--- name: ai-prompt-set-design description: > Design and audit prompt sets for AI search visibility monitoring (GEO). Covers the rule that outranks the rest, never pooling brand-named prompts into a discovery score, the prompt types worth covering, funnel and persona segmentation, intent clusters instead of single prompts, wording rules, source mining, and what a normal visibility number looks like for a brand of your size. Use when building a prompt set, auditing an existing one, deciding what to add or retire, or explaining why a visibility number looks wrong. Platform neutral: works with any monitoring tool, or with a spreadsheet. --- # AI prompt set design A prompt set is the measurement instrument. Every rate, rank, and share you report is a statement about the questions you chose to ask, so a set built badly produces a dashboard that is precise about the wrong market. This skill is platform neutral. The discipline belongs to the instrument, not the vendor. It applies to any AI search monitoring tool, and to a spreadsheet and patience. The final section covers what refd does and what you still have to do by hand. ## The rule that outranks everything else **Never pool brand-named prompts into a discovery or share-of-voice number.** A prompt that names your brand hands you the mention. That is a constant, not a measurement, and pooling it into a headline inflates the number with information you already had before you spent anything. One study of 102,025 prompt responses across 102 brands and five engines, with 149,912 citations, collected between March and May 2026, measured the effect directly. Split by whether the brand name appears in the prompt: | Engine | Branded prompts | Unbranded prompts | | --- | --- | --- | | ChatGPT | 94.2% | 22.1% | | Gemini | 94.0% | 18.7% | | Perplexity | 98.5% | 23.9% | | Claude | 100.0% | 51.5% | | Grok | 100.0% | 12.0% | Branded prompts were roughly 9% of first-run responses and recognised the brand at 94% or better on every engine. Including them lifted the pooled per-tier figure from 43.6% to 52.2% for mid-market brands and from 11.4% to 17.3% for small ones. The paper computes its headline result on unbranded prompts only, and says why in the text. That is the entire argument. A branded prompt answers "what does the assistant say about you", which is worth measuring. It is not evidence about whether an assistant would have found you. Branded prompts are not worthless. They are the only source of accuracy checks, competitive framing, and objection data, because no unbranded prompt can produce those. Keep them, keep them a minority of the set, and never let them touch a discovery aggregate. There is no universal right share, because the right share depends on how many buying decisions you need to watch. If you can only report one number, report the unbranded one. **Check this before anything else.** If your tool cannot separate the two cohorts in reporting, the headline is uninterpretable and no amount of prompt editing will repair it. Fix the reporting first. Changing prompts will not fix a broken aggregate. ## What a prompt set is A fixed list of questions, asked on every run, held stable long enough to compare. Twenty to thirty is a legitimate starting size for one brand in one market. The set defines the denominator of every metric derived from it, so changing it mid-campaign destroys the comparison that makes the metrics worth having. The taxonomy below is a practical scaffold, not a finding. No study tells you which five buckets to use. What the research does support is the first rule, persona conditioning, and clusters over single prompts, all of which are separated out below because each one has evidence behind it. ## The prompt types worth covering 1. **Category discovery.** "Best X for Y", "what are the best X". Where new buyers meet the category. Highest reach, hardest to win, and where most competitors live. 2. **Use-case fit.** "X for a [persona] with [constraint]". Qualified intent. This is where a strong fit beats a bigger brand. 3. **Comparison and alternatives.** "A vs B", "alternatives to X", "how does X compare". Highest commercial intent. Naming a competitor is legitimate and worth tracking. Naming the brand you are measuring is not, because that prompt belongs in the branded cohort. 4. **Brand-defining.** "Is X any good", "what is X", "how much does X cost", "what is the learning curve for X". What the assistant says about you specifically. This is the branded cohort. Segregate it. 5. **Problem and job to be done.** "How do I solve Z", "will X actually save me time". Captures buyers who do not yet know the category exists. Two more worth adding when you have budget: - **Market perception.** "What is a [category] and how should I evaluate one?" Names no brand at all. Tests whether the buying framework itself favours you, which is upstream of any prompt you could write about yourself. - **Implementation and proof.** "How do I measure whether X is working", "how do I prove X pays for itself". Authority-building, frequently cited, and usually missing from most sets. ## Funnel coverage Prompts must span the journey, because absence early is invisible and absence late loses the sale. | Stage | Buyer state | Prompt shape | Share | | --- | --- | --- | --- | | Awareness | Has the problem, does not know the category exists | Concern, doubt, objection, job to be done | 15-20% | | Consideration | Building a shortlist | Category discovery, use-case fit | 35-45% | | Evaluation | Choosing between vendors | Comparison, alternative | 20-25% | | Purchase | Ready now, or nearly | Where to get it, price, trial, onboarding friction | 10% | | Brand-defining | Verifying you specifically | Accuracy, framing, objections | remainder | These proportions are a starting distribution, not a measured optimum. Two failure shapes are worth naming because they are common and they are invisible: - A set that is all category discovery will show you losing every comparison and never tell you why. - A set that is all comparison will show you winning the evaluation stage while being invisible at the moment buyers form a shortlist. Skip the purchase stage only if you genuinely do not sell direct. ## Personas: measured, not optional **A blended persona view hides the fact that you win one segment and lose another, and the size of the effect is large enough to justify separate prompts per segment.** An audit of 2,000 runs across 10 personas, 8 prompts, 3 model configurations, and 10 repetitions per cell measured what happens when the same question is asked of a different buyer. Prefixing the user message with a persona dropped the overlap between recommended brand sets by 0.12 to 0.20 against a within-persona reference, with clustered 95% confidence intervals excluding zero on all three model cells. The effect was sharply stratified by how established the brand is. Category leaders held 80% same-brand consistency as the persona changed. Mid-market brands swapped up to 75% of their recommendation set. The authors' conclusion is the rule: any measurement of AI brand perception has to condition on the persona supplying the query, because a protocol that aggregates across personas systematically hides that variation. The same question to two audiences can return entirely different answers, and a single average over both is a number about nobody. Practically: - For every broad category prompt, add one variant per priority persona. - Persona prompts must carry the constraint that makes the segment real. Weak: "best voice tool" Strong: "best Mac voice automation app for software developers who live in the terminal" - Keep the personas as separate, labelled prompts so you can read visibility per persona. If you cannot separate them in reporting, the variants cost you queries and buy you an unreadable average. - If you sell across regions, add region-specific phrasing where it changes the answer. One variance decomposition of 12,933 responses across 8 languages and 3 models found query language accounted for 26.5% of single-response variance against 1.5% for brand identity, so a single-language set is blind to a large systematic effect. ## One prompt is not a measurement **Treat one buyer intent as an attribute, not as a prompt.** A study of roughly 6,000 paraphrase runs against roughly 6,000 same-prompt rerun controls, on OpenAI and Anthropic models, measured how much of an answer is driven by the question's exact wording. Using Jaccard overlap between the sets of recommended brands: | Comparison | Overlap | 95% CI | | --- | --- | --- | | Same prompt, rerun | 0.50-0.61 | baseline | | Two paraphrases of one intent | 0.288 | 0.215-0.361 | | Paraphrase that adds a constraint | 0.135 | 0.098-0.175 | Paraphrasing the same buying intent moves the answer *further* than rerunning the identical prompt does. The wording is the dominant input to which brands surface, ahead of the intent behind it. The authors conclude that prompt-by-prompt mention tracking is structurally unstable as a unit of measurement, and that the fix is a different unit, not a bigger list. The academic prompt-sensitivity literature agrees on the mechanism and names the axis: across task types, paraphrasing is the variation that moves open-ended generation most, while template changes move multiple-choice tasks most. Recommendation questions are open-ended generation. What to do: - Write 3 to 5 phrasings per intent and score them together. - Read at the attribute level. The individual prompts do the measuring; the cluster holds the read you act on. - Expect variance inside a cluster. That variance is the most actionable thing in the whole exercise, because it tells you which wording to keep and which to retire. - Useful cluster axes: "best" versus "top", category label versus job description, broad versus segment-specific, capability-framed versus outcome-framed. A caution on the same evidence: the same variance decomposition found brand by prompt interaction to be near zero once other terms were removed. So a brand's *relative standing* is fairly stable across paraphrases even while the *set of brands recommended* is not. Expect rewording to reshuffle who is in the answer, and expect it to move your own rank far less. Read clusters for reach, not for position. ## Wording rules These are cheap to follow and they dominate results. **Write what a buyer types, not what an analyst writes.** Third-person spec phrasing is the most reliable predictor of zero visibility in a mature set. Compare: - Fails: "What level of offline or on-device processing do voice assistants offer for privacy?" - Works: "Can I use a voice assistant without my recordings going to the cloud?" One is a specification question. The other is a person protecting something they care about. Buyers ask the second. - **One question per prompt.** A bundled ask produces a muddy answer and an unreadable trend. "Developers and knowledge workers" is two audiences in one prompt. - **Evergreen, with no year stamps.** A year in the prompt breaks before and after comparability. Measure year-stamped variants as their own cluster. - **Customer language, not internal taxonomy.** Take phrasing from real users, not from your product categories or your sales deck. A generated set is a starting draft to edit, not a finished instrument. - **No near-duplicates.** Two prompts asking the same decision in slightly different words is waste. Cluster them instead. - **Neutral wording on discovery prompts.** Do not praise, hint, or name the brand you want to be found for. - **Freeze and version.** Once a set is live, hold it stable. A trend line across a changed population is not a trend line. If you must evolve the set, add on a stated date and treat the two periods as two series rather than one trend. ## Buying questions versus informational questions These shapes measure different things, and the distinction is easy to miss. **A buying question** ("best AI visibility tracker") makes the model name a specific product in nearly every run. It measures your recommendation rate. **An informational question** ("how do I get my site cited in ChatGPT") usually makes the model name nobody but cite sources almost every run. It measures whether your content gets picked up. A set of only buying questions hides your citation problem. A set of only informational questions hides whether you get recommended. Include both and read them separately, because the fixes differ. Recommendation gaps need authority and comparison coverage. Citation gaps need crawlable facts, structured data, and extractable answer blocks. ## Where prompts come from In priority order. Do not skip to inventing questions. 1. **Your own search data.** Export unbranded queries and rewrite each as a conversational question. This is proven demand, already filtered for you. 2. **Frontier models themselves.** Ask the models people actually use what real users ask about the topic. This surfaces intent before it reaches a keyword tool. 3. **People Also Ask and autocomplete.** Real aggregated queries, already phrased the way humans phrase them. 4. **Competitor pages.** The businesses ranking and getting cited for your topics have, in effect, published a validated question list. Read the pages that beat you and extract the questions they answer. 5. **First-party voice.** Sales calls, support tickets, on-site search, reviews, community forums. No competitor has this, and for objection prompts it beats anything you can infer. 6. **Query fan-out.** Run a few starter prompts and collect the follow-up questions the model generates. Those are the model's own prediction of what a user wants to ask next, already phrased naturally. 7. **Reddit and forums.** Heavily represented in model training data, so Reddit-style phrasing is phrasing models are unusually well tuned to. ## Calibration: know what normal is Do not report a rate without a baseline. Absolute numbers mean nothing out of context. The same 102,025-response study reports a clear brand-stature ladder in first-run visibility, restricted to unbranded category prompts: | Tier | Brands | Unbranded visibility | 95% CI | | --- | --- | --- | --- | | Global household names | 11 | 72.9% | 60.1-84.2% | | Established mid-market and regional | 36 | 43.6% | 36.4-50.9% | | Niche and small brands | 55 | 11.4% | 4.2-20.3% | About 30 points per step, significant at p<0.001 with Cohen's d above 1.3. Read that table with its own caveats, which the authors state: it is a single-author preprint, not peer reviewed, the author is affiliated with a commercial measurement tool, the data is that tool's own production database, and the cohort is convenience-sampled from four verticals with no claim to category representativeness. Read the shape of the gap, not the absolute rates. A different vendor measuring a different market will report different numbers. The part that survives the caveats is the practical one: **a niche brand at 11% to 15% is at the expected baseline, not failing.** Before treating a number as a problem, establish what comparable brands actually get. Three more findings from that dataset, carrying the same caveats: - The ranked "best-of" listicle is the most-cited page format, at about 21% of all citations. If you publish no ranked lists, you are leaving the highest-yield citation format available to you. - Sentiment framing flips about 6.7 times more often than mention does. Do not build a target, a cadence, or a decision on sentiment movement. Mention is the stable signal. - About 78% of citations went to corporate websites, but only 2.9% pointed back to the tracked brand's own domain, while 75.2% went to third-party brand pages such as a competitor's site or a business directory. Treat that as a question to check in your own data rather than a settled fact, because citation extraction differs between tools. If it holds for you, the implication is that being cited usually means being cited by someone else, and the lever is third-party authority rather than more messaging on your own site. When a brand *is* mentioned, watch the gap between mention rate and citation rate. High mention with near-zero own-domain citation means the model knows the name and will not retrieve the site, which is an extractability problem rather than an awareness problem. ## Budget for how little a single cycle can show Set expectations before you read your first month. An open-source probe that sends 20 buyer-intent prompts to a provider 5 times each reports the smallest change its own design can distinguish from sampling noise as 21 to 35 percentage points, at conventional significance and power, raising the floor to about 33 points if the threshold has to read in both directions. Those are design calculations for that probe's budget rather than universal constants, but they are the right order of magnitude to assume: in a small prompt set, a swing of a few points is noise, and buying more repeats of the same prompts barely moves the floor. Two practical consequences: - Run for a full cycle before reading anything. A month of answers is the minimum that says something, and it is still thin. - Prefer changes you can see in several prompts at once over changes in one prompt. One prompt moving is the paraphrase variance from the section above. ## Workflow 1. **Establish the brand's attributes.** What should it be known for? Attributes drive which prompts matter. Get this wrong and you generate data with no strategy to read it against. 2. **Enumerate the buying territory.** Segments, use cases, constraints, objections, geography, price posture. 3. **Build the core set.** Discovery and use-case first, then comparison, then a small brand-defining slice. 4. **Cluster.** Group into attributes, add phrasings, mark which prompt speaks for which attribute. 5. **Label every prompt** with its cohort, type, attribute, funnel stage, and persona. This is what makes the set auditable later. 6. **Run and baseline.** A full cycle minimum. 7. **Read by cohort, never blended.** Unbranded visibility is the headline. The branded cohort is a separate accuracy and framing report. Buying and informational questions separately. 8. **Iterate on gaps, not on noise.** Act where you are consistently invisible on an attribute you could credibly serve. Leave small swings alone. ## Auditing an existing set Work through these in order. Most sets fail at the first two. 1. **Are cohorts separable in reporting?** If the tool pools branded into discovery, the headline is uninterpretable. Fix reporting before changing the set. 2. **Compute the branded share and judge it against the two jobs.** Too much and the headline is carrying self-reference. Too little and you have no accuracy, framing, or objection data at all. There is no published target share, so judge it on what the unbranded half can still answer: if you could delete every branded prompt and lose nothing you report, the branded side is too small. 3. **Split into unbranded, competitor-named, and brand-named**, and report each separately. 4. **Map every prompt to a funnel stage.** Look for the stages with no coverage. 5. **Map every prompt to a persona.** Count how many carry a real constraint. One or two in a set of thirty is effectively none, and the measured persona effect above says that is leaving a real effect unread. 6. **Group into attributes.** Any attribute with a single prompt is unmeasured. 7. **Classify wording.** Count the analyst-voiced prompts among your zero-visibility prompts. If the failures cluster stylistically, the fix is rewriting, not new content. 8. **Check competitor coverage.** Each tracked competitor needs its own alternative, head-to-head, and differentiator prompts, or you cannot read share of voice against them. 9. **Check for near-duplicates**, and for prompts retired for the wrong reason. Never delete buyer-journey coverage to hit a prompt count. Segregate it. 10. **Verify new prompts have aged.** Anything added recently has no baseline. Do not judge it. ## Anti-patterns - Pooling brand-named prompts into a discovery score - Reporting one rate with no cohort label and no baseline - A single prompt per attribute, then concluding the attribute does not work - A blended persona average hiding which segment you lose - Deleting buyer-moment prompts to hit a prompt count instead of segregating them - A set that is all "best X" prompts - Inventing questions from internal taxonomy instead of mining real sources - Reading sentiment movement as a trend - Judging a prompt in its first cycle - Reordering the set mid-cycle and calling the change a trend - Acting on a few-point move in a small prompt set ## Output format for a drafted set Deliver a table plus per-prompt reasoning. Columns: `prompt`, `cohort`, `type`, `attribute`, `funnel stage`, `persona`, `wording note`, `action`. Then state explicitly: - Cohort counts and the branded share - Which funnel stages and personas are covered, and which are not - Every attribute with only one prompt, as a coverage gap - Which prompts are variants of which attribute - What to retire, and where each retired prompt should go instead - The cadence: how long to run before the set can be read **Reasoning per prompt is mandatory.** A set without a stated rationale per row cannot be reviewed, and an unreviewable set accumulates cruft. Where a prompt exists to answer a specific question, say which question, and which result would change a decision. ## What refd does, and what you still do by hand The discipline above is not vendor-specific. This section is specific to refd, and it is split honestly into the two halves. ### refd handles - **Prompt cohorts.** Every prompt is classified as `brand_defining` (the text names your brand), `alternative` (it names only a tracked competitor), or `discovery` (it names neither, so the rate is unprompted visibility). Classification runs the same alias matcher the scorer runs, so "this prompt names the brand" means exactly what "this answer mentions the brand" means, and a prompt naming both you and a rival lands in `brand_defining`, which is the biased case the cohort exists to isolate. Two further cohorts, `problem` and `market_perception`, are declared rather than derived: telling a problem-shaped question from a broad discovery one is a judgement about buyer intent that no keyword settles, so the product does not guess and you set them. Rows written before the column existed are classified on the first read that needs it, a `kind` you set explicitly is never overwritten, and an unclassified prompt is counted rather than dropped. - **Cohort-filtered rates.** A cohort rate is computed over that cohort's own answer cells, not re-weighted from the blended figure, so it is a rate for that cohort rather than a share of the whole. A filter that matches no prompt yields null rates instead of quietly falling back to the blend. - **A headline that names its population.** Every aggregate returns a `headline` object stating the cohort it was measured over, in `headline.population` and `headline.scope`, and the measures sit inside it rather than at the top level. A number can never be read as organic visibility when it was measured over brand-named questions. - **Asking nothing gets you discovery.** A prompt naming the brand scores near 1.0 by construction, so the default population is the discovery cohort, not the blend. The blend is still there, labelled and marked deprecated, for when you genuinely want it. A workspace with no discovery prompts falls back to all cohorts and says so in `headline.population` rather than reporting an empty population quietly. - **Every cohort at once.** `get_visibility_overview` returns all five cohorts side by side in `byCohort`, always over the full prompt pool, and the digest carries the same split. One call answers "how visible am I when nobody asked by name" without grouping prompt ids by hand. - **A filtered prompt read.** `get_prompt_performance` takes a cohort filter and applies it to the prompt list, the per-surface breakdowns, and the zero-visibility list alike, so a prompt the filter excluded cannot leak back in through a different field. - **Filtered prompt writes.** `add_prompt` and `update_prompt` classify the prompt they write and accept an explicit cohort override, and `list_prompts` reports each prompt's cohort alongside the per-cohort counts. - **A starting set.** Onboarding drafts prompts spread across Discovery, Evaluation, Comparison, Decision, and Authority, capped at five per category and steerable by total count and free-text focus. The generator is explicitly told that most questions must not name the brand, which is the cohort discipline applied at draft time. Treat the output as a draft to edit, not a finished set. A standard workspace holds up to 25 active prompts and 3 of the 5 surfaces (ChatGPT, Perplexity, Gemini, AI Mode, AI Overview); administrator accounts are not capped. - **A cohort control in the dashboard.** The Overview page has a cohort picker that narrows the whole page, and the Prompts table shows each prompt's cohort inline where a row can read near 100% while the workspace average does not. - **Frozen sets.** A run records the prompt ids and exact texts it will use, plus the tracked entity set, before collection starts. Editing a prompt while a run is in flight cannot change what that run measures. - **Set-change detection.** Share of voice, position, and competitor comparisons pause automatically when the tracked competitor set changed anywhere inside the compared span, because those metrics are relative to that set. Trend charts mark the break. - **Shared-cell comparison.** Two periods are compared only over the cells, one prompt on one surface, that both actually answered. A partial run cannot manufacture a change. - **Noise floors.** refd needs at least 4 shared cells before it reports any change at all, and at least 3 positioned or classified mentions before it reports a position or sentiment change. Thresholds are stated in the glossary and calibrated to seven-day windows rather than single runs, because at one sample per cell a day-to-day comparison is mostly sampling wobble: 5 points on mention and citation rates, 4 points on share of voice, 5 points on sentiment shares, and a quarter of a rank on position. - **Prompt retirement without deletion.** A prompt with history is deactivated and keeps its past results, then can be reactivated. - **Independent mention and citation.** They are stored and reported separately, and an answer can carry either, both, or neither. - **A valid empty result.** A Google query that returns no AI Overview is recorded as a valid observation, not a failed fetch, and coverage counts it. - **Named versus unnamed, as a standing signal.** The Home agent raises it as a suggestion when visibility on brand-named prompts leads visibility on prompts that do not name the brand by 20 points, which is where a headline starts being carried by self-reference. - **Sentiment discipline.** Classification runs on a queue hop after the answer is stored. Unclassified mentions are excluded from the distribution rather than counted as neutral, and a pending classification is shown as pending. ### Reading a cohort safely through the API `get_visibility_overview`, `get_competitor_landscape`, `get_citation_sources`, and `get_prompt_performance` all accept a cohort filter and apply it to everything they return. The digest does not, on purpose. It is a whole-workspace rollup that already carries every cohort side by side, so it has no cohort argument to get wrong. To read one cohort from the digest, read `sections.prompts.cohorts`. Whichever route you take, read the population out of the response rather than assuming it: `headline.population` and `headline.scope` on an aggregate, and the per-cohort rates inside `byCohort` or `sections.prompts.cohorts`. ### You still do by hand - **Versioning.** A run records which prompt ids and texts it froze, and refd can tell you which prompts entered or exited between two runs, but there is no prompt-set revision number and no way to filter history by set version. Keep your own set version and the dates you changed it. Note also that a cohort filter over historical answers reads each prompt's current cohort, not the one it had when the answers were collected, because the run freeze covers prompt text and nothing else. - **Funnel and persona coverage.** refd measures the prompts you give it. It does not tell you which funnel stages or personas your set is missing, whether two prompts cover the same intent, or whether one prompt is carrying an attribute alone. That audit is a spreadsheet, and the checklist above is the audit. - **Persona reading.** If you sell to more than one segment, you write the per-persona prompts and you read the per-persona results yourself. Nothing in the product segments results by persona. - **Category rollups.** Prompt categories are stored on every prompt and the dashboard can filter the prompt table by them, but no aggregate is computed per category anywhere. Rolling rates up by category is a join you have to make. - **The three-cohort breakdown on a screen.** The dashboard filters to one cohort and shows each prompt's cohort, but it does not display every cohort side by side, and it does not show the scope label the API attaches to a headline. For both, read `byCohort` or the digest. The rule that survives all of this: prefer a platform change over a prompt change when a metric is being misread. Deleting prompts to compensate for a reporting gap destroys exactly the data that segmentation would have preserved. ## Sources Every empirical claim above traces to one of these. Vendor-affiliated and preprint work is labelled as such, with the authors' own stated limits, and is never presented as a settled market-wide rate. - `https://arxiv.org/abs/2606.20065`: *Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines.* 102,025 prompt responses, 102 brands, 3,508 runs, 149,912 citations, five engines, March to May 2026. Source for the branded versus unbranded recognition table, the stature ladder, the 21% listicle citation share, the 78% / 2.9% / 75.2% citation breakdown, and the 6.7x sentiment volatility. Single-author preprint, not peer reviewed, author-affiliated with a commercial measurement tool, convenience sample, authors state no category representativeness. - `https://arxiv.org/abs/2605.27440`: *Paraphrase Brittleness in Production Retrieval-Augmented Commercial Recommendation.* Roughly 6,000 paraphrase runs and 6,000 same-prompt rerun controls on OpenAI and Anthropic models. Source for the Jaccard table and the conclusion that prompt-by-prompt tracking is structurally unstable. - `https://arxiv.org/abs/2605.30207`: *Persona Conditioning of Brand Recommendations in Retrieval-Augmented Commercial Chat.* 2,000 runs, 10 personas, 8 prompts, 3 model configurations, 10 repetitions. Source for the persona effect sizes and the prominence stratification. - `https://arxiv.org/abs/2607.13304`: *Where Does the Noise Come From? A Variance-Components Decomposition of Non-Determinism in LLM Brand Answers.* 12,933 responses, 20 brands, 8 languages, 3 models. Source for the 26.5% language variance share and the near-zero brand-by-prompt term. - `https://arxiv.org/abs/2410.02185`: *POSIX: A Prompt Sensitivity Index for Large Language Models.* Source for paraphrasing being the highest-sensitivity axis in open-ended generation. - `https://github.com/chirag23177/geo-probe`: open-source probe reporting a minimum detectable effect of 21 to 35 percentage points at 20 prompts and 5 repetitions. A design calculation for one budget, not a constant. - `https://arxiv.org/abs/2311.09735`: Aggarwal et al., *GEO: Generative Engine Optimization*, KDD 2024. The peer-reviewed entry point for the field, cited by the work above. For the platform-specific claims, see the companion skill for the refd MCP server at `https://refd.ai/skills/refd/SKILL.md`, and the metric definitions refd itself reports at `https://refd.ai/glossary`.