Methodology
Prompt research is not keyword picking. It is a loop.
In AI search the smallest unit of demand is no longer a keyword. No platform publishes query-level data, and the answers change daily — re-run the same prompt a minute later and it changes again. This page is the executable version: where to find prompts, how to estimate “volume” and commercial value, how to design the mix in a prompts map, how to measure in a way that holds up statistically, and how to feed the answers back into new scenarios, pain points and segments.
- For
- Brands · agencies · content and comms teams
- Version
- v1.0
- Reviewed
- 2026-11-06
- Engines tracked
- ChatGPT · Google AI Mode · Google AI Overviews · Gemini
In one paragraph
There is no AI search volume to look up, so build a relative demand index from four proxy types and rank at topic-cluster level. Score commercial value with a multiplicative four-factor model so high-volume, zero-intent questions drop out on their own. Keep branded prompts at 10–15% and report them separately — the median brand scores 95% on branded prompts and 7% on unbranded ones, so the mix, not your performance, decides the headline. Measure on a 10–28 day rolling window, never a single day. Then mine the answers for the attributes, competitors and sub-queries that tell you which segment to open next.
95% / 7%
Median mention rate for the same brands on branded vs unbranded prompts. Mix branded prompts into one total and the score stops meaning anything.
~65%
Share of the sources cited in AI answers that change from one day to the next. A single-day snapshot is not a measurement, it is a draw.
≥ 7
Effective observations a single prompt needs before the standard error on brand mention rate drops below 0.10.
Key takeaways
- There is no AI “search volume”: no platform publishes query-level data and there is no Search Console equivalent. Every number on the market is a model estimate or a sample extrapolation — use a relative demand index with its sub-scores visible, aggregated to topic-cluster level.
- Branded prompts belong in their own group at 10–15% of the list. Median mention rate is 95% on branded prompts versus 7% on unbranded ones, and 98% of brands score higher on branded — so mix decides the headline number, not performance.
- Score value multiplicatively, not additively: PVS = 100 × (D/5)^0.30 × (I/5)^0.30 × (W/5)^0.20 × (E/5)^0.20. Any factor at zero drops the prompt out of the pool, which is exactly what you want for high-volume, zero-intent questions.
- Never conclude from a single day. Re-runs within the same minute overlap only 0.32–0.43 on cited sources — about 65% of sources change daily. A single prompt needs ≥7 effective observations, or a 10–28 day rolling window instead.
- Diagnose the gap type before acting: coverage, position, perception or evidence. For recommendation intent the answer cites third-party roundups, so editing your own landing page is the wrong move.
Why keyword research does not transfer
In SEO, the thing you research and the thing you optimise are the same object: the keyword a user typed. You can see it in Search Console, buy its volume from a tool, and build a page for it. AI search splits that into three layers — and the layer you can see is not the layer where the user speaks.
The user’s prompt
Not observableWhat a person actually typed. No platform publishes it and there is no Search Console equivalent, so demand strength here can only be estimated through proxies.
Fan-out queries the model issues
Partly observableThe sub-queries a model sends out to answer the prompt. Some of them surface as question-shaped queries in Search Console, and some can be collected directly. This is the layer where the competitive field is actually decided.
The answer surface
Fully observableWhich brands get named, which URLs get cited, which product cards appear. The only layer you can verify — and the only place a result can be confirmed.
LLM providers currently offer no native monitoring tool equivalent to Search Console. Basic metrics such as which search queries were used, and their query volume, are therefore no longer directly observable in the GEO ecosystem.
That sentence is the premise of everything below. Any product promising you an “AI search volume” is giving you a model estimate or a sample extrapolation, not a platform count. Accept it, and the method becomes defensible.
Eight stages, two feedback loops
Run prompt research as a one-off project and it expires in three months: the answers move, the competitors move, and so does the category boundary as the model sees it. It should be a pipeline with fixed deliverables and fixed decision rules.
- 01
Scope
Which categories, markets and buying moments are in play this quarter — and which are explicitly out.
- 02
Collect
A candidate table of real phrasings, each tagged with source, evidence tier, original location and entry date.
- 03
Expand
Each core prompt stress-tested against the eight fan-out variant types, plus an intent label.
- 04
Score
A demand index per topic cluster and a four-factor value score (PVS) per prompt.
- 05
Assemble
A prompts map with an explicit branded / unbranded / competitor mix, a filled coverage matrix, and a frozen core set.
- 06
Measure
Rolling-window brand-level metrics with per-engine baselines. Daily numbers are alerts, not findings.
- 07
Act
Each action bound to one prompt cluster, one gap type and an observation window of at least four weeks.
- 08
Feed back
New phrasings, new competitors and new segments mined out of the answers themselves.
Monthly loop: 08 → 02
New phrasings and newly co-occurring competitors found in the answers go back into collection, so the candidate table keeps growing from evidence rather than from brainstorming.
Quarterly loop: 08 → 01
A segment that clears all three tests (its own sub-query theme, its own competitive set, its own cited sources) goes back into scope as a first-class segment.
Eight source types: real phrasings first, volume later
The most common failure is writing prompts in a meeting room: the team imagines how users “would” ask, and the whole pool ends up in marketing language nobody actually uses. Sources come in three evidence tiers. Use all three — but never treat tier C output as evidence of demand. It exists to fill coverage.
A First-hand evidence
Left behind by a real user or a real platform.
B Derived inference
Inferred from an adjacent signal.
C Model-generated
Another model’s guess. Use it for coverage, never as evidence of demand.
| Source | Tier | What it tells you | Known bias |
|---|---|---|---|
| Question-shaped queries in Search Console | A | Long, constraint-carrying full sentences. A sizeable share are grounding queries issued while an AI composed an answer — real L2 residue. | Google surfaces only; nothing appears until an impression has already fired. |
| Real fan-out queries | A | The sub-queries a model actually issued for a question, their frequency, and the brands that showed up alongside them. | Needs collection capability; a sample, never the full population. |
| Support tickets and pre-sales chats | A | Pain points in the customer’s own words, blockers in the decision, and who they are comparing you against. | Skewed toward people who already converted; the awareness-stage question is missing. |
| Site search and return reasons | A | Attribute-level demand — size, material, fit, use case. Closest thing you have to the matching conditions on a product card. | Only covers people who already reached your site. |
| Community threads (Reddit, Xiaohongshu, niche forums) | A | How people ask about reputation, and the real reasons a product gets recommended. These pages are themselves frequently cited by models. | Audience skews technical and young; high noise. |
| People Also Ask, autocomplete, keyword volume | B | Relative heat across topics, and the skeleton of a question. | Biased toward short keywords; systematically understates conversational long prompts. |
| Competitor answer mining | B | Which phrasings name your competitors and who gets cited there — this hands you the attack surface directly. | Reflects only what exists today; says nothing about demand nobody covers yet. |
| LLM-generated candidates | C | Fast coverage of the long tail and of other languages; fills blank rows in the coverage matrix. | Produces sentences that look like users, not sentences users said. |
The only deliverable of this stage is a candidate table with source tags: every prompt records its source, evidence tier, original location and entry date. Without those four fields none of the later scores can be audited.
“What is this prompt’s search volume?” — the honest answer
Nobody knows. The platforms do not publish it and no third party has the full population. Every “AI search volume” or “prompt volume” on the market falls into one of four proxy types, each with its own systematic bias:
| Proxy type | How it is produced | Known bias | Verifiable? | Good for answering |
|---|---|---|---|---|
| Search-derived | Models Google keyword volume plus People Also Ask into an “AI-adjusted volume”. | Short, informational keywords; understates long conversational prompts. | High — you can recompute it | Relative ranking between topics |
| Panel-derived | Buys de-identified real AI conversations from a user panel. | Panel demographics and geography; cannot be independently verified. | Low | What phrasings look like, and how intent is distributed |
| Machine-side residue | Frequency of the sub-queries models issue, plus question-shaped queries in Search Console. | Only covers what already fired. | High — the platform emits it | How wide the competitive field is (L2) |
| First-party demand | Tickets, site search, sales calls, return reasons. | Skewed toward people who already arrived or already bought. | Highest | Pain points and the constraints that matter |
What to do: stop chasing a volume, build a demand index
Normalise each of the four proxy types to 0–100, then weight them by verifiability and sum. Three disciplines make it work:
Do not call it “search volume”
Call it a Demand Index. Change the name and the team stops using it for revenue forecasts.
Keep the sub-scores visible
An index of 72 should still show whether it is “search-derived high, first-party zero” (an industry buzzword your customers do not actually use) or “first-party high, search-derived zero” (a real pain point that has not yet become public search behaviour — usually the cheapest thing worth taking).
Aggregate to the topic cluster before ranking
A volume estimate for one prompt is extremely noisy. Only at cluster level (5–10 prompts) is it stable enough to sort on.
Query fan-out: one question becomes many
A user asks one thing, and the model does not simply search that sentence. It rewrites, generalises, specifies and follows up, issuing a batch of sub-queries and synthesising an answer from what returns. Which means: you are not competing with that one sentence, you are competing with the whole cluster it expands into.
Google’s related patents sort these synthetic queries into eight types (original names below). One caveat matters: Google has never confirmed that any patent is a live implementation — officially it says only that it “issues a large number of queries”, and it never publishes how many sub-queries a prompt produces. Every third-party “fan-out tool” is another model simulating it. So the eight types are a stress-test checklist, not a reconstruction of the algorithm.
Equivalent
The same question in different words. If a rewrite loses you, your page is matching a phrase rather than a meaning.
Follow-up
The next turn of the conversation. Answering turn one and not turn two hands the second answer to someone else.
Generalization
The broader question the prompt sits inside. Tests whether you exist at category level at all.
Specification
The same question with a constraint added — budget, persona, use case. Constraints change the brand set that gets named.
Canonicalization
The standard form of a colloquial or misspelled query. Needs a canonical definition passage somewhere you own.
Language Translation
The same question in another language. The sources cited change completely — translated conclusions do not transfer.
Entailment
What the question implies but does not say. “Is it safe for sensitive skin” hides inside “recommend a sunscreen”.
Clarification
The disambiguating query a model issues when the prompt is ambiguous. Entity confusion shows up here first.
Figure 3. The eight variant types are drawn from Google’s published patent text — Thematic Search (US12158907B1), Search with Stateful Chat (US20240289407A1) and others — as compiled by iPullRank. What matters in practice: rewrite your core prompt each of these eight ways and see which rewrites you disappear from. That type is your content gap.
Two hard rules
One page per cluster, not one page per variant
Building a separate page for every fan-out variant is exactly what Google’s scaled content abuse policy describes: generating many pages for search variations primarily to manipulate rankings or AI answers. The working approach is a single authoritative page that can be sourced from in all three directions at once — a standard definition passage, a comparable-parameters table, and the concrete constraints (price band, persona, use case) all on the same page.
The follow-up is the next battlefield
Follow-up and Entailment decide the second turn. After “which brand should I buy”, the next line is usually “any allergens in it?” or “safe during pregnancy?”. Fail that turn and the model swaps in a different brand to answer it — so those phrasings belong in the pool as prompts of their own, not as appendages to the first one.
Collect the sub-queries models actually issue: query fan-out analysis
The intent ladder: classify by what the answer looks like, not by the SEO taxonomy
Porting informational / navigational / transactional over is useless, because it does not tell you what to prepare. In AI search, intent decides the shape of the answer, the shape decides who gets cited, and who gets cited decides where you need to show up. Sorted along that chain, six intents have six completely different ways to win.
| Intent | Typical phrasing | Answer shape | What the model mostly cites | How you win |
|---|---|---|---|---|
| ① Definition | What is X / how is it different from Y | An explanatory passage, often with no retrieval at all | Encyclopaedias, authoritative glossaries, trade press | A definition passage that can be lifted whole, plus an entity page. You are competing to be the term’s explanation of record. |
| ② How-to | How do I do X / how do I set it up | Numbered steps | Documentation sites, tutorials, official help centres | Step-structured docs where each step stands alone and states its prerequisites. |
| ③ Recommendation | Best X / which brand should I buy | A brand list, usually 3–7 names | Third-party roundups and reviews — not your own site | Get into the roundups the model already cites. Your own site cannot win this category. |
| ④ Comparison | A vs B / alternatives to A | A comparison table or a point-by-point contrast | Comparison articles, alternatives lists, community threads | Your own comparison pages — including the dimensions where you lose — plus third-party review coverage. |
| ⑤ Validation | Is X any good / what do people say / any catches | A reputation summary with pros and cons | Review sites, Reddit and Xiaohongshu, complaint and refund threads | Genuine reputation work. This is a communications problem, not an SEO problem. |
| ⑥ Transaction | How much / where to buy / is it in stock | Product cards with prices | Structured product data, product feeds, retailer sites | Field completeness, readable price and stock, correct category. |
Category ③ is the most frequently misdiagnosed. A team finds it is absent from “best X” and reflexively rewrites its own landing page — while for that intent the model cites almost nothing but third-party roundups. Get this chain wrong and a whole quarter of effort lands nowhere.
Prompts map: the design rules
By now the candidate pool may hold hundreds or thousands of prompts. Assembling the list decides three things: what goes in, in what proportion, and when it may change. This is the easiest stage to fool yourself at — because how you mix the list is how the score comes out.
7.1 Branded prompts: report separately, never fold into a total
This is the single biggest trap in the industry, and the data on it is hard. Across a tracking sample of 3,989 brands and roughly 23 million answers (the 30 days to 2026-10-15, Searchable’s own data), median brand mention rate was:
Sample: 3,989 brands, each with at least 30 branded and 30 unbranded answers.
In the same dataset, 98% of brands scored higher on branded prompts. Put differently: adding one more branded prompt to the list almost always raises the total while your real position has not moved at all.
The consequence of the mix can be calculated directly. Take a brand mentioned in 95% of branded answers and 10% of category answers. Same brand, same week, three lists, three conclusions:
Worked example from Carlos Barraza’s public write-up (illustrative figures). The conclusion does not depend on the specific numbers: the bulk of a visibility score is decided by the mix of the list, not by your actual performance.
7.2 Three-way mix: industry ranges and our recommendation
| Category | HubSpot | Conductor | This guide | The question it answers |
|---|---|---|---|---|
| Unbranded (category / scenario / problem) | 50–60% | ≈ 75% | 55–65% | When a stranger asks about this, are you in the answer at all |
| Competitor-facing (A vs B, alternatives to X) | 20–30% | (counted as unbranded) | 20–30% | When you are put up for comparison, are you treated as a credible option |
| Branded | 10–20% | ≤ 25% (ceiling) | 10–15% | Is the model’s description of you accurate, and is it negative |
The gap between the two sources is mostly about where competitor prompts belong: Conductor counts them as unbranded, which produces 75/25. Converted to a common caliber, both agree that branded should stay at or under a quarter. Ecommerce brands can push competitor prompts down to 20% and move transactional intent onto the unbranded side, because product-card competition happens on category and scenario phrasings.
7.3 Coverage matrix: all six dimensions filled
The mix can be right and the whole list still be concentrated in one intent or one persona. The second test when assembling a list is that no row in the coverage matrix is blank:
Topic cluster
3–5 priority topics to start, 5–10 prompts each; 10–30 topics at enterprise scale. Within a topic, collapse synonymous rewrites into 1–2 representative sentences — do not pad with 20 paraphrases.
Intent
At least one prompt for each of the six intents above. Recommendation and comparison should make up most of the unbranded side; give definition prompts a small share — they rarely produce a result.
Persona and constraint
The same question once per budget, company size, region and use case. Constraints directly change which brands get named — this is a variable unique to AI search, and it must be covered.
Market and language
Full coverage in your main markets, a reduced subset in secondary ones. The same sentence cites entirely different sources in another language — you cannot reuse a conclusion by translating it.
Engine
Run the same list on every monitored engine, but set the baseline per engine (see the measurement section). One threshold across engines is not a comparison.
Bound asset
Name the page that *should* be cited for each intent cluster. If you cannot name one, that is a content gap, not a monitoring item.
7.4 Scale ladder and freeze discipline
| Stage | Prompts | Topics | Frozen / exploratory | Change rule |
|---|---|---|---|---|
| Starting out (month 0–1) | 25–50 | 3–5 | 100% frozen | Collect data only for the first month; change nothing. |
| Growing (month 2–6) | 150–300 | 5–12 | 70 / 30 | Quarterly change ≤ 20%, added in dated batches. |
| Enterprise / multi-market | 500+ | 10–30 | 70 / 30 per market | An independent frozen set per market; never merge reports across markets. |
The frozen set is the comparison baseline
Not one character changes inside a quarter — punctuation included. A rewritten prompt is a new prompt in the data; its history does not carry over, and the chart needs a break marker.
The exploratory set absorbs the trial and error
New phrasings, new segments and new markets enter there first. Two consecutive months of demonstrated volume and intent earn promotion into the frozen set.
Retirement rule
A prompt that has triggered no decision and no action for three straight months should be retired. It is not monitoring anything; it is producing noise.
7.5 Writing rules
- Write whole questions, not keywords
- Prompts in AI search average 70–80 words; Google searches run 3–4 (SimilarWeb data, via iPullRank). A list made of three-word phrases like “best sunscreen” is not measuring a real scenario.
- Put the constraints in the sentence
- Budget, scale, region, use case, persona — these are what the model actually filters brands on. “A facial sunscreen under 300 yuan for sensitive skin” and “recommend a sunscreen” are two different prompts with two different brand sets in the answer.
- No bare follow-ups
- “Which one is cheaper?” on its own is useless in a list because it has no context. Write it out as a standalone question carrying the full background.
- Collapse synonymous rewrites
- Keep 1–2 representative sentences per intent. Twenty paraphrases add no coverage; they just weight one intent several times over in the total.
- One intent per sentence
- “How does your pricing compare to X, and which integrations do you support?” must be split in two. Combined, the answer shape is mixed and nothing can be attributed.
Commercial value: four factors, multiplicative
“High volume” is not the same as “worth doing”. AI search is full of high-volume, zero-intent questions (“what is GEO”) and of low-volume, high-value ones (“an ERP under 50k that handles cross-border settlement”). Score additively and volume drags the zero-intent questions up; score multiplicatively and any factor at zero takes the prompt out of the pool — which is exactly the behaviour you want.
PVS = 100 × (D/5)0.30 × (I/5)0.30 × (W/5)0.20 × (E/5)0.20
- D Demand strength (1–5)
- The demand index from the previous section, taken at topic-cluster level, never per prompt.
- I Intent purity (1–5)
- How close the phrasing sits to deciding who to buy from. Definition prompts score 1; recommendation and comparison 4; transaction 5.
- W Winnability (1–5)
- Whether you are already mentioned, how concentrated the citations are for that phrasing (more concentrated is harder), how deep the incumbent’s moat runs, and whether you already own a citable asset.
- E Unit economics (1–5)
- AOV × margin × repeat rate for what the phrasing actually sells. For B2B, deal size and sales-cycle length.
Any factor scored 0 → out of the pool, no score
Worked examples
“A facial sunscreen under 300 yuan for sensitive skin”
D=4 I=4 W=3 E=4
PVS ≈ 76
Lower demand, far higher intent and value.
“What is generative engine optimization”
D=5 I=1 W=4 E=1
PVS ≈ 43
The higher-volume prompt ranks lower. That is the intended behaviour.
Score a prompt
Set the four factors and read the score. Multiplicative, so a single zero takes the prompt out of the pool.
PVS
76
High priority
Scored — now split the actions on two axes
PVS only sorts; it does not tell you what to do. Cross it with current visibility and the four quadrants call for four completely different kinds of investment:
High value · visible
Defend
Into the frozen set for routine monitoring; watch for slipping position and for cited sources being swapped out. The risk here is being squeezed out slowly with nobody noticing.
High value · not visible
Attack
This quarter’s main front. Start by finding out who gets cited for those phrasings — the attack surface is usually in third-party sources, not on your own site.
Low value · visible
Harvest
Hold position, add nothing. Do not be seduced by the high visibility number; it contributes no result.
Low value · not visible
Park
Leave in the exploratory set on low-frequency observation. Some appreciate as the category evolves — emerging scenarios usually show up here first.
Measurement validity: why today’s number is not a finding
This is the one part of prompt research with rigorous public research behind it, and the part most often skipped. A University of St. Gallen study from April 2026 did two things: ran the same prompt set daily for 45–46 days, and re-ran the same prompt up to ten times within a single day — the second of which separates the model’s own randomness from genuine external ranking change.
The conclusion is blunt: re-run in the same minute, and the overlap in cited sources is about as low as it is a day apart.
| Measurement | Citation-source overlap (Jaccard) | Brand-set overlap (Jaccard) | What it means |
|---|---|---|---|
| Between consecutive days | 0.34 – 0.42 | 0.45 – 0.59 | About 65% of cited sources change every day |
| Re-runs within the same day (≤10) | 0.32 – 0.43 | 0.33 – 0.48 | The variance comes mostly from model randomness, not from ranking change |
Source: arXiv:2604.07585 — 4 industries × 8 prompts × 4 engines (the study tested ChatGPT, Gemini, Google AI Mode and Perplexity), Swiss German-language market, 2026-01-24 to 2026-03-25. Brand figures cover only industries with ≥70% brand recognition.
Four operating rules that follow
- 01
At least 7 effective observations per prompt
Bootstrap convergence analysis puts the standard error on brand mention rate below 0.10 at n=7 (95% CI ±0.158) and below 0.08 at n=8; citation-source coverage needs n≥8. In practice, if each prompt runs once per engine per day, you must accumulate observations across days in a rolling window rather than compare single-day figures.
- 02
A 10–28 day rolling window, never a single day
The same study’s rolling-window analysis puts the standard error on a single brand’s mention rate below 0.10 at d=10 days and below 0.05 at d=24. Daily reports are for spotting anomalies, not for judging trend. A one-week window is still too noisy at brand level.
- 03
Brand level for KPIs, URL level for diagnosis
Day-to-day stability of the brand set (0.45–0.59) is clearly higher than that of cited sources (0.34–0.42). So the metric you report should be brand mention and ordering, with “which URL got cited” demoted to an attribution tool — watching a single URL rise and fall is mostly reading noise.
- 04
Set the baseline per engine
Citation concentration (Gini) averaged 0.715, but engines differ markedly: in that study Google AI Mode was the most concentrated (0.782) and Perplexity the most dispersed (0.671). In a concentrated engine a handful of domains absorb most citations — the same “20% visibility” means two very different degrees of difficulty. The same study also recorded 57.8% of ChatGPT runs carrying no citations at all (definition questions do not trigger retrieval), so missing citations do not mean missing mentions.
Definitions and formulas for every metric: AI visibility metrics
From gap to action: diagnose the type first
“We are not visible on this prompt” is four completely different diseases with four completely different cures. Misdiagnose the type and the investment lands nowhere.
Gap A
Coverage gap — never mentioned at all
Signal: Your mention rate for that phrasing sits near zero while five other brands fill the answer.
Cure: First check whether the model has any entity knowledge of you at all — can it answer a branded prompt correctly? If the entity does not hold up, producing content is futile; fix the entity and the structured facts first.
Gap B
Position gap — mentioned, but late
Signal: Mention rate is fine, but average position is low, or you only appear at the tail of long answers.
Cure: Supply comparable parameters — price band, who it suits, concrete specs. Models order on attributes they can align; a vague brand story does not sort to the front.
Gap C
Perception gap — mentioned wrongly or negatively
Signal: Branded prompts return distorted descriptions, the wrong category, or negative framing.
Cure: A communications problem. When a model summarises reputation it cites review sites and communities, not your own statements. Do not hand this one to the SEO team.
Gap D
Evidence gap — competitors cited, you are not
Signal: You get mentioned occasionally, but every source in the answer is a roundup or review your competitors are in.
Cure: Show up in those sources. List the domains cited most for that phrasing and sort them by whether you can realistically get in — this is the only approach that works for recommendation intent.
Every action must be bound to one prompt cluster and one observation window of at least four weeks. An unbound action cannot be verified, and next quarter it turns into “it probably helped”.
Cadence: who looks at what, and how often
| Period | What to watch | What may change | Forbidden |
|---|---|---|---|
| Daily | Anomaly alerts: a whole engine going dark, or a high-value prompt reading zero three days running. | Nothing | Drawing conclusions from a single day; editing prompts |
| Weekly | Direction of the rolling 7-day window; whether a competitor is rising across a whole intent cluster. | Exploratory set additions and removals; action scheduling | Touching the frozen set; changing report calibers |
| Monthly | Brand-level metrics on the rolling 28-day window; new phrasings and new competitors from answer mining. | Promotions and demotions; new prompts in dated batches | Deleting a prompt that has history |
| Quarterly | Recompute PVS across the whole list; redistribute the four quadrants; test whether a segment holds up. | Frozen-set change ≤ 20%; new markets and segments; retire prompts idle for three months | Rebuilding the entire list at once (all history becomes incomparable) |
90 days
Day 1–30
Build the list and freeze it
Scope → collect from the eight source types → a 25–50 prompt starting list → freeze everything and collect data. Draw no conclusions this month; the point is to accumulate enough window length. Deliverables: a source-tagged prompt table plus the target page bound to each prompt.
Day 31–60
First judgement
The 28-day window is in: run the first four-quadrant split, queue 3–5 actions by gap type, each bound to a prompt cluster and a 4-week observation window. Start the exploratory set at the same time to fill blank rows in the intent and persona matrix.
Day 61–90
Verify and scale the list
Verify the first batch of actions by comparing the 28-day windows before and after — not single days. Scale to 150–300 prompts, start the answer-mining loop, and decide whether a new segment opens next quarter.
Let the answers tell you: new scenarios, pain points, segments
This is where the loop earns its keep, and the step most teams skip. The thousands of answers you collect are not just “was I mentioned” — they are a map of the market as the model sees it. Four ways to mine them:
Attribute mining → how the model classifies you
Count the qualifiers that appear right next to your brand in answer text. If the model keeps saying you “suit small shops on a tight budget”, that is the positioning it has assigned you — and it directly decides which phrasings recommend you and which skip you. To change the positioning, start with this word list. Those same words are your new scenario and pain-point candidates: write them up as prompts and put them in the exploratory set.
Co-occurrence mining → the real category boundary
The brands that appear alongside you in answers are the competitive set the model believes in, and it is often not the list in your own competitor analysis. An unexpected brand means the model has filed you under a different category — either bring that category’s phrasings into your list, or go fix the entity knowledge.
New sub-query frequency → early signal of an emerging scenario
Among the sub-queries models issue, the ones that did not exist last month and keep recurring this month are demand forming on the user side. They usually have no public search volume yet — which is precisely why they are the cheapest moment to enter.
Three tests before a segment counts
Not every high-scoring topic cluster deserves to become its own segment. All three must hold: ① it has its own sub-query theme (not a rewrite of an existing cluster); ② it has its own competitive set (the co-occurring brands differ clearly from the main list); ③ it has its own cited sources (the roundups and reviews in its answers are a different set). Miss one and it is just a variant of an existing cluster — leave it in the exploratory set.
Ten anti-patterns
Every one of these has a real case behind it. Print it and put it on the wall.
| Anti-pattern | Why it is wrong | Instead |
|---|---|---|
| 1 Branded prompts folded into one total | Median 95% vs 7% — the total is decided almost entirely by the mix | Report branded prompts as their own group; read only accuracy and sentiment there |
| 2 Comparing visibility while editing the list | The caliber drifts, so every rise and fall is a mix change | Frozen set untouched for the quarter; a rewrite is a new prompt, with a break marker on the chart |
| 3 Concluding from a single day’s number | Re-runs in the same minute overlap only 0.32–0.43 | A 10–28 day rolling window; daily only for alerts |
| 4 Simple-averaging daily rates | Record counts differ by day, so the weights are wrong | Read the totals row, or re-weight by record count |
| 5 Treating a missing day as zero | Not collected ≠ not mentioned; it drags the window average down | Exclude missing days from the denominator |
| 6 A page for every fan-out variant | Lands squarely in scaled content abuse | One authoritative entity page covering the whole cluster |
| 7 One threshold across engines | Citation concentration differs markedly (0.671–0.782) | An independent baseline per engine |
| 8 Watching a single URL rise and fall | URL level is the least stable — 65% of sources change daily | Brand level for KPIs, URL level for attribution only |
| 9 Editing your own site for recommendation intent | Those answers cite third-party roundups | List the cited domains first, then sort by whether you can get in |
| 10 Watching yourself and not the field | With no denominator you cannot tell whether 20% is good | Compare the competitive set on the same list and the same window |
Sources and calibers
Every figure on this page carries its original source and data period. Items drawn from vendor data are flagged — those cannot be independently verified, so keep the original wording and sample description when you quote them.
Measurement stability · every statistical claim
Schulte, Bleeker & Kaufmann, Don’t Measure Once: Measuring Visibility in AI Search (GEO), arXiv:2604.07585, 2026-04-10. 4 industries × 8 prompts × 4 engines — the study tested ChatGPT, Gemini, Google AI Mode and Perplexity — Swiss German-language market, 2026-01-24 to 2026-03-25. Brand figures cover only industries with ≥70% brand recognition.
arxiv.org/abs/2604.07585Branded vs unbranded mention rateVendor data
Searchable’s own tracking data: 3,989 brands, roughly 23 million answers, the 30 days to 2026-10-15. Median 95% vs 7%; 98% of brands scored higher on branded prompts. Vendor data.
searchable.comMix guidance 50–60 / 20–30 / 10–20
HubSpot, Which AI prompts should you track? Also the source for 25–50 prompts to start, 5–10 per topic, and ≤20% quarterly change.
blog.hubspot.comMix guidance 75 / 25 and the scale ladder
Conductor Academy, AI Prompt Tracking. Branded ≤25%; 10–30 topics; roughly 100 prompts per topic at enterprise scale.
conductor.comMix worked example 52.5% / 19% / 10%
Carlos Barraza, Your AI Visibility Score Is Only as Good as Your Prompt List. Illustrative figures.
barrazacarlos.comThe eight fan-out variant types and the patents
iPullRank, Expanding Queries with Fan-out. Compiled from US12158907B1 (Thematic Search), US20240289407A1 (Search with Stateful Chat) and WO2024064249A1. Google has never confirmed that any patent is a live implementation. The same piece cites SimilarWeb for the 70–80 words vs 3–4 words contrast.
ipullrank.comProduct-card rates 87% / 91%Vendor data
cloro.dev’s study of 3,312 shopping-intent prompts. Vendor data; treat as directional only.
cloro.devCommercial value of AI referral traffic
Adobe Analytics: generative-AI referral traffic up 693.4% year over year in the 2025 holiday season, with a 31% higher conversion rate for AI-referred visits.
news.adobe.comCounter-evidence (read it alongside)
Amsive’s analysis of 54 sites: no consistently significant conversion-rate difference between LLM referrals and organic search (4.87% vs 4.60%), and LLM traffic made up just 0.24% of sessions. Treating an AI conversion premium as established fact is risky.
Click behaviour
Pew Research Center: visits where a traditional result was clicked fell to 8% when an AI summary was present versus 15% when it was not; only 1% clicked a link inside the summary. 900 US adults, March 2025.
pewresearch.org
Where this maps onto GEOly
The methodology does not depend on any single tool, but each stage has a direct counterpart in GEOly:
Stage 02 · Collect
AI Research takes a domain and seed terms and returns prompts already grouped into topic clusters with intent labels, ready to import into monitoring. The free AI Prompt Generator does a lighter version of the same thing.
Stage 03 · Expand
Query fan-out analysis collects the sub-queries models actually issue, groups them into Demand Themes, and uses Query Breadth to compare how wide a retrieval surface each brand covers.
Stage 05–06 · Assemble and measure
Prompts & Query Analytics, topic-cluster analysis and the cross-engine matrix. Every prompt runs once a day on each of ChatGPT, Google AI Mode, Google AI Overviews and Gemini, aggregated on a weighted 30-day window — which is exactly how the rolling-window rule above gets implemented.
Stage 07 · Act
Citation source analysis (which domains get cited most for a phrasing), competitor co-occurrence, category whitespace and GEO Audit give your team the evidence to work from. Metric definitions live on the AI visibility metrics pages; the public source leaderboard is Cited.
Stage 08 · Feed back
Answer-level records with brand co-occurrence and category momentum — the inputs for deciding whether a new segment holds up.
GEOly tracks ChatGPT, Google AI Mode, Google AI Overviews and Gemini. Perplexity was retired on 2026-10-22 and Microsoft Copilot on 2026-11-03; Claude and Grok have never been tracked.
Run the loop on your own brand
Look up where your brand stands in AI answers today, then see which prompts and which cited sources decide it.
FAQ
- How do I find the search volume of an AI prompt?
- You cannot, in the sense that keyword volume exists. No AI platform publishes query-level data and there is no Search Console equivalent, so every “AI search volume” figure on the market is a model estimate or a sample extrapolation. The workable substitute is a relative Demand Index built from four proxy types — search-derived, panel-derived, machine-side residue and first-party demand — each normalised to 0–100 and weighted by how verifiable it is. Keep the sub-scores visible, aggregate to topic-cluster level before ranking, and never call the result a volume.
- What share of my tracked prompts should contain my brand name?
- 10–15%, and they must be reported as their own group. Branded prompts score far higher by construction: across 3,989 brands and roughly 23 million answers, the median mention rate was 95% on branded prompts versus 7% on unbranded ones, and 98% of brands scored higher on branded (Searchable’s own data, the 30 days to 2026-10-15). So adding one more branded prompt reliably lifts the total while nothing about your real position has changed. Read branded prompts for accuracy and sentiment — whether the model describes you correctly — not for visibility.
- What is query fan-out and how do I research it?
- Fan-out is the model rewriting one prompt into a batch of sub-queries — generalising, specifying, following up — then synthesising an answer from what comes back. So you are not competing with one sentence but with the whole cluster it expands into. Google’s related patents describe eight variant types (Equivalent, Follow-up, Generalization, Specification, Canonicalization, Language Translation, Entailment, Clarification), though Google has never confirmed any patent is a live implementation. Use the eight as a stress-test checklist: rewrite your core prompt each way and see which rewrites you disappear from. The type you vanish under tells you what kind of content gap you have.
- How many prompts should I track, and how often can I change the list?
- 25–50 across 3–5 topics to start, fully frozen for the first month; 150–300 across 5–12 topics once you are running, split 70% frozen core and 30% exploratory; 500+ across 10–30 topics at enterprise scale, with an independent frozen set per market. The frozen set does not change inside a quarter — punctuation included — because it is your comparison baseline. Quarterly change should stay at or under 20%, added in dated batches. A rewritten prompt is a new prompt in the data, so mark the break on the chart.
- Why should I not compare today’s AI visibility number with yesterday’s?
- Because most of the day-to-day movement is model randomness, not ranking change. A St. Gallen study (arXiv:2604.07585) re-ran the same prompt up to ten times within a single day and found citation-source overlap of 0.32–0.43 — essentially the same as the 0.34–0.42 overlap between consecutive days. The usable reading is a rolling window: the standard error on one brand’s mention rate falls below 0.10 at 10 days and below 0.05 at 24. Use daily numbers for alerts only. That study tested ChatGPT, Gemini, Google AI Mode and Perplexity.
- We are not mentioned for “best X”. Should we rewrite that landing page?
- Usually not — that is the most commonly misdiagnosed case. For recommendation intent, the sources a model cites are almost entirely third-party roundups and reviews, not your own site. Before touching any page, identify which gap you actually have: no mentions at all (fix the entity and structured facts first), mentioned but ranked late (supply comparable parameters), mentioned but described wrongly (a communications problem), or competitors cited and you are not (get into those sources). List the domains cited most for that phrasing and sort them by whether you can realistically get in.
Related
- AI Research
- Prompts & Query Analytics
- Share of Model
- GEO statistics
- GEO for Chinese brands going global
- All solutions
GEOly · Prompt Research Loop v1.0. A methodology document; all third-party figures remain the property of their original publishers.