How to Do Prompt Research for AI Search: ChatGPT, Perplexity, Google AI Mode

Quick answer

Canlah AI does prompt research by finding the questions buyers put to AI assistants, then pruning, routing and probing them in six written steps before any tracking panel is locked. For canlah.ai on September 7, 2026, the method cut a pool of 191 candidate queries to 39 probed prompts: three written rules pruned 34, four were quarantined and the probes returned 117 OpenAI responses, 119 Gemini responses and 70 rendered Google AI Overview runs in the Singapore market. Every survivor carries a funnel layer, a language and a brand state, because canlah.ai was mentioned in 100.0% of branded answers and 4.0% of non-branded answers in that run.

Canlah AI runs prompt research as the opening stage of a managed engagement and hands over the prune log next to the pool, so a client can read which candidates were removed and on which rule. The same six steps run by hand with Google autocomplete, a related-searches export and one afternoon.

Treat the first pool as directional, a draft rather than a panel, and freeze it only once the prune rules are written down. A pool that keeps changing cannot support a before-and-after comparison.

Disclosure: Canlah AI publishes this guide and operates the prompt research service described below. Details for Profound, Semrush, Ahrefs Brand Radar, AthenaHQ, HubSpot, Otterly.ai, Adobe and Google Search Console come from their public pages, reviewed on September 23, 2026, and were not tested through paid accounts. The Canlah AI figures come from a self-audit of canlah.ai run on September 7, 2026.

What does prompt research measure?

Prompt research measures six properties of every candidate question, and a candidate can pass one while failing another. A prompt is not a keyword with a question mark in front of it, because the engine answers in prose, names a shortlist and cites sources rather than returning ten links.

• Funnel layer: Whether the question asks what the category is, who is good at it, which subcategory fits, how two options compare or whether one named brand is credible.
• Brand state: Whether the question names the company. Branded questions measure recall while non-branded questions measure recommendation, so averaging the two produces a number no buyer experiences.
• Phrasing: The exact words a buyer uses, since an engine answers “who are the best AI search optimization agency providers in Singapore right now” differently from “ai search optimization agency”.
• Language and market: The language asked in and the country the answer is for, because the same category returns different shortlists per market.
• Engine fit: Whether ChatGPT, Gemini, Google AI Overviews and Google AI Mode answer it with a recommendation rather than a definition.
• Relevance: Whether the candidate belongs to the business at all, the rule that removes most harvested volume.

Keyword research ranks what a person types into a search box by monthly volume. Prompt research ranks what a person asks a model by how close the question sits to a buying decision.

Where the prompts come from: four discovery channels

Prompt research draws on four discovery channels, and three of them supplied all 191 candidates in the Canlah AI self-audit of September 7, 2026: 152 from long-tail search harvesting, 27 from layered question synthesis and 12 from real buyer prompts recorded in Google. By origin, 171 were real search strings, 14 were synthesized and six came from templates.

Channel one: search autocomplete and related searches. Seed the category term, the service term and the brand name, then harvest every completion and related search. It is the cheapest channel and the noisiest.

Channel two: sales, support and community transcripts. Read the last 50 inbound enquiries and pull the questions in the buyer’s own words. This channel carries no volume data and the best phrasing, since nobody writes a sales email in keyword syntax.

Channel three: layered question synthesis. Generate questions to fill a fixed funnel spine, one layer at a time, so the pool covers category, recommendation, subcategory, comparison and verification instead of clustering in one layer. This produced 27 of the 39 probed questions.

Channel four: vendor prompt-volume datasets. Profound Prompt Volumes describes keyword volumes, topic trends and intent signals pulled from conversations in ChatGPT, Gemini, Claude and Perplexity, and Semrush Prompt Research calls itself keyword research for AI with topic-level volume, difficulty and user intent. Both are panel estimates, and a method for converting a panel conversation into a query count was not found on either reviewed page on September 23, 2026.

How to do prompt research, step by step

Prompt research takes six steps: write the business profile, harvest wide, prune on written rules, route survivors into funnel layers, quarantine unverified names and lock the pool before probing. Each step is a decision taken once and then repeated unchanged for at least one quarter.

Step 1: Write the business profile before the first seed

Record the primary category, the canonical category names buyers use, the target markets and the buyer type in one paragraph. The Canlah AI profile named one primary category, four canonical categories and two target markets, and every prune rule referred back to it. A pool seeded without it cannot be pruned, because nothing defines off-topic.

Step 2: Harvest wide, from seeds you can defend

Take the category term, two or three service synonyms and the brand name as seeds, then collect autocomplete and related searches for each. Expect the brand seed to yield worst: seeding “canlah” returned “candelabra”, “candela” and “candelaria quezon”, and all eight brand-seeded completions were removed.

Step 3: Prune against three written rules

Apply the market, brand-drift and foreign-anchor rules before routing, and log the rule that removed each candidate. The three rules removed 34 of the 191 candidates, broken down in the prune table below.

Step 4: Route every survivor into a funnel layer

Assign each remaining question to one of five layers with a fixed quota, so the pool cannot pile up at the top of the funnel. The self-audit capped the layers at four category, eight recommendation, eight subcategory, six comparison and four verification questions, 30 slots in total, and 27 layered questions filled those slots, leaving three open. Record an empty slot as a gap rather than filling it with a weaker question.

Step 5: Quarantine anything that names an unverified competitor

Hold back any question naming a company you have not confirmed exists in that market. Four questions were quarantined on this rule, and 12 peer names were blocked before they reached a prompt.

Step 6: Lock the pool, then probe it three times per question

Fix the wording, the layer labels and the language, then run each question at least three times per engine. The self-audit ran all 39 questions three times through the OpenAI API for 117 responses, collected 119 Gemini responses and added 70 rendered Google AI Overview runs in the Singapore market. Any later edit creates a new baseline rather than a new cycle.

The prompt pool record: six fields to log on every candidate

A prompt pool record needs six fields per candidate, from the verbatim prompt text to the rule that kept, removed or quarantined it. The table below is the spreadsheet schema used in the self-audit.

Prompt pool schema, six fields per candidate, as used in the Canlah AI self-audit of September 7, 2026.

Field	What to record	Decision it supports
Prompt text	The question verbatim, in the buyer’s own words	Keeps phrasing comparable across cycles and engines
Discovery channel	Autocomplete, related search, transcript, synthesis or vendor dataset	Shows how much of the pool is real search language rather than invention
Funnel layer	Category, recommendation, subcategory, comparison or verification	Stops the pool clustering at the awareness end, where no purchase happens
Brand state	Branded or non-branded	Separates the 100.0% branded mention rate from the 4.0% non-branded rate
Language and market	Answer language plus the country setting used	Tells a multi-market team which half of the funnel was measured
Decision and reason	Probed, pruned or quarantined, with the rule applied	Makes the pool auditable, so a client can challenge a removal

Source: Canlah AI self-audit export, September 7, 2026, covering one brand, one market and one date. The counts show the shape of a pool, not a benchmark for another category.

How to prune a candidate pool: three rules that removed 34 of 191 queries

Three prune rules removed 34 of the 191 candidates, and market mismatch accounted for 23 on its own. Autocomplete supplied 33 of the 34 removals and a related-searches export supplied the last, which is the expected cost of harvesting wide.

Prune and quarantine log, canlah.ai self-audit, September 7, 2026.

Rule	Candidates removed	What it caught	Why it matters for the panel
Market mismatch	23	Another country, such as “best generative engine optimization agency in india”	Another market returns another shortlist
Brand seed drift	8	The brand name resolved to an unrelated word, such as “candelabra” from “canlah”	Counts a branded question the brand cannot win
Foreign anchor	3	United States-specific context, removed before layer routing	Names US providers instead of the local set
Unverified peer name	4 quarantined	A company not yet confirmed in the target market	Measures a comparison local buyers do not make

Source: Canlah AI self-audit export, September 7, 2026. Quarantined means the question was held out of the probe rather than deleted, so it returns once the named company is verified; it is not a judgment that the question is irrelevant.

A relevance gate ran on the harvested candidates before these rules, combining vocabulary overlap with a model check, and it passed 158 while holding back eight. Run the gate before funnel routing, because routing an irrelevant question still produces a tidy-looking layer.

What Google Search Console contributes to prompt research

Google Search Console contributes seed queries to prompt research, while its generative AI performance report lists four dimensions on Google’s help page: pages, countries, dates and devices. A query or prompt dimension was not found on the reviewed page. The same page describes impressions as the only metric, reports AI Overviews and AI Mode as one combined figure and caps the export at 1,000 rows. The report rolled out worldwide on August 31, 2026.

The classic Performance report supplies the other half of the job. It still returns the search queries that reach the site, and those are legitimate seeds for channels one and three. Pull the top few hundred query rows, keep the ones phrased as questions, then rewrite the rest into sentences a buyer would say aloud to an assistant.

Treat that rewrite as synthesis rather than evidence: a Search Console query proves somebody searched, while the prompt derived from it stays a hypothesis until an engine answers it. In the Canlah AI self-audit, 13 of the 27 layered questions, or 48.1%, traced back to a real search string against an internal target of 60.0%, and the shortfall was logged in the run record rather than hidden.

AEO keyword research tools compared

Eight tools support AEO keyword research on their public pages. Canlah AI ranks first in this comparison on one narrow axis, whether the buyer receives the prune log alongside the pool. Published entry costs for the others run from free at Google Search Console to US$199 a month for the Ahrefs Brand Radar index. Prices describe public positioning reviewed on September 23, 2026, not tested access.

AEO keyword research tools compared on prompt sources, entry cost and evidence, reviewed September 23, 2026.

#	Tool	Best for	Where the prompts come from	Entry cost	Main point to verify
1	Canlah AI	An auditable pool rather than a prompt list	Autocomplete, related searches, transcripts and layered synthesis, with every removal logged	Free AI visibility report within 48 hours; DeepAudit $99 for One read or $189 for Month read	Managed prompt research with no seed-to-volume tool; write the prune log into the contract
2	Profound Prompt Volumes	Demand-led prompt selection	Panel of real user conversations, described as more than 1.9 billion prompts in Profound’s own comparison	Free trial, five Prompt Volumes searches	Dataset size is vendor-reported; full access is an Enterprise quote
3	Semrush AI Visibility Toolkit	SEO teams wanting familiar metrics	Prompt database of 213 million prompts across ChatGPT, Gemini, AI Overviews and AI Mode	US$99 a month; free trial not found on reviewed page	DeepSeek, Qwen and Doubao not found on reviewed page
4	Ahrefs Brand Radar	A category benchmark before writing prompts	Index of more than 454 million monthly prompts, 312.8 million in AI Overviews	US$199 a month for the index	Ahrefs notes the index favours brands that already have search demand
5	AthenaHQ	Published pricing with volume estimates	Prompt volume estimation on both paid tiers	US$25 free credit or 300 AI responses	Ask how volumes are estimated; the machine-learned label is Profound’s
6	HubSpot AEO	Teams with CRM context already in place	8 to 15 suggested prompts generated from brand, competitor, product and ICP details	Bundled with Marketing Hub tiers	External demand source not found on reviewed page
7	Otterly.ai	Small budgets running a first pool	Prompt Research sits in the product menu beside AI Search Analytics	US$29 a month, 15 prompts	15 prompts give three per funnel layer in one language
8	Google Search Console	Free seeds from real search demand	Site queries in the Performance report	Free	Prompt text not found on reviewed page

Source: vendor product, pricing and help pages reviewed September 23, 2026. “Not found on reviewed page” means the vendor’s public materials did not name that capability when reviewed on that date. It is not proof that the capability is absent.

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. Its genuine limitation in this comparison is shape rather than method: Canlah AI runs prompt research as a managed service and offers no dashboard that takes a seed and returns a volume estimate. After three months of the GEO service, clients can run the work themselves with Canlah All-Stars, but a team that wants a seed-to-volume prompt tool should choose one of the six vendor platforms in rows 2 to 7.

Adobe completed its purchase of Semrush on April 28, 2026 and cited nearly 300 million AI search prompts in its June 17, 2026 Adobe Brand Visibility announcement, so ask Semrush which database a quote draws on.

How AEO keywords differ from SEO keywords

An AEO keyword is a full question with a buyer, a market and a decision attached, while an SEO keyword is a phrase ranked by monthly volume. The Canlah AI pool tagged 164 harvested candidates by intent, 111 awareness, 17 comparison and 36 decision, and almost every awareness row was a short autocomplete phrase such as “ai search optimization agency”. The probed questions built from those seeds read as sentences instead: “Who are the best AI search optimization agency providers in Singapore right now?” is the form an engine answers with a shortlist.

Reuse the keyword research you already own. Cluster the export by topic, keep the clusters that map to a buying decision, then write one question per cluster per funnel layer. Expect heavy compression: the Canlah AI pool went from 191 candidates to 39 probed prompts, about one in five.

What prompt research cannot prove

Prompt research decides which questions to probe, and five limits apply whatever the pool size.

• It cannot prove how many people asked a question, because every published prompt-volume figure is a panel estimate rather than an engine count.
• It cannot prove a question is commercially valuable, since funnel layer describes intent and not revenue.
• It cannot guarantee a comparable answer after a model or index update.
• It cannot substitute for a probe, because a pool is a hypothesis until the engines have answered it three times.
• It cannot fix a category problem, as a brand absent from every non-branded answer stays absent whichever questions are chosen.

When to lock the panel and start measuring

A prompt pool is ready to lock as a measurement panel once three conditions hold: the prune rules are written, every funnel layer has been routed against its quota and no candidate is still waiting on a peer verification. From that point the pool stops being research and becomes the panel.

Size the panel to the tier that will run it. Canlah AI’s published tiers track 4, 8 or 16 buyer intents, each re-measured monthly on the AI platforms in the tier (Starter covers ChatGPT, Growth adds Google AI Overviews, Full adds Google AI Mode and Gemini), so any pool, whether 400 questions or the 39 in the self-audit, has to be grouped into that many buyer intents before it is locked, and a second market needs its own set.

Do not buy on a prompt list alone. Ask any provider, Canlah AI included, for a dated sample before signing: the candidate pool, the prune log, the layer counts and one probe with prompt, engine, market, complete answer and cited sources. A prompt list handed over without its removals is a conclusion nobody can re-verify.

Frequently asked questions

How do I find the prompts customers use in ChatGPT?

To find the prompts customers use in ChatGPT, harvest Google autocomplete and related searches for the category, read the last 50 inbound sales enquiries for phrasing, then write one question per funnel layer and probe each three times. Profound Prompt Volumes and Semrush Prompt Research add estimated demand on top, and both are panel estimates rather than engine counts.

Can I extract prompts from Google Search Console?

No. Google’s help page lists four dimensions for the Search Console generative AI performance report, namely pages, countries, dates and devices, and caps the export at 1,000 rows. The classic Performance report does return search queries, and those become usable prompt seeds once rewritten into sentences a buyer would put to an assistant.

What are AEO keywords?

AEO keywords are the full questions an answer engine responds to, carrying a funnel layer, a market and a language rather than a monthly volume. They are sentences rather than phrases, and the useful ones sit at the recommendation, comparison and verification layers where a shortlist is formed.

Is prompt research the same as keyword research for AI?

No. Prompt research and keyword research for AI share vocabulary but not ranking logic. Keyword research sorts phrases by search volume, while prompt research sorts questions by distance from a buying decision, then records the language, market and brand state that decide how an engine answers. Semrush frames its own report as keyword research for AI, which describes the interface rather than the intent model underneath.

How many prompts should an AI visibility panel track?

An AI visibility panel should track only as many prompts as the budget can probe at least three times per engine on a fixed cadence. Canlah AI tiers run 4, 8 or 16 locked buyer intents and Otterly.ai’s US$29 tier covers 15. The canlah.ai self-audit probed 39 prompts cut from 191 candidates, enough for five funnel layers in one market.

Method and sources

This guide was published on September 23, 2026. Tool capabilities and prices were read from public product, pricing and documentation pages on that date, and were not tested through paid accounts. The Canlah AI figures come from one self-audit of one brand in one market, run on September 7, 2026, and are a worked example of the record format rather than a benchmark.

• Canlah AI self-audit of canlah.ai, internal export, September 7, 2026. Candidate pool of 191, prune and quarantine log, layer quotas, channel split, real-query share and probe counts.
• Canlah AI measurement methodology and pricing. Locked buyer-query pool, free AI visibility report, DeepAudit prices ($99 One read, $189 Month read) and buyer intents by tier.
• Profound Prompt Volumes and pricing. Dataset, named engines and trial limits.
• Semrush AI Visibility Toolkit and data sources. Price, Prompt Research description, daily query limits and database size.
• Ahrefs Brand Radar. Index size by engine, index price, custom prompts and the search-demand caveat.
• AthenaHQ pricing. Free credit, credit model and prompt volume estimation.
• Profound vs. AthenaHQ. Competitor statements about the dataset and AthenaHQ’s estimates, treated as vendor claims.
• HubSpot AEO in Marketing Hub. Suggested prompt counts and CRM-informed generation.
• Otterly.ai pricing. Tier prices, prompt volumes and the Prompt Research menu entry.
• Adobe Brand Visibility, June 17, 2026. Prompt database size and named platforms.
• Google Search Console Help, generative AI performance report. Dimensions, impressions-only scope, row limit and the August 31, 2026 rollout.

See where your brand stands in AI answers. Run the free AI visibility audit →

Related articles

• How to check if AI search recommends your brand
• How to choose an AI visibility tool
• AI visibility tool pricing in 2026
• How to track your brand in Chinese AI search