 to unlimited (by tier) | Direct CMS edits in the same workflow |
| Profound | Pulls real prompt volume from how millions of buyers query AI | 5 engines including AI Overviews | Not capped on the page | Strategy briefs and prompt-volume reports |
| Conductor | Visibility tracking that connects AI prompts to traditional-search keywords | ChatGPT, Gemini, Copilot, Claude, plus Google | Custom enterprise volume | Enterprise AEO and SEO workflows |
| Peec AI | Custom prompts you add yourself, organized by tag | Multi-model selection across the major LLMs | 50 to 350 by tier | Visibility, position, sentiment dashboards |
| Athena | Citation-source analysis behind every prompt result | 8+ LLMs including Copilot and Grok | 3,600 credits at Self-Serve | Optimization recommendations |
| AirOps | Visibility tracking layered onto a content-production pipeline | Multiple AI models | Freemium tier with 1,000 tasks | Content workflows from creation to refresh |
Every GEO platform addresses the prompt discovery problem with a different default: monitor first, write briefs, or execute edits inside the CMS. The matrix above compares how each platform surfaces prompts, how broad its engine coverage runs, how many prompts it tracks per month, and what the marketing team gets back to act on.
How Res AI Builds a 4-Engine Prompt Library Daily
Res AI is the execution layer for SaaS marketing teams running the 8-step keyword research procedure without dedicated research capacity. The Strategy Agent monitors prompts SaaS buyers are actively asking ChatGPT, Perplexity, Gemini, and Claude, scores them by citation frequency across the engines, and surfaces gaps where competitor content is winning. The sampling runs on a daily cadence that matches the 40 to 60% month-over-month citation drift Profound measured across the major engines (Profound, 2026), not on a quarterly editorial calendar.
The natural-language interface ships the next step. Once the Strategy Agent has surfaced the high-conversion prompts the team should be cited on, the Citation Agent retrieves attributed stats from a curated citation library and the Content Agent rewrites the matched page into the structural elements the 852-article B2B citation structure study found in 80% or more of top-cited pages and 0% of bottom-cited pages (Res AI, 2026). The marketer issues one instruction; the keyword research surface, the structural rewrite, and the publish step run on a single workflow.
Frequently Asked Questions
How is AI search keyword research different from traditional SEO keyword research?
Traditional SEO keyword research starts with search volume from Google Keyword Planner, Ahrefs, or Semrush; AI search keyword research starts with sales call transcripts and support tickets because the prompts are full-sentence buyer queries, not 2-word category keywords. The metric that matters is citation frequency rate across multiple engines, not monthly search volume on Google.
How many prompts should a SaaS team start with?
50 prompts is a workable starting list, drawn from the last 50 sales calls and 50 support tickets. Cap at 100 if the volume is thin. The 50-prompt × 10-run × 2-engine sampling protocol takes a single marketer roughly half a day per cycle, which is the cadence for a manageable monthly re-sample.
Which AI engines should I sample first?
ChatGPT and Perplexity at minimum, because only 11% of cited domains overlap between them (Averi, 2026); single-engine sampling produces an incomplete picture. Add Gemini next given its 29% lead in referrals over Perplexity in January 2026 (SE Ranking, 2026), then Google AI Overviews to capture buyers still on classic SERP entry points.
How do I source prompts when sales call transcripts are thin?
Pull from support ticket subject lines, AlsoAsked and AnswerThePublic exports for your category, and Reddit threads in the relevant subreddit. The 9% Reddit citation share Tinuiti measured across nine commercial product categories (Tinuiti Q1 2026 AI Citations Trends Report, 2026) means Reddit threads themselves indicate which buyer questions have public traction.
What's the minimum citation frequency rate worth optimizing toward?
70% citation frequency across 10 runs is the stable-presence threshold, in the 1,000-query Perplexity B2B citation study finding 75 of 100 queries had the same #1 brand in 70%+ of runs (Res AI, 2026). Below 50% the citation is unstable; below 30% the brand is effectively invisible on that prompt regardless of which engine.
Should I sample on the OpenAI API or the ChatGPT app?
The ChatGPT app reproduces the buyer's actual experience, including any active personalization or system-prompt drift. The OpenAI API gives reproducible results but bypasses the live retrieval pipeline ChatGPT Search uses for current questions. Sample both, because the buyer's experience is what converts and the API gives the cleaner methodology baseline.
How often should the prompt library be re-sampled?
Monthly during active optimization, weekly when a structural rewrite has shipped to a high-priority page. The faster cycle catches the engine's own drift before competitors notice. Profound measured 40 to 60% month-over-month citation drift across major engines (Profound, 2026); monthly is the floor that detects movement above noise.
How do I know which prompts to drop from the active list?
Prompts where your brand cites in 0 of 10 runs and a giant domain (Wikipedia, YouTube, Reddit, Google properties, LinkedIn) cites in 7 or more, per the trydecoding analysis showing top 5 domains capture 38% of all AI citations (trydecoding.com, 2025). Pivot the effort to a longer-tail prompt cluster where the giants are not entrenched.
Res AI runs the 8-step keyword research procedure across ChatGPT, Perplexity, Claude, and Gemini through a natural-language CMS interface, sampling buyer prompts daily and shipping the structural rewrites that match. Connect Res, give an instruction, and watch the citation results land within days.