How to Choose AI Search Prompts to Track in AI Visibility Tools: Coverage, Count, Rotation

Quick answer

Under this method, Canlah AI recommends tracking 30 to 40 AI search prompts written as buyer questions across five funnel layers, from unbranded category, shortlist and segment questions to branded comparison and verification questions, with branded questions capped at about a third of the set. Every question runs at least three times per engine, so the output is a range rather than one answer. The method’s base quota is four category-awareness questions, eight main-shelf questions, eight segment questions, six brand-comparison questions and four brand-verification questions, which is 30 questions before any candidate testing.

Canlah AI ran that panel against its own site on September 7, 2026: 39 questions across six layers, three runs each on two engines, for 234 planned samples of which 233 returned. The four category-awareness questions scored a 0.0 mention rate on both engines while the four brand-verification questions scored 1.0 on both. A branded-heavy panel would have reported that brand as visible.

Treat the panel as an instrument rather than a scoreboard. A prompt set that changes between cycles cannot support a before-and-after comparison, and no panel size proves causation, stable share of voice or a future recommendation. Freeze the set first and argue about the score afterwards.

Disclosure: Canlah AI publishes this guide and operates the managed measurement service described below. Prompt allowances for Ahrefs Brand Radar, Otterly.ai, Peec AI, Profound, Scrunch AI, Semrush, AthenaHQ, Rankscale and LLMrefs come from public pricing pages reviewed on September 23, 2026 and were not verified through paid accounts. The Canlah AI figures come from a self-audit of canlah.ai run on September 7, 2026, internal and not independently audited.

How many prompts should you track for AI search?

For one brand in one market and one language, 30 to 40 tracked prompts is the working band, and the binding constraint is runs rather than questions. A panel is priced in observations: questions multiplied by runs per question multiplied by engines. Forty questions at three runs on two engines produces 240 observations per cycle. The same 40 at one run on one engine produces 40 observations and no range at all.

Published plan limits are slot counts, not observation counts, and the two diverge. Ahrefs Brand Radar includes five AI prompts on Lite, 10 on Standard, 20 on Advanced and from 83 on Enterprise. Otterly.ai sells 15 prompts at US$29, 100 at US$189 and 400 at US$489. Peec AI lists 50 on Starter, 150 on Pro and 350 on Advanced. Semrush states 25 on the AI Visibility Toolkit, 50 on Semrush One Starter and up to 200 on Advanced.

A 20-slot plan leaves a 30-question funnel panel 10 questions short. That gap usually surfaces after the contract is signed rather than during the demo, so count the panel before comparing plans.

What a tracked prompt set has to cover

A tracked prompt set has to cover five buyer layers, because a brand can score 1.0 in one layer and 0.0 in the next. The layers map to what a buyer already knows when the question is typed, from no vendor in mind to checking one vendor’s reputation. Coverage failures are directional: a panel skewed toward the bottom two layers flatters the brand, and one skewed toward the top two understates it.

The five-layer buyer funnel, the method’s base quota and what each layer proves. Self-audit of canlah.ai, September 7, 2026.

Layer	Questions	Branded	Canlah AI self-audit result	What the layer proves
L1 Category awareness	4	No	0.0 on both engines; cited Wikipedia, arXiv and Google Search Central	Whether the engine explains the category, and which sources it trusts to do so
L2 Main shelf	8	No	0.0 to 1.0 on OpenAI; 0.0 on Gemini	Whether the brand appears on an unbranded “who are the best” shortlist
L3 Segment shelf	8	No	0.0 to 0.67 on OpenAI; 0.0 on Gemini	Whether the brand appears once the question names a segment, budget or engine
L4 Brand comparison	6	Yes	1.0 on both, from 3 questions rather than 6	How the engine positions the brand against alternatives
L5 Brand verification	4	Yes	1.0 on both engines	Whether “is this company legitimate” returns accurate facts

Source: Canlah AI self-audit of canlah.ai, September 7, 2026, three runs per question on an OpenAI leg and a Gemini leg; internal and not independently audited. Mention rates are the range across the questions in a layer, never a single run, so “0.0 to 1.0” means at least one question returned no mention and at least one returned a mention in every run.

Ten of the 30 questions are branded, which is a third of the panel, and that ratio matters more than the total. A brand that fills 20 slots with branded questions buys a dashboard reporting 1.0 every month while never detecting the shelf it is missing from.

Scrunch AI is the one vendor among the nine reviewed whose public pricing page names funnel stages as a tracked dimension, on both Core and Enterprise. A funnel-stage dimension was not found on the other eight vendors’ reviewed pages on September 23, 2026, which leaves the layer design to the buyer on those plans.

Where the questions come from

Tracked AI search prompts come from four sources: People Also Ask, Google autocomplete, intent templates and sales-call transcripts. A panel built from only one of them is biased in a predictable direction, and the Canlah AI self-audit drew its 39 questions from the first three, tagging each with its source.

• People Also Ask: the real phrasing of category and cost questions. It produced all four L1 questions, including “How much does hiring an AI search optimization agency typically cost in Singapore?”
• Autocomplete: short, unpolished head questions such as “best generative engine optimization agencies?” that no internal team writes by hand.
• Intent templates: the comparison and verification questions search data rarely exposes, such as “Does this company actually deliver verified results, or is it mostly hype?”
• Sales-call transcripts: the objection wording buyers use with a salesperson and then repeat to an assistant. This source is first-party and already in hand.

One test settles the balance: count how many questions a person who has never heard of the company could type. In the self-audit, 32 of 39 passed, because only the seven L4 and L5 questions named the brand.

How to choose prompts for LLM tracking, step by step

Choosing prompts for LLM tracking takes six steps, from writing the layer quota to freezing the panel before the first cycle. The order matters, because the freeze has to come before the first measurement rather than after the first disappointing one.

Step 1: Write the layer quota before writing any questions

Decide the count per layer first, then fill it: four, eight, eight, six and four on a 30-question panel. A blank list produces a panel weighted toward whatever the team already worried about, usually brand reputation.

Step 2: Harvest from People Also Ask and autocomplete before writing anything

Pull the live phrasings and edit them lightly. A question rewritten into marketing language stops matching how buyers ask, and in the self-audit all four L1 questions came from People Also Ask.

Step 3: Cap branded questions at a third of the panel

Branded questions answer a real procurement worry and inflate every aggregate score. Keep 10 of 30 at most, report them separately and never let them set the headline.

Step 4: Record the conditions alongside the question

Log engine, model version, market, language, account state, date and run index on every observation. The Canlah AI self-audit recorded a Gemini fallback from gemini-flash-latest to gemini-flash-lite-latest, the detail that explains a score change three cycles later.

Step 5: Set runs per question before the first cycle

Three runs per question per engine is the practical floor for a range, which puts a 30-question panel on two engines at 180 observations. One run records an answer that existed once, not what a buyer is likely to meet.

Step 6: Freeze the panel and open a separate candidate layer

Lock the counted set for the full comparison window. New ideas go into an add-on layer probed and disclosed on its own, as the 12 candidates in the self-audit were, so they cannot move the frozen panel’s numbers.

Prompt allowances across nine AI visibility tools

Prompt allowances across the nine platforms in this comparison run from five slots on Ahrefs Brand Radar Lite to 400 prompts on the US$489 Otterly.ai plan below Enterprise, and three of the nine bill by something other than a prompt slot. AthenaHQ and Rankscale bill by credit and LLMrefs by keyword, and the billing unit decides how expensive rotation is, which makes it the most important line in a quote for a team running a funnel panel.

Prompt allowances, billing units and rotation cost for Canlah AI and nine AI visibility platforms, reviewed September 23, 2026.

Tool	Best for	Prompt allowance at the stated price	Billing unit	Cost of rotating a question	Main point to verify
Canlah AI	Teams that want the panel designed, run and archived by a managed team	4, 8 or 16 buyer intents by tier, drawn from the five-layer funnel, plus a candidate layer	Managed engagement, not per prompt; GEO service from $1,500 a month per AI platform	A scope decision recorded in the report	Which tier covers the buyer intents and AI platforms you need; get anything beyond the plan quoted in writing
Ahrefs Brand Radar	Teams already on Ahrefs that want AI prompts beside their SEO data	5 on Lite, 10 on Standard, 20 on Advanced, from 83 on Enterprise	Prompt slot plus monthly checks	Free within the slot count; the 20-slot Advanced plan is 10 short of a 30-question panel	Checks run 150 to 600 a month below Enterprise
Otterly.ai	Low-cost entry at US$29 with a published top-up price	15 at US$29, 100 at US$189, 400 at US$489, from 1,000 on Enterprise	Prompt slot	Free; 100 extra slots cost US$99	Claude, Google AI Mode and Gemini are add-ons on every tier
Peec AI	Wide panels at shallow depth, up to 350 slots on Advanced	50 on Starter, 150 on Pro, 350 on Advanced; plan prices: Not found on reviewed page	Prompt slot, 3 models per plan	Free within the slot count	Custom prompt setup is listed on Enterprise
Profound	Enterprise teams that want nine answer engines and tailored tracking	50 unique prompts on the free trial, daily for seven days; Enterprise count: Not found on reviewed page	Custom on Enterprise	The pricing page describes the trial prompt set as fixed	Tailored tracking and nine answer engines are listed on Enterprise
Scrunch AI	Buyers who want funnel stages as a named tracked dimension	125 unique prompts on Core at US$250	Prompt slot, 5 licences	Free within the slot count	Core covers 1 country, 1 persona, 1 language and 4 LLMs
Semrush	Teams already on Semrush that want AI prompts inside the same licence	25 on the AI Visibility Toolkit, 50 on Semrush One Starter, up to 200 on Advanced	Prompt slot per licence	Free within the slot count	Page lists ChatGPT, Google AI Mode and Gemini; other engines: Not found on reviewed page
AthenaHQ	Deep sampling on a narrow panel, billed per AI response	300 credits free, 3,600 on Starter at US$295	Credit, where 1 credit equals 1 AI response	Free; every extra run costs a credit	A 30-question panel at 3 runs on 5 models burns 450 credits a cycle
Rankscale	Frequent rotation, because stored search terms are unlimited	1,200 credits at US$99, 5,500 at US$385, 12,000 at US$780	Credit, unlimited stored search terms	Free, because terms are unlimited	Credit burn scales with the engines selected per prompt
LLMrefs	Broad keyword coverage at one US$79 price	500 keywords on the single All in One plan at US$79	Keyword, with fan-out prompts per keyword	Free within 500 keywords; swapping a keyword replaces its generated prompts	Check the generated prompts against the buyer questions you wrote

Source: vendor pricing and product pages reviewed September 23, 2026, figures as published and not tested through paid accounts. “Not found on reviewed page” means the vendor’s public materials did not name that detail when reviewed on September 23, 2026. It is not proof that the capability is absent.

Slot billing and credit billing create opposite rotation incentives. On a slot plan, swapping a question costs nothing and adding a run costs a tier upgrade. On a credit plan, storing a question costs nothing and every extra run bills. Price the credit plans if you sample deeply and the slot plans if you hold a wide panel at shallow depth.

How to rotate the panel without breaking the comparison

A tracked prompt panel should be rotated only through a separate candidate layer, which held 12 questions in the Canlah AI self-audit, with the counted panel changed only at a declared boundary. Every question added or removed mid-window changes the denominator, so the cycles either side are not comparable even though a dashboard will plot them on one line.

The self-audit is the worked example. The 12 buyer-prompt candidates were probed as a separate sixth layer and disclosed on their own, so they would not alter the main funnel layers’ mention-rate basis. Those 12 candidates returned a 0.0 mention rate on both engines, which made them useful as a diagnosis and useless as a headline.

Three rotation rules hold. Retire a question only when the market changed, not when the score disappointed. Replace it from the same layer, so the quota survives. Record the date of every change, because a rotation nobody wrote down becomes an unexplained step in the trend six months later.

How to interpret five common panel results

Five panel results recur often enough to need a standard reading: a blank category layer, a wide range inside one layer, one blank engine, a score that moves after a panel edit and a comparison layer with no peers. Canlah AI met four of the five in its own September 7, 2026 self-audit.

Blank category layer, full verification layer: the engine knows the brand when asked by name and has no reason to raise it otherwise. This was Canlah AI’s own result, with 0.0 in L1 and 1.0 in L5. Read it as an evidence problem in the category layer, not a reputation problem.

Wide range inside one layer: one question returns a mention in every run while another returns none. The layer is not the unit of the fix. Inspect both questions, compare the sources cited for each and target the shelf question.

One engine visible, the other blank: the self-audit returned mentions on the OpenAI leg at L2 and L3 and none on the Gemini leg. Engines retrieve from different source sets, so the fix is a source problem in one corpus rather than a site-wide content problem.

Score improves after a panel edit: suspect the edit before crediting the work. Compare only the questions present in both cycles, and report that subset separately.

Comparison layer with no comparable peers: Canlah AI’s brand-comparison layer dropped from six questions to three because no same-tier-verified peer passed the check, so it fell back to category phrasing naming no rival. Record the shortfall rather than padding the layer.

What a tracked prompt set cannot prove

A tracked prompt set of any size cannot prove five things: stable share of voice, causation, future mentions, signed-in consumer answers or the accuracy of what engines repeat.

• It cannot prove a stable share of voice, because a panel samples a distribution rather than measuring a rank.
• It cannot prove causation between a content change and an answer change without repeated observations on a controlled timeline.
• It cannot guarantee future mentions, because model and index updates change retrieval independently of what the brand publishes.
• It cannot reproduce a signed-in consumer session if the runs used an API, so the two surfaces are compared separately.
• It cannot replace legal, factual or compliance review of what the engines repeat about the company.

The Canlah AI self-audit carries its own limits. The runs used a compressed sampling window with no time spacing between same-intent samples, so intraday variance is not represented. One of 234 samples failed on a read timeout. The result is a single-day baseline for one brand in one market.

When to move from a manual panel to a managed one

A spreadsheet panel is sufficient up to about 180 observations a cycle, which is 30 questions at three runs on two engines and one person’s day of collection and coding. The same panel across five engines and two languages is 900 observations, at which point collection drifts and the layer quotas erode. Past that point, the choice is a self-serve platform from the table above or a managed service.

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Every engagement starts from a locked buyer-query pool built on the five-layer quota above, and the client keeps the raw prompt, answer, citation list and timestamp archive. Its pricing page prices the GEO service per AI platform: Starter from $1,500 a month (ChatGPT, 4 buyer questions), Growth from $3,000 a month (adds Google AI Overviews, 8 buyer questions) and Full customised (adds Google AI Mode and Gemini, 16 buyer questions).

The main limitation is scope. Canlah AI runs the panel as a managed service and offers neither single sign-on nor API access; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. A team that wants to configure its own prompt tracker from day one should buy one of the platforms above.

Do not choose a measurement model from this article alone. Ask any provider, Canlah AI included, to run your frozen panel in your market and language and to hand back the raw answers before you commit.

Frequently asked questions

How many prompts should I track for AI search?

Thirty to forty tracked prompts cover one brand in one market and one language, split across five funnel layers with branded questions capped at about a third. Below 20 questions the layer quotas collapse and one question carries the score. Above 60 in one market the panel duplicates intent rather than adding coverage.

Which AI search prompts should I track first?

Track the unbranded main-shelf prompts first, because they are what a buyer types before any vendor is in mind. In the Canlah AI self-audit those eight questions produced the widest range of any layer, from no mention to a mention in every run. Add branded questions once the unbranded layers are filled.

Can I change my tracked prompts every month?

No. Every tracked prompt added or removed mid-window changes the denominator, which makes the cycles either side incomparable even though a dashboard will plot them on one line. Rotate a separate candidate layer instead, and move questions into the counted panel only at a declared boundary with the date recorded.

How many runs per prompt are enough?

Three runs per prompt per engine is the practical floor, because a single run records one sampled answer rather than a frequency. Sample deeper on the prompts that drive a decision, which is why credit-billed platforms such as AthenaHQ and Rankscale suit narrow panels.

Do branded prompts inflate an AI visibility score?

Yes, reliably. In the Canlah AI self-audit the seven branded questions returned a 1.0 mention rate on both engines while the 20 unbranded funnel questions ranged from 0.0 to 1.0. A branded-heavy panel reports a healthy number for a brand absent from every shortlist.

Does the AI visibility tool decide my prompt set, or do I?

The buyer writes the prompt set on most slot-billed platforms, but three reviewed pages limit that control. Profound’s pricing page describes the trial prompt set as fixed and lists tailored prompt tracking on Enterprise, Peec AI lists custom prompt setup on Enterprise and LLMrefs generates prompts from the keywords you track. Confirm current scope with the vendor before building a panel around it.

Method and sources

The prompt-allowance figures were taken from each vendor’s public pricing or product page on September 23, 2026 and record what those pages stated that day. No paid account, private demonstration or API was used, and no vendor reviewed this article before publication. The funnel quota, layer results and sampling figures come from a Canlah AI self-audit of canlah.ai executed on September 7, 2026, an internal run on one brand in one market, not independently audited.

• Ahrefs Brand Radar. Prompt counts per plan, check allowances, update frequency.
• Otterly.ai pricing. Prompt tiers, the US$99 add-on for 100 prompts, engine add-ons.
• Peec AI pricing. Prompt counts, three-model selection, Enterprise custom prompt setup.
• Profound pricing. Trial prompt count, trial prompt-set terms, Enterprise scope.
• Scrunch AI pricing. Core prompt count, funnel stages, country and language limits.
• Semrush AI Visibility. Prompt counts per plan, tracked engines, refresh cadence.
• AthenaHQ pricing. Credit allowances and the credit-to-response definition.
• Rankscale pricing. Credit allowances, unlimited search terms, per-engine burn.
• LLMrefs. Keyword tracking unit and auto-generated fan-out prompts.
• Canlah AI self-audit of canlah.ai, September 7, 2026, internal and not independently audited. Layer quotas, mention-rate ranges, question sources, sample counts, recorded failures.
• Canlah AI AI-info page. Locked buyer-query pool, sampling depth and evidence-archive scope.

See which buyer questions AI engines answer without you. Run the free AI visibility audit →

Related articles

• How to choose the right AI visibility tool
• AI visibility tool pricing in 2026
• How to evaluate a GEO vendor: the evidence checklist
• What are the best AI search monitoring tools?