How we measure brand visibility in AI search.

For each brand, we run a fixed set of buyer-intent prompts against ChatGPT, Perplexity, Gemini, Claude, and Grok in parallel. Every (prompt × engine) verdict is classified one of three ways:

• Cited. The brand was named as a recommended option.
• Mentioned. The brand appears but is not recommended as a top choice.
• Invisible. The brand does not appear in the response.

Verdicts are aggregated through a weighted formula across four buyer-intent dimensions to produce a single MVI score (0-100) per brand. The full leaderboard, per-engine breakdowns, and category top-cited brands are published openly at /research.

Four terms recur throughout the research and on this page. Each links to the underlying evidence from the panel. The terms are not aspirational, they map to measured behavior.

GEO is the practice of optimizing brand content, entity data, and citation sources so generative AI assistants like ChatGPT, Perplexity, Gemini, Claude, and Grok recommend the brand in their answers to buyer-intent prompts. It overlaps with SEO and AEO but is built around the synthesized-answer format AI assistants return, not the ranked-list format Google returns.

Generative AI assistants do not return ranked pages. They return synthesized answers built from extracted entity facts, citation sources, and the model's training-time exposure to the brand. GEO targets all three. Entity data shapes how the model represents you. Citation sources shape what the model says about you. Content rewrites at the buyer-intent level shape whether you appear in comparison and evaluation answers, where the actual purchase decision happens.

What GEO does not include.GEO does not cover ChatGPT's web-scraped content (that is page-level optimization, indistinguishable from SEO). It does not cover Google AI Overviews specifically (a separate signal pulled from Google's ranking, not from a generative model). It does not cover paid placement inside AI assistants. GEO is the organic-citation problem inside generative assistants.

Evidence: AI engines disagree on the same buyer prompt (Finding 01) · Different engines decide different funnel phases (Finding 05) · 47% of brand × prompt cells are invisible to all 5 engines (Finding 07)

AEO is the practice of optimizing structured data, FAQ schema, Speakable schema, and on-page content so answer engines (voice assistants, AI assistants, featured snippet engines) surface a brand or page as the direct answer to a question.

AEO targets the answer surface, the moment a query gets a single direct response instead of a ranked list. That happens on voice assistants reading aloud, on Google's featured snippets, on AI assistants synthesizing an answer, and on schema-backed answer cards. The optimization techniques converge: structured data so the engine knows what is answerable, FAQ pairs so question-shape queries get direct extraction, Speakable schema so voice engines know which text to read, concise factual content so the answer is short enough to surface.

Why AEO matters now. Answer engines are eating the click. The buyer who asked Siri, Alexa, or ChatGPT a question and got a direct answer never visits a results page. SEO competition for those queries is over before it begins because there is no list to rank in. AEO is how brands stay in front of buyers who do not click.

Evidence: Perplexity has the highest mention-share in our panel (32%) · Speakable cssSelector pattern in this site's schema

MVI is a 0-100 proprietary score per brand combining citation rate across 5 AI engines and 4 buyer-intent funnel phases (Discovery, Filtered Discovery, Comparison, Evaluation), with Wilson 95% confidence intervals. It is the single number that measures AI search visibility.

The score is normalized to 0-100 within each segment so a 73 in skincare means the same thing as a 73 in luxury watches. Cross-segment comparison is the headline use case. The thresholds map to behavioral consequence (Default choice, Repeat use, First encounter, Not yet cited), not arbitrary cutoffs. See the behavioral-stage framework section below for the tier-by-tier breakdown.

Cadence. The same canonical buyer-intent prompts run on the 1st of each month against all 5 engines × 500 brands across 25 segments. Snapshots are date-stamped and archived to git so before/after comparisons are reproducible. Methodology version is bumped only when the prompt set or scoring formula changes.

Where to see MVI: Free /check tool (60-second MVI for any brand) · MVIP roster (every brand on the panel) · Per-segment leaderboards

Persona-Tuned MVIis the mapou Visibility Index weighted by a brand's actual buyer mix instead of the baseline persona. It is the AI search visibility number that matches who is actually asking, not the average of everyone.

Baseline MVI averages across a generic shopping-frame buyer. In our panel the leader brand changes under at least one buyer persona in 20 of 22 categories tested. The baseline number is real but represents only one slice of the buyer population. For categories where rankings move under persona, a brand needs the persona-weighted number to make budget decisions tied to actual buyer mix.

We measure MVI per brand under each of 12 personas (baseline plus 11 named buyer signals: budget, premium, professional, first-time, values-driven, gift-giver, time-pressed parent, brand-skeptic, gift-receiver, subscription-buyer, and two extended). The brand's actual buyer-mix percentages weight those persona-specific MVIs into a single score that matches the brand's real audience.

Where to see it: Free Persona Explorer (rankings under each persona) · Finding 08: leaders flip in 20 of 22 segments · Persona-Tuned MVI with custom buyer-mix weights ships in app.mapou.ai for paid customers.

We test the five AI assistants real shoppers use. Engines are weighted equally in the index regardless of market share, to avoid the methodological gymnastics of weighting by share-of-traffic (which itself shifts month to month).

• ChatGPT (OpenAI gpt-5.4-mini, routed via Perplexity Agent API), monthly
• Perplexity (sonar, native Sonar API with bundled web grounding), quarterly
• Gemini (Google gemini-3-flash-preview, routed via Perplexity Agent API), monthly
• Claude (Anthropic claude-haiku-4-5, routed via Perplexity Agent API), monthly
• Grok (xAI grok-4-1-fast-non-reasoning, routed via Perplexity Agent API), quarterly

Cadence policy. Three engines (ChatGPT, Gemini, Claude) are tested every month with the full canonical prompt set. Perplexity Sonar and Grok are tested quarterly because each per-call web search costs $0.005 on the Perplexity Agent API and stacking five engines at monthly cadence pushed recurring cost outside the budget for a self-funded research program. Quarterly Perplexity and Grok runs are date-stamped, archived separately, and the resulting MVI score for each segment carries a freshness indicator showing the most-recent run including each engine. We disclose this trade-off explicitly rather than running the full 5-engine panel monthly and absorbing the cost into the line items, because transparency about cadence is more useful than smoothing it over.

Retrieval architecture. Four of the five engine cells (ChatGPT, Gemini, Claude, Grok) are routed through the Perplexity Agent API with the web_searchtool enabled, capped at one search invocation per response. This gives every monthly cell live web grounding from the same continuously-refreshed Perplexity index. The Sonar cell uses Perplexity's native Sonar API (web grounding bundled into token cost). Differs from native consumer behavior at chatgpt.com, claude.ai, and gemini.google.com (which use Bing-backed, Anthropic-internal, and Google-Search-backed retrieval respectively). The methodological gain is uniform retrieval across cells; the trade-off is the routed retrieval is not the exact retrieval a real consumer sees on each vendor's native app. Disclosed for completeness.

Analyzer.Brand citation extraction from each engine's response usesgpt-5.4-nano via direct OpenAI API with strict JSON schema for verdict classification (cited / mentioned / invisible) and competitor-name extraction.

Model fallback.Each routed engine cell uses Perplexity's Model Fallback feature with a primary + secondary model chain. If the primary is temporarily unavailable (e.g.,gemini-3-flash-preview gets deprecated mid-quarter), the request auto-fails-over to the secondary at no additional fee. Billed at the model that actually serves the request. Telemetry captures which model served each call so we can detect when fallbacks fire and disclose any model-mix shift on time-series charts.

Search recency. Every routed call passes search_recency_filter: "year"to the web_search tool. This restricts retrieval to content from the last 12 months. Removes long-tail outdated reviews while preserving the live-web reality our customers care about. As a side effect, it narrows the search target and reduces searches per call.

Search budget. Every routed call instructs the model to use exactly oneweb_search invocation per response (enforced via system prompt instruction, not API hard limit). Per-run telemetry confirms compliance — the average searches-per-call across the 25-segment monthly canonical is published with each run in data/research/runs/[run_id]/_telemetry.json.

Microsoft Copilot and Google AI Overviews are not yet included; both lack clean public APIs. We're evaluating SerpAPI integration as a future addition.

Each engine cell uses the cost-efficient mini/haiku/flash variant of its vendor's model lineup, not the flagship. We chose this to keep monthly canonical cost within budget for a self-funded research program AND because direct A/B testing showed the cheaper variants produce equivalent brand rankings.

Direct A/B test (May 2026): we ran the ChatGPT cell on gpt-5.4-mini ($0.75/$4.50 per 1M tokens) and gpt-5.5 ($5/$30 per 1M tokens) head-to-head across three segments (mens-fashion, fs-banking, beauty-skincare), 60 brands × 20 prompts × 1 engine cell × 2 model variants. 1,200 atomic verdicts per variant. Result:

• Spearman ρ = 0.889 average across the three segments (rank correlation of the two leaderboards). 1.00 would be identical rankings; 0.0 would be independent.
• Top-3 brands stable in 2 of 3 segments; the third (mens-fashion) had a cosmetic top-1 swap with both candidates within a 4 MVI point window in both versions.
• Top-10 overlap = 90% on average. Long tail of the leaderboard is durable.
• Mean MVI |Δ| = 7.8 pts; p90 = 18 pts; max = 21 pts. Middle-pack brands show higher cross-model variance than top or bottom of the leaderboard.

Decision: stay on the cheaper variant. The flagship costs ~10× more per token and runs ~3× slower per call, but produces equivalent rankings within the noise floor of a single monthly snapshot. Absolute MVI scores in the middle of the leaderboard should be read as ranges (±5-10 pts) rather than point estimates.

Re-run criteria: when a new flagship lands (gpt-5.6, claude-opus-5, etc.), when a paying customer asks for tighter precision in the middle of the leaderboard, or when a model deprecation forces the choice. Raw audit data committed under data/research/ab-tests/2026-05-09T02-*/.

The MVI is composed across four buyer-intent phases. Each phase tests a different relationship to brand visibility. The exact weighting is part of mapou's proprietary methodology.

mapou describes AI visibility as a sequence of three behavioral stages plus a pre-encounter state. The MVI score maps to where a brand sits in this sequence right now.

The MVI score answers “how visible is this brand?” but most strategic questions need a layer up. Five derived metrics computed from the same data, surfaced on every per-segment report and aggregated across all segments on the State of AI Search:

• Cross-engine divergence (0–1). One minus the mean pairwise Spearman rank correlation across the five engines. Near 0 = all engines rank brands similarly; near 1 = the engines act as independent channels. Tells you whether AI search is one channel or five for your category.
• Effective number of brands. The headline concentration metric, computed as 1 / Σ(MVI share²), the inverse Simpson index (equivalently a Hill number of order 2, equivalently 10,000 / HHI on a 0–1 share basis). Reads as “AI effectively recommends X brands out of N tracked.” Self-normalizing for sample size, so a 10-brand segment is directly comparable to a 20-brand one. Reported alongside the top-2 share for narrative clarity. Tells challenger brands whether the segment is dislodge-incumbents or show-up-at-all territory. (Gini coefficient is also computed and stored for back-compat; not surfaced in the UI because it carries an income-inequality analogy that doesn't map cleanly to AI-search visibility.)
• Discovery → Evaluation leakage(per brand, ±1). Citation rate in Discovery prompts minus citation rate in Evaluation prompts. Positive value = the brand is awareness-rich and conversion-poor in AI answers. Surfaces brands AI knows but won't recommend at the buying moment.
• Mention-but-not-cited gap (per brand, 0–1). Share of total visibility that came as mention only. AI uses the brand as context but rarely surfaces it as the answer. Recognition without recommendation.
• Kingmaker engine per phase. For each segment and funnel phase, the engine where the citation-rate gap between most-cited and least-cited brand is widest. That's where positioning matters most. Win that engine, win that phase.

All five run at render time on the published research data. Numbers move when AI behavior moves, no recompute step needed.

Each brand has 100 atomic observations per monthly run. With sample sizes this large, we can compute meaningful confidence intervals on the underlying citation rate.

We use Wilson 95% confidence intervals, narrower and more accurate than normal approximations for binary outcomes. Every brand on the leaderboard shows MVI ± CI bounds. Brands with similar MVI but non-overlapping CIs are statistically distinguishable.

Engine agreement: we report what fraction of (brand, prompt) pairs all engines agreed on. Higher agreement = higher confidence in the verdict.

The MVI is built on a fixed prompt set generated once per category and reused every monthly run. The same brands are tested every month. Methodology version is locked (currently v1.0); changing the prompt set, weights, or thresholds requires a version bump.

All results are persisted with run-ID timestamps and immutable archives. Brand-level MVI history is queryable from the day this methodology was first run.

Re-running the same prompts every month means MVI deltas are paired comparisons, not noise. A brand moving from MVI 42 → 51 month-over-month reflects real signal change, not random LLM variability.

Every monthly report ships with a small set of falsifiable, dated predictions about what the next run will show. Each prediction names a metric, a threshold, and a direction, and gets graded green or red against the data when that next run lands. Predictions are committed before the data exists, not chosen post-hoc, the same discipline empirical science papers use to prevent narrative-fitting.

The current cycle's active claims and the running track record are public on the State of AI Search page. Segment-scoped predictions also surface on the relevant per-segment report.

Why this matters: a methodology that can only be confirmed retroactively is not a methodology, it is a storytelling apparatus. Pre-registration is the difference between an analysis that could have come out differently and one that could not.

Honest disclosure of what the measurement does not measure:

• Snapshot, not continuous. Each run is a moment in time. AI assistants are stochastic; results vary minute-to-minute.
• Region-specific. All API calls originate from US East. Brand visibility likely differs in other markets.
• Coverage gaps. Microsoft Copilot, Google AI Overviews, and voice assistants are not yet tested.
• Conflict of interest disclosed. mapou is a consultancy that benefits from AI search adoption. Scores are not adjusted for client or non-client status.
• One snapshot is not a trend. Single-month scores are diagnostic. Time-series gets meaningful at month 3+.
• Calibrated to high-intent queries. All 20 canonical prompts target commerce decisions across Discovery, Filtered Discovery, Comparison, and Evaluation. Findings should not be generalized to navigational or broad informational search behavior, which still rewards traditional SEO.
• Upstream influence, not downstream performance. MVI measures presence in AI-mediated discovery, analogous to impression share in early paid search. It is not a proxy for conversion lift, branded-search uplift, or assisted revenue. Tying citation rate to outcomes requires first-party clickstream and incrementality testing, which we ship as paid engagements.
• Hallucination not yet scored. Every (brand × prompt × engine) verdict is currently classified as cited / mentioned / invisible. We do not yet flag when an engine misattributes a product, confuses categories, or invents a brand. A failure-mode classification is on the v1.2 roadmap.
• API surface, not personalized chat. The canonical MVI run uses stateless engine APIs: no account, no chat history, no Memory, no custom instructions, no browsing personalization. Same prompt and same model produce the same distribution of answers regardless of who runs them, within sampling error. We do not capture how chatgpt.com, claude.ai, gemini.google.com, or perplexity.ai behave for a logged-in user with personalization signals enabled. Those consumer surfaces can vary per user. We measured the size of that effect via API system-prompt variants (no consumer-app scraping): leaders flip in 5 of 5 segments tested under at least one persona vs baseline, mean top-3 overlap drops to 67%. See Finding 08 on the macro page.
• Shopping-assistant framing, by design. The MVI baseline uses a shopping-assistant system prompt ("answer the user's question naturally and recommend specific brands and products where relevant"). This is the right scope for measuring commerce signals, but it is not zero-frame neutral. We isolated the framing effect by comparing it to a minimal "answer the user's question concisely" prompt across 5 segments: shopping framing inflates citation rates by 20-40pp on top brands and changes the leader in 4 of 5 segments. Baseline measures "what AI shopping assistants recommend," not "what AI engines say in raw form." Both are valid; the MVI deliberately measures the former.

Some claims in this methodology and on the research pages are well-powered inferential statements; others are exact descriptive observations of our specific panel. Disclosing the difference, layer by layer:

• Per-brand MVI scores have Wilson 95% confidence intervals. Computed and displayed on every leaderboard. With n=100 verdicts per brand at the engine level (20 prompts × 5 engines), differences of 5-10 percentage points are well outside sampling noise.
• Per-engine personalization sensitivity (Finding 09) has bootstrap 95% CIs. Each engine has n=270 (segment × persona) overlap observations across the panel. The 8-percentage-point spread between the most stable engine (Claude) and the most sensitive (Gemini) clears every pairwise confidence interval. The ranking is statistically real, not artifact.
• Per-cell citation rates (single segment × single engine × single persona) have wider precision.At n=20 prompts per cell, a 50% citation rate has a Wilson 95% CI of roughly 30-70%. Anywhere we cite specific small differences between cells (for example, "ChatGPT 70% vs Grok 70%"), readers should treat the comparison as suggestive within ±10pp of noise. Large effects (citation rate 0% baseline → 90% under one persona) survive this noise.
• Aggregate counts on n=27 categories are exact for our panel, descriptive not inferential. When we say "6 of 27 categories have 3+ different baseline leaders across the 5 engines," that is an exact count of our specific 22-category panel. As a generalization to all consumer categories, the binomial 95% CI on 6/22 is roughly 12-48%. The directional finding holds; precise extrapolation does not.
• "Leader held" / "unmovable leader" claims are descriptive at n=10 personas. When a brand holds #1 across all 10 buyer personas tested, that is literally true at n=10 — but you cannot statistically distinguish "always" from "9 times out of 10" with this sample size. The accurate phrasing is "held #1 across all 10 buyer personas tested," not "is unmovable." We are tightening any "unmovable" language in marketing copy to match this distinction.
• Cross-engine baseline disagreement is descriptive of the prompts we ran, not the population. The 6 categories with 3+ distinct baseline leaders are exact for our 20 canonical prompts per segment. A different (also legitimate) prompt set could produce different counts. We ship the same 20 prompts every month so time-series readings are comparable to themselves.

In short: directional claims and large-effect findings are well-supported. Precise small-effect claims and aggregate proportions on n=27 should be read as descriptive observations of the panel, not population parameters. Where we use phrases like "always" or "every category" we mean "in every case we tested," with the sample size implied by the surrounding context.

Methodology designed by Arvin Nundloll, formerly Director of Strategy and Business Development at Comcast Advertising, with prior roles at NBCUniversal, Amazon, and DIRECTV. MBA, William & Mary. Based in New York City. About →