Most brands are invisible to most AI engines

For CMOs, heads of ecommerce, retail media leads, AI search and GEO practitioners, and analysts covering AI-mediated discovery. The method is summarized at the end of this report; the full framework is on the methodology page.

The single-number AI visibility myth is dead. One brand. One Panel Score. One leaderboard. That framing reads cleanly in a board deck and falls apart inside the data. Across 25 segments and 500 brands, the same brand has different visibility on every engine, drops between funnel phases, and shifts under buyer persona. Twelve separate measurements per brand, not one.

Buyer signals move the leaderboard. In 32 of 36 categories, the #1 brand changes when AI receives a buyer signal it did not have on the baseline run. The persona-tuned Panel Score for a brand's actual buyer mix, not the average buyer, is the only useful visibility metric in categories where rankings move with persona.

Most brands are invisible to most engines. 52% of (brand × prompt) cells in the panel are invisible to all 5 engines simultaneously. The ceiling for "improvement" is much higher than most brand teams assume because the floor is much lower than they think.

Five engines, five answers, five investments. Mean engine agreement is 52%, well below the threshold where optimizing for one engine carries to the others. The right engine to optimize for first depends on which buyer phase your category leans on, and that varies by category.

Frame conditioning matters. Removing the shopping-assistant system prompt flips the #1 brand in 8 of 22 segments. An "AI search visibility" number is always frame-conditioned. Vendor benchmarks that publish different leaders are usually using a different frame.

What most AI search reports leave out and this one includes: a 12-persona, 5-engine personalization layer (Findings 08 and 09), bootstrap 95% CIs on the cross-engine sensitivity table, and five falsifiable predictions on record that get graded next month.

Across 25 segments, average engine agreement is 0.52. The five engines rank brands the same way only about half the time. Marketers planning AI visibility as one channel are wrong by construction.

Most-divergent segments (engines disagree the most about who's best)

• Investing & Wealthdivergence 0.88, engine agreement 0.12
• Fintech & Neobanksdivergence 0.69, engine agreement 0.31
• Smart Home & Connected Devicesdivergence 0.66, engine agreement 0.34
• OTAs & Travel Bookingdivergence 0.65, engine agreement 0.35
• Business Telecomdivergence 0.64, engine agreement 0.36

Most-consensus segments (the engines mostly agree)

• Home Appliancesdivergence 0.10, engine agreement 0.90
• Mass-Market OEMsdivergence 0.15, engine agreement 0.85
• Luxury & EV OEMsdivergence 0.20, engine agreement 0.80

Method: rank every brand within each engine by citation rate, then compute the mean pairwise rank correlation (Spearman) across all engine pairs. Engine agreement = mean correlation; divergence = 1 − agreement. Near 0 agreement means the engines are effectively independent channels.

In a winner-take-all segment AI effectively recommends only 3 of 12 tracked brands; in an open one it spreads citations across most of the field. Same Panel Score shares, two very different competitive worlds.

Most-concentrated segments (winner-take-all)

• TVs5.5 effective brands of 20 tracked, top 2 take 46%
• Audio (Headphones, Earbuds, Speakers)6.0 effective brands of 20 tracked, top 2 take 44%
• Credit Cards6.3 effective brands of 20 tracked, top 2 take 42%
• Athleisure & Active Lifestyle6.4 effective brands of 20 tracked, top 2 take 44%
• Laptops7.5 effective brands of 20 tracked, top 2 take 35%

Most-spread segments (visibility is widely distributed)

• Men's Fashion14.3 effective brands of 20 tracked, top 2 take 22%
• Residential Telecom13.4 effective brands of 20 tracked, top 2 take 20%
• Business Telecom13.3 effective brands of 20 tracked, top 2 take 22%

Method: effective number of brands = 1 / Σ(Panel Score share²), the inverse Simpson index. Self-normalizing for sample size, so a 10-brand segment and a 20-brand one are directly comparable. Equivalent to 10,000 / HHI on a 0-1 share basis.

For each brand, the gap between citation rate in Discovery prompts (top-of-funnel: best brands in X) and Evaluation prompts (decision: which one should I buy). Positive gap = brand is awareness-rich and conversion-poor in AI answers.

Top demand-leak brands

• Four Seasons (hotels)Discovery citation rate beats Evaluation by 48 percentage points
• loanDepot (mortgages)Discovery citation rate beats Evaluation by 45 percentage points
• Vrbo (online travel agencies)Discovery citation rate beats Evaluation by 42 percentage points
• Charlotte Tilbury (makeup)Discovery citation rate beats Evaluation by 33 percentage points
• Drunk Elephant (skincare)Discovery citation rate beats Evaluation by 33 percentage points
• TCL (TV brands)Discovery citation rate beats Evaluation by 33 percentage points
• Rivian (luxury and EV brands)Discovery citation rate beats Evaluation by 32 percentage points
• Singapore Airlines (airlines)Discovery citation rate beats Evaluation by 31 percentage points

Method: for each brand, share of total visibility (cited + mentioned) that came from mentions only. High values mean the brand has positioning in AI's conceptual map but isn't earning the recommendation slot.

Brands frequently named as context (“... unlike X, this product offers...”) but rarely surfaced as the answer. Recognition without recommendation.

Top mention-but-not-cited brands

• Samsung Galaxy Book (laptops)mentioned in 5, cited in 0, 100% mention-only
• MetLife Auto & Home (P&C insurers)mentioned in 6, cited in 0, 100% mention-only
• Going (online travel agencies)mentioned in 6, cited in 0, 100% mention-only
• Wealthsimple (investing platforms)mentioned in 30, cited in 2, 94% mention-only
• Roku TV (TV brands)mentioned in 12, cited in 1, 92% mention-only
• NerdWallet (mortgages)mentioned in 11, cited in 1, 92% mention-only
• Volvo (luxury and EV brands)mentioned in 19, cited in 2, 91% mention-only
• La Mer (skincare)mentioned in 44, cited in 6, 88% mention-only

Across 25 segments, here's which engine is the kingmaker (largest spread between top and bottom brands) at each buyer-intent phase. Win that engine, win that phase.

Method: for each segment and funnel phase, we look at each engine's gap between its most-cited and least-cited brand. The engine with the biggest gap is the kingmaker for that phase, it's where small differences in brand strength produce large differences in AI visibility. Counts show how many of the 25 segments each engine wins as kingmaker for that phase.

25% of tracked brands have a cross-engine citation-rate spread of 50 percentage points or more. The spread is the gap between the brand's best and worst engine. Two brands with identical Panel Score can be on completely different visibility paths: one durable across engines, one engine-dependent. Averages alone hide this.

Method: per brand, citation-rate spread = max(per-engine citation rate) - min(per-engine citation rate) across the five tested engines. Computed only for brands with at least three engines reporting a citation rate. 500 brands total. 42% have spread under 15pp (durable), 25% have spread 50pp or more (engine-dependent).

Across 10,000 (brand, prompt) cells where four or more engines responded, 52% are invisible to all five engines simultaneously, and only 3% are cited or mentioned by all five. Cross-engine consensus invisibility is the rule, not the exception. The recommendation set you actually compete for is much narrower than the brand list suggests.

Engines do not just disagree on rankings, they cite differently. Some confidently recommend, others hedge with mentions, others cite few brands per answer. Recognizing the personality of each engine tells you which engine to optimize for first if you can only target one.

Method: invisibility bucket counts (brand × prompt) cells with at least 4 engines reporting and tallies how many missed the brand entirely. Engine personality table aggregates over every (brand × prompt × engine) result. Mean brands per answer is bounded above by the analyzer's 6-competitor extraction cap, so the column compares relative behavior across engines rather than the true ceiling.

We tested it. 36 segments × 6 personas via API system-prompt variants (no consumer-app scraping). The leader brand changed in 32 of 36 segments under at least one persona. Mean top-3 overlap with baseline: 69%. makeup was the most volatile (27% overlap), ce-smartphones the most stable (93%).

Each cell is the brand cited most often across the 20 canonical prompts for that segment, when the AI engine is given that persona's system prompt. Cells shaded olive flag a leader change vs baseline. The full ranked leaderboard per (segment × persona) is in the underlying JSON.

Method: 5 segments × 6 personas × ChatGPT × 20 canonical prompts (the same prompts used for the monthly Panel Score run). Personas are API system-prompt variants, fully visible and replicable. We do NOT scrape consumer chat apps; that violates ToS and breaks reproducibility. "Top-3 overlap" counts how many of the baseline top-3 brands stay in top-3 under each non-baseline persona, averaged across the 5 personas. "Leader holds" is true only if the #1 brand is the same as baseline for every persona. Note: "baseline" uses the shopping-assistant system prompt, not zero-frame neutral. We separately measured the framing effect (shopping vs minimal prompt) and disclose it on the methodology page. Run id: 2026-05-08-1526. Methodology version 1.0.

The largest first-party AI personalization panel published. 37 segments × 5 engines × 12 buyer personas × 20 canonical prompts = 50,000 verdicts. Two findings stack: engines disagree on the right answer even before any persona signal (8 of the 22 categories with data from all five engines have 3+ distinct baseline leaders), AND each engine personalizes differently when a persona signal arrives. Mean top-3 overlap with baseline: 74% on the most stable engine (claude), 66% on the most sensitive (gemini). Same buyers, 8-point spread.

95% CI computed via 1,000-iteration bootstrap on the per-(segment × persona) overlap observations (n=270 per engine). The 8-percentage-point spread between the most stable engine (Claude) and the most sensitive (Gemini) clears every pairwise CI. *Leader heldcounts segments where the same brand was #1 across all 10 buyer personas tested. With n=10, this distinguishes "held in our panel" from "mathematically unmovable". See methodology caveats.

Engines disagree even before personalization

Without any persona signal, the 5 engines do not converge on the same #1 brand in 8 of 37 categories. Three or more distinct baseline leaders across the panel.

The 2 categories where all 5 engines agree on baseline

Even in these "universal agreement" categories, persona signals split the leaderboard. Most extreme: home-services has all 5 engines agreeing on Roto-Rooter as the baseline leader, but produces 10 distinct leaders across the 60 (engine × persona) cells in our panel. Universal baseline agreement does not survive personalization.

Method: 37 segments × 5engines × 12 personas × 20 canonical buyer-intent prompts × 264 tracked brands. Personas are API system-prompt variants, fully visible and replicable. We do NOT scrape consumer chat apps; that violates ToS and breaks reproducibility. Top-3 overlap measures how many of the baseline top-3 brands stay in the top-3 under each non-baseline buyer persona, averaged across the 10 buyer personas (excluding the bare framing-isolation persona). Note: "baseline" uses the shopping-assistant frame; the bare-vs-baseline framing-effect study is disclosed on /methodology. Explore any brand × persona × engine cell in the Persona Explorer.

Finding 09 showed engines personalize at different intensities. Finding 10 looks inside the funnel: which buyer signals move which phases the most? 18,665 per-prompt verdicts across 4 funnel phases × 5 buyer personas. The most volatile cell is Premium × Evaluation (32% leader-held); the most stable is Professional × Discovery (68%). The pattern: aspirational personas (premium, values-driven) move rankings hardest, especially at Evaluation. Practical personas (professional, first-time) preserve them.

Each cell is the share of (segment × engine) combinations where the baseline #1 brand stayed #1 under that persona at that funnel phase. Higher = persona doesn't move that phase. Lower = persona disrupts that phase. Shading runs from none (65% or more held, resilient) to dark olive (under 35%, volatile).

What this tells you about each persona

Method: 13,035 per-prompt verdicts from the persona robustness panel (run 2026-05-05-1424 + 2026-05-05-1659 + 2026-05-05-2256 + extension runs). For each (segment, engine, persona, phase): rank brands by citation count within phase. Compare each persona's top brand to the baseline persona's top brand at the same (segment, engine, phase). Aggregate per (persona × phase) is mean across (segment × engine) of "leader held." The 4 phases (Discovery, Filtered Discovery, Comparison, Evaluation) come from the 20 canonical buyer-intent prompts per segment, distributed 6/6/4/4. The 5 buyer personas shown are the public-tier subset; full 10-persona analysis is in the paid panel via app.mapou.ai.

Every other finding in this report uses a "shopping assistant" system prompt, the frame buyers experience inside ChatGPT shopping mode, Perplexity, and AI-search retail integrations. We tested what happens with a minimal prompt instead ("answer concisely"). Across 22 segments × ChatGPT × 440 prompts, removing the shopping frame flipped the #1 brand in 8 of 22 segments (36%) and dropped mean cite rates by 8 percentage points. Top-3 overlap stays high (82%), the consideration set is roughly the same, but who AI calls #1 changes. The implication for methodology: an "AI search visibility" number is always frame-conditioned. Picking the right frame for the question matters.

Method. Run id 2026-05-06-2027 (2026-05-08). For each of 22segments × ChatGPT, the same 20 canonical prompts were issued twice: once with the “shopping assistant” system prompt used everywhere else in Panel Score, once with a minimal prompt (“Answer the user's question concisely”). All other variables held constant. Leader = #1 brand by citation rate. Top-3 overlap = brand-name match. The 8 flipped segments are listed; the 14 unchanged segments kept the same #1 across both frames. The 7.8-percentage-point cite-rate inflation under the shopping frame is consistent across categories: shopping framing makes engines more confident citers, not just different ones.

Finding 08 used buyer-intent personas (budget, premium, professional). Finding 12 uses demographic personas: gender × generation, age ranges per Pew Research. Across 23 segments × 8 demographic cells, the leader changed in 15 of 23 segments (65%) under at least one demographic. Mean baseline-match rate across cells: 64%. Gen X men (46-61) matched baseline least often (52%). ChatGPT only.

Each cell is the brand cited most often across the 20 canonical prompts for that segment, when the AI engine is given that demographic's system prompt (age range only, gender named). Cells shaded olive flag a leader change vs the buyer-baseline (no demographic signal). Top 12 most-divergent segments shown.

Method: 37segments × 8 demographic cells × ChatGPT × 20 canonical prompts. Personas are API system-prompt variants naming an age range (Gen Z 18-29, Millennial 30-45, Gen X 46-61, Boomer 62+ per Pew Research) and a gender. The system prompt does not name a generation label; that is descriptive only. We do NOT scrape consumer chat apps. “Baseline” here is the buyer-baseline from Finding 08 (no demographic signal, shopping-assistant frame). Top 12 most-divergent segments shown; full 37-segment table is paid-tier. Run id from the Phase 2 panel, methodology version unchanged.

The findings above are abstract. Here's what they look like for a single segment, using the same May 2026 data. 20 brands tracked. Run the same lens on your category to find the equivalent picture.

01. Engines disagree. Engine agreement of 0.41across the five engines. ChatGPT's top brand is not Gemini's top brand. Treat each engine as its own surface.

02. Concentrated. AI is effectively recommending 11.4 brands out of 20 tracked. CeraVe and La Roche-Posay alone take 28% of citations. The other 18 brands fight over the rest.

03. Discovery to Evaluation leak. Drunk Elephant loses 33percentage points between “best skincare” prompts and “which should I buy” prompts. AI surfaces the brand at discovery and drops it at the moment of decision.

04. Mentioned, not cited. La Mer is mentioned 44 times but cited only 6. AI uses it as a reference point in answers about other brands but rarely recommends it directly.

05. Kingmakers split. Claude decides Discovery, Grok decides Filtered Discovery, ChatGPT decides Comparison, ChatGPT decides Evaluation. A ChatGPT-only plan misses any phase the others kingmaker.

How to read this for your category.High divergence, few effective brands and a named leak brand add up to a market that's up for grabs at the moment of decision. Defending Evaluation surface (comparison content, head-to-head reviews, structured buying-advice answers) is where share moves.

84% of tracked brands sit at first encounter or below. The recommendation set AI consistently surfaces is much narrower than the brand population.

One row per segment. Find your category. High divergence plus few effective brands means a split, winner-take-most space where engines disagree on who to recommend and retuning AI visibility can move real share quickly. Click any segment for the full leaderboard, per-engine breakdowns, and analyst note.

The findings above are research. Here is what to do with them. Each item is a portfolio-level decision, with the mapou service that operationalizes it. Numbers reflect the 36-segment monthly panel, refreshed on the 1st of every month.

1. 1

   Treat AI engines as separate channels

   Engines disagree on baseline #1 in 8 of 22 categories with data from all five engines. Single-engine optimization is a single-channel investment. Plan budget by engine, not by "AI."

   How mapou helps: Start with an AI Visibility Audit to map your position per engine.

2. 2

   Visibility is buyer-mix dependent

   Leaders flipped in 32 of 36 segments in the persona study under at least one buyer persona. Your baseline Panel Score is a calibration. Build the persona-corrected number into the board deck.

   How mapou helps: Persona-Tuned Panel Score surfaces the visibility number that matches your actual buyer mix.

3. 3

   Personalization is real and measurable

   AI search is not winner-takes-most in the abstract. It is winner-takes-most per engine, per persona. Brands that are invisible at baseline can be the top pick inside a specific persona, and the leader you see today depends on who is asking.

   How mapou helps: GEO & Citation Architecture restructures your entity data so AI surfaces you in the personas that fit your positioning.

4. 4

   Visibility tier is the conversation

   Panel Score 75+ is default-choice. 50-74 is repeat-use. 25-49 is first-encounter. <25 is not yet cited. The right strategy differs at each tier. A first-encounter brand needs visibility infrastructure. A default-choice brand needs to defend against persona-driven challengers.

   How mapou helps: AI Visibility Audit identifies your tier; GEO & Citation Architecture moves you up.

5. 5

   Time-series wins

   A monthly tracked Panel Score catches changes a one-shot audit misses. AI search retrieval policies, model upgrades, and competitor activity all move the leaderboard between snapshots. Static GEO is yesterday's SEO.

   How mapou helps: Executive Advisory keeps your monthly Panel Score tracked, with a CMO-level read on what changed and why.

Three implications drawn directly from the findings above. Each is observable in the panel data, not extrapolated from a single segment.

Falsifiable, dated, gradeable. Each prediction names a metric and a threshold. Next run we publish green or red against every claim, accumulating a track record we can be held to.

Predictions are committed before the next run, not chosen post-hoc. Each is graded against the same fixed prompt set, with the result and methodology version stamped into the record. Disagree with any of these? Run your own check and bring data.

Every AI answer in this report is currently classified one of three ways: cited (recommended), mentioned (named without recommendation), or invisible (not surfaced). The implicit fourth category, which we are actively building toward v1.2, is misrepresented: the brand is named, but the AI's description of it is wrong.

Track record matters more than headline claims. The classifier infrastructure (sample, prompt schema, persistence) is in place; what is not in place is the review pipeline that gets us to publishable confidence. We would rather ship a half-finished section that tells the truth about the gap than a polished section full of false positives.

AI answer engines are not the first surface to change how shoppers find brands, and they will not be the last. Reading these findings against prior search shifts and against the boundaries of what we measured keeps the takeaways honest.