Prompts Aren't Keywords: How to Choose Prompts to Track

Prompts aren't keywords, so choosing which prompts to track works differently from building a keyword list. A keyword names a topic, while a prompt describes a buying situation with the company, its stack, and the conditions a vendor must meet. Knowing how to choose prompts to track starts from those situations.

In Semrush's US user panel, at least 65% of the ChatGPT prompts matched no keyword in its database for most of October 2024 to February 2026.

For a B2B SaaS, fintech, or AI company, the choice should focus on the bottom of the funnel, where a prospect asks for a shortlist, a comparison, or an alternative. Those prompts carry the conditions that rule vendors in or out, such as an accounting stack or a security certification, and show whether a company makes the shortlist. This guide shows how to choose them, how many to run, and how to read the results.

Key takeaways

• Choose prompts from the buying situations that decide deals, such as a finance lead at a 200-person German company on Xero that needs ISO 27001.
• Keyword data shows topic demand, while sales notes and deal records supply the situations and wording to track.
• Requirement prompts measure shortlist inclusion, because the answers show which products were kept, flagged, or dropped once conditions were added.
• Start with one canonical prompt per buying situation, then test rephrasings during the baseline and keep the variants that change the results.
• Read shares across repeated runs and several weeks, engine by engine, because one answer says little about where a company stands.

What is prompt tracking, and why aren't prompts keywords?

Prompt tracking, also called AI visibility tracking or AI prompt tracking, is the practice of running a fixed set of prompts through AI search engines on a schedule to measure brand visibility. Each answer is checked for whether it mentions, cites, or recommends a company. It is the measurement side of answer engine optimization (AEO) and generative engine optimization (GEO), and it samples answers from large language model (LLM) engines.

The set is a sample, never a census, because the major engines do not publish the prompts people type or how often they type them. Bing Webmaster Tools offers one first-party view, listing a sample of a website's grounding queries, the search phrases Microsoft's AI experiences used to retrieve the pages they cited.

Google describes the move from keywords to longer requests. On May 19, 2026, Elizabeth Reid, vice president of Search, introduced the biggest upgrade to the Search box in over 25 years, because "curiosity doesn't always fit into keywords."

Google's US usage analysis, published the same day, put the average AI Mode search at triple the length of a traditional Search query. The spend management request used later in this guide runs to 26 words against 10 for the plain question, and the extra words carry the company context, the requirements, and the requested output.

G2's March 2026 survey shows how often vendor research now starts in a chat. It asked 1,076 people who make or influence B2B software purchases how they research, and 51% said they start with an AI chatbot more often than Google.

In the same survey, 69% said a chatbot led them to a different vendor than they had planned, the shortlist effect a prompt set has to measure.

Four dated changes from the past six months show where the measurement is heading:

• May 19, 2026: Google said AI Mode had passed one billion monthly users.
• June 16, 2026: Bing added Intents and Topics to its AI Performance report in a global preview, sorting the grounding queries behind a website's AI citations.
• October 12, 2026: Ahrefs published an analysis of the fan-out queries behind logged-out ChatGPT responses.
• October 13, 2026: Bing launched Topics Portfolio in beta, grouping a website's grounding queries into themes with citation-based estimates of each topic's size.

How does an AI engine turn a prompt into a shortlist?

Four steps lead from what a prospect types to the answer a tracker records, and each one shapes what a tracked prompt must contain.

1. The prompt carries the buying situation. Software research along the customer journey includes category discovery, comparisons, and requirement checks, so a prompt set needs questions for each stage it reports on. G2 found that two-thirds start their AI research with a category or competitor query, and one in five starts with requirements or process. Respondents also use chatbots to draft requests for proposals (RFPs), examine pricing and packaging, and validate product fit.

In Profound's panel of 7.5 million ChatGPT conversations, commercial ones averaged 2.43 turns by June 2026 and had the lowest single-turn share of the three intents it classified. A tracked prompt is a synthetic summary of those turns, so we write each one to carry the company context and the requirements that define the situation.

2. When the engine searches, it fans the prompt out. OpenAI's help page says that when ChatGPT searches through its search partners, it typically rewrites the question into one or more targeted queries, the sub-queries behind the answer. After reviewing the first results, it may send more specific ones. Google calls this query fan-out. AI Mode breaks a question into subtopics and issues many searches at once, and its Deep Search can issue hundreds.

These targeted queries belong to answers where ChatGPT searches, and Semrush found web search turned on for 34.5% of sampled queries in February 2026. Ahrefs measured recorded fan-out queries separately. In its sample of 66.6 million free, logged-out responses, those containing fan-outs averaged about 1.6 each.

A tracker should record whether each run searched and keep the searches and citations it can see, because those records show which web evidence supported an answer. Keyword research supplies demand context, grounding queries show the phrases an engine used to retrieve pages, citations show the pages it referenced, and the unit we track stays the prompt.

3. Requirements narrow and reorder the shortlist. Once the searches return pages, the engine weighs candidates against what the prompt asked for. Rand Fishkin of SparkToro saw it with headphones, where requests for gaming, podcasting, or noise suppression models returned very different lists from the general request.

A month-end close comparison on October 11 showed the same in B2B software. Once the prompt required NetSuite, EU data residency, and SOC 2 Type II, ChatGPT ranked a different vendor first, citing its public documentation of SOC 2 Type II and EU hosting. In our coding, each product ends up kept, kept with a condition flagged to confirm, or dropped. A requirement prompt therefore measures selection and a category prompt measures recognition, so requirement prompts form a group of their own in the set.

4. Each answer is one sample, and repeated runs show the spread. The last step explains why one reading misleads. In SparkToro's study with Gumshoe, 600 volunteers ran 12 prompts through ChatGPT, Claude, and Google's AI 2,961 times. The study put the chance that ChatGPT or Google's AI returns the same brand list twice at under 1 in 100, with Claude only slightly more likely to repeat itself.

The sources behind the answers rotate as well. In SISTRIX's analysis, 54% to 59% of the domains Google AI Mode cited each week were new since the week before, across six countries. A stable core of one to five domains still held in 86.5% of AI Mode prompts, while the domains around it rotated. Repeated runs estimate how often a company appears, and a window of several weeks shows whether that frequency holds.

What makes a prompt worth tracking for AI visibility?

Those four steps give a prompt four ways to earn a slot.

1. It sits close to the decision. A prompt close to the decision asks for a shortlist with conditions, a comparison, or an alternative. We give these prompts priority because the report measures software evaluation, and their weight comes from the decisions they represent and the requirements recorded in deals.

What to do: Give most slots to shortlist, comparison, and alternative prompts, and give informational queries a slot only when the report has a use for them.

• ❌ What is a payout API?
• ✔ Which payout providers let a UK marketplace pay sellers in 30 countries through an API that works with Stripe Connect? Compare fees.

2. It carries the requirements that decide deals. A requirement prompt records the role, company size, region, stack, and hard requirements that define a buying situation for the ideal customer profile (ICP). These conditions narrow the pool of candidates. In the Don't Measure Once study's repeated-run analysis, source overlap varied by prompt from below 0.2 to above 0.8, where 1 means identical sets, and the authors associated that spread with query specificity. We propose testing whether requirement prompts give steadier B2B selection rates than category prompts run under the same settings.

What to do: Pick the three to five requirements that most often decide deals for the segment, and combine them as prospects do.

• ❌ best spend management software
• ✔ Shortlist three spend management platforms for a 200-person company in Germany with several legal entities that uses Xero and needs ISO 27001. Include pricing for each.

3. It comes from what prospects asked. The wording comes from call notes, demo forms, lost deals, and RFPs, because those hold the prospect's own words. Glen Allsopp, head of marketing strategy and research at Ahrefs, notes in his guide that "platforms don't directly share the queries users are entering, or how often," so a keyword-built list may track phrases nobody types. We start from closed-lost notes that record a specific requirement or objection, then check whether it decided the deal.

Ahrefs still anchors most of its own tracked prompts on search terms, because those carry search volume data. Keyword demand picks the topics, and the prospect's words carry the situation.

What to do: Keep the prospect's nouns (multi-entity, subsidiaries, Xero), role, and stack, write in conversational wording, and leave the tracked company's name out of unbranded prompts.

• ❌ ai customer service agent zendesk eu
• ✔ A 50-agent support team on Zendesk needs customer data to stay in the EU. Which AI agents resolve tier-1 tickets inside Zendesk and provide a current SOC 2 Type II report?

4. It stands for one buying situation, whatever the wording. A canonical prompt is one wording chosen to stand for a buying situation. In SparkToro's panel, 142 people each wrote their own prompt for the same headphone need. Their wordings had an average semantic similarity of 0.081, which SparkToro likens to two dishes that share peanuts and little else. Yet across 994 answers, Bose, Sony, Sennheiser, and Apple appeared 55% to 77% of the time.

The effect weakens in wide categories, where the leading names for brand design services peaked at visibility in the 30s and 40s, in percent, so the method needs a check.

What to do: Treat the check as a pilot procedure we propose. Run two or three phrasings through the baseline weeks, compare their selection share by engine, and keep a variant only when it tests a distinct requirement or its share stays different across the baseline weeks.

• ❌ Five tracked prompts for one situation, from best contract management software Salesforce to which contract tool integrates with Salesforce
• ✔ One canonical prompt, Which contract management platforms integrate with Salesforce and support single sign-on for a 300-person SaaS company?, with two rephrasings run for four weeks

What happens when a prompt adds hard requirements?

The four properties are easiest to see on one request. On October 11, 2026, we ran a comparison across five B2B SaaS categories on ChatGPT, Claude, Gemini, Google AI Mode, and Google AI Overviews. We later shared the results on X, and the spend management pair shows every step.

1. The two prompts. The plain question asked What is the best spend management software for mid-sized companies? The requirement prompt asked for three platforms for a 200-person company in Germany with several legal entities, on Xero, needing ISO 27001, with pricing.
2. The plain answers. Ramp ranked first on ChatGPT, Claude, Gemini, and AI Mode. Payhawk appeared on two engines.
3. What the requirement answers read. Gemini cited two vendors' pricing pages. AI Mode cited 19 sources: Payhawk's own page, Xero's security page, review platforms, directories, and other vendors' posts.
4. What they decided. Ramp made none of the five lists, which our post put down to its waitlist for European companies. Payhawk made all five, and Moss, Mobilexpense, and Pleo were added.
5. What was flagged. Claude kept Payhawk with "ISO 27001 (Confirm with vendor)" beside it, although Payhawk documents the certification in its trust portal.
6. What a tracker logs. Each answer is coded selected, selected but flagged, or not selected.

Across the five categories, the plain-question leader stayed first in 9 of 25 engine-category pairs once three requirements were added. It ranked lower in 8 and dropped out of 8, a reshuffle like the one SparkToro saw with headphones. The requirement prompt also asked for three products with pricing, so the comparison describes the shift without isolating the requirements. As we wrote in the post, "Being visible for a category and being selected for a specific requirement are very different things."

Read more: Agentic SEO guide

How to choose prompts to track for AI visibility

To choose prompts to track, a company starts from the conversion it cares about, collects the questions prospects ask just before it, and writes one prompt per buying situation, grouped and sized to its run budget.

1. Write down the conversion event and the segments. The set is built around one event, such as a demo request, a trial, or a sign-up, and the ICP segments behind it, so a prompt nobody in those segments would type before that event gets no slot.

2. Collect the last questions and the requirements. The raw material sits in call recordings, CRM notes, demo-form fields, RFPs, security questionnaires, and closed-lost reasons, sorted by ICP segment and persona. These records reflect the prospects the company has recorded, so review-platform filters, competitors' review pages, and Reddit threads add other requirements to investigate, including questions absent from them.

Three sources then check the wording and topics: the question-shaped queries in Google Search Console that earned the website impressions, Bing's grounding queries with the page cited for each, and keyword data for demand. Tracker-suggested prompts, such as the prompt research reports Profound launched in June 2026, serve as a coverage check.

3. Write one canonical prompt per buying situation. Each situation gets one prompt with its segment's three to five deciding requirements, checked against two or three rephrasings as factor 4 describes.

4. Sort the prompts into four groups. Each group answers a different question and is read on its own.

5. Record how often each buying situation comes up in deals.

Each distinct buying situation gets one prompt and a record of its share of recent deals, won and lost. The count comes from a CRM field or a call recorder's topic tags, and the deal shares are applied once, in a separate roll-up for leadership, with branded prompts left out of it.

6. Size the set from the run budget. Answer volume equals prompts times engines times runs per prompt times locations. In a worked example, 40 prompts (24 requirement, 8 competitor, 4 category, and 4 branded) on 3 engines at 7 runs from one location come to 840 answers per reading.

The Don't Measure Once study advises "A large prompt portfolio of diverse queries" as well as repeated runs. We run the set daily, one run per prompt per engine, and read it weekly, so each reading holds about seven runs of every prompt.

Repeated runs matter most when a decision rests on one prompt. Across a 753-prompt US portfolio on seven platforms over two weeks, Profound measured visibility of 78.7% with one daily run and 80.4% with ten. Its explanation: "Averaging one draw across a large portfolio does most of the work that extra runs would have done."

7. Choose the engines and the market. The engines are the ones the ICP uses, and the runs come from the ICP's market and language where the tracker allows. A prompt about a German company runs from Germany where possible, and the run location is recorded with each answer. In SISTRIX's citation-drift analysis, 74% of the domains ChatGPT search cited each week were new in Germany, against 42% in France.

8. Freeze the set, then review it each quarter. The set stays fixed through a baseline period, because an edited prompt breaks its own trend line, and each quarter new deal notes decide which prompts stay.

How to read and report prompt tracking results

Once the set runs, three measures of prompt performance carry the report, and the tracking setup guide covers the tool side.

Visibility rate shows recognition. Visibility rate, which some trackers call mention rate, is the share of valid completed answers that name the company, reported by group and engine. Tracker dashboards weight their own visibility scores differently, some by position, so the report states which measure it uses. The branded group also tracks sentiment and factual accuracy. Rand Fishkin of SparkToro concluded that "visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric".

How to measure: Count mentions across all runs of a group's prompts on one engine, and report average position with the rule used for runs that omit the company, next to the visibility rate.

Selection share shows who gets picked. Selection share, the measure we proposed in our post, counts how often a company makes the shortlist once specific requirements are added. It counts only requirement prompts, and any share of voice measure needs its own competitor set and denominator.

How to measure: Divide the completed requirement answers that shortlist the company by all completed requirement answers on the same engine. Flagged answers count in selection share, and flagged share is flagged answers divided by the same total.

The fix list names what to publish.

Review flagged and dropped answers before assigning work. A documented error or a missing requirement on the website supports a content task for the pricing, integration, security, or comparison page, while a real product limit belongs with product or sales. The month-end close answer shows what a documented position looks like: it cited the leader's security page, certifications page, subprocessor list, and NetSuite integration page.

How to tell when a movement is worth acting on. Kevin Indig, who writes Growth Memo, sets the posture in one line: "Prompt tracking should be closer to polling and focus groups than SEO rank tracking." The Don't Measure Once study shows why, since the margin of error on a brand's rate narrows only as runs per prompt rise.

Compare equivalent periods with the same prompts, engine, market, and collection settings. Report the underlying counts, and check whether a movement persists across prompts and across successive weekly readings. The study's seven-run margin describes its own sampling exercise, so a claim that a change is significant needs uncertainty calculated for the actual comparison of two periods.

Each engine gets its own baseline, because SISTRIX found that only 17% of the domains cited by AI Overviews and AI Mode for the same prompt appeared in both.

A worked example for an invented spend management platform, Vendor A, shows what one weekly reading reports.

Changes refer to selection share on requirement rows and visibility rate on category and competitor rows.

For the CEO, selection share sits next to sign-up events from AI referrers such as chatgpt.com and perplexity.ai in Google Analytics 4, the demo form's self-reported source, and pipeline. This guide does not validate selection share against pipeline, so the numbers sit side by side as context. A movement earns attention when it persists across prompts and weeks on the same engine.

How we choose prompts to track at Omnius

Omnius, an SEO and AI search company for SaaS, fintech, and AI businesses, treats a prompt set as a sample of buying situations. AI answers vary from run to run, so a score built on one answer claims more certainty than it has. Our work starts with what prospects ask and ends with the facts engines could not confirm. In the October 11 comparison, where the answers found no first-party price, ChatGPT filled the gap with third-party purchasing estimates and labeled them. A missing price therefore goes on the fix list.

• Buying situations from sales calls, CRM notes, and lost deals.
• Canonical prompts in four groups, with deal weights applied once, in the leadership roll-up.
• Repeated runs per engine from the ICP's market, with searched runs logged.
• Selection share next to visibility rate, with a fix list of unconfirmed conditions.

Atomic AGI, the AI search analytics software built inside Omnius, runs the prompts and keeps each answer for those readings.

More on finding the conditions engines cannot confirm: AI search audit.

FAQ

How many prompts should a company track?

About 40 prompts across four groups, the size of this guide's worked example, makes a workable first set for one product and one market, and the run budget sets the final count. With thin deal data, a company starts with the three or four buying situations sales hears most and adds groups as notes build up.

Which AI engines should a B2B company track?

A B2B company should track the engines its ICP uses, each with its own baseline. In G2's survey ChatGPT led in every industry and company size, so most sets start there and add the engines the ICP's own usage points to, such as Google AI Mode.

Can SEO keywords be used as prompts?

SEO keywords make useful inputs rather than finished prompts, because they show which topics people search in Google. The tracked prompt itself should read like the prospect's request, and the phrases an engine searched show up in records such as Bing's grounding queries.

Should a company use its tracker's suggested prompts and prompt volumes?

Suggested prompts make a good coverage check. Each one earns a slot only after a check against the target segment and the report's purpose. Suggestions drawn from panel data show topic interest, while a prompt volume needs a stated population, collection method, and time window before it guides a choice. Panel demand alone does not show what this company's prospects ask before converting.

Methodology

The conclusions come from documentation, studies, and one comparison we ran, read together.

• How we reached the conclusions. We read the platform documentation of OpenAI, Google, Anthropic, and Microsoft Bing on how a prompt becomes searches, the studies on run-to-run variation and prompt language listed below, and the guides that rank for the query. The worked spend management example comes from a comparison we ran on October 11, 2026, with our own tracking software: one plain and one requirement prompt in each of five B2B SaaS categories, one run per engine on ChatGPT, Claude, Gemini, Google AI Mode, and AI Overviews, from a US location, so the Germany prompt shows the US view of that request. We described it on X on October 17, 2026. The screenshots were captured on October 27, 2026.
• Limits. One run per engine on one day illustrates the answers returned and establishes no stable rates, and the requirement prompt also asked for three products with pricing. The run counts and margins come from consumer verticals in one market. That one canonical prompt stands for its rephrasings is an inference the proposed phrasing check tests, and whether requirement prompts read steadier needs a separate repeated-run comparison. This guide does not validate selection share against pipeline. The Vendor A reading is invented to show the rule.
• Sources and their interests. Omnius sells SEO and AI search services and ran the comparison. Semrush, SISTRIX, Ahrefs, and Profound sell AI visibility or SEO software; Gumshoe, SparkToro's partner in the study, sells AI tracking; G2 sells visibility on its review platform; the Don't Measure Once data came from a GEO measurement start-up's export, and one author discloses an affiliation with that company.

The sources, grouped by publisher:

• OpenAI: Searching the web with ChatGPT
• Google: AI Mode at I/O 2025; AI Mode US insights; Search at I/O 2026; AI features documentation
• Anthropic: Web search tool
• Microsoft Bing: AI Performance; Intents and Topics; Topics Portfolio
• Studies and guides: Profound on commercial conversations; Growth Memo, AI halftime report; Ahrefs on choosing prompts; Ahrefs on ChatGPT fan-out queries; Profound, Is once a day enough?; Profound Prompt Research Reports