The Shattered Mirror: Why Every AI Sees a Different World
The Shattered Mirror: Why Every AI Sees a Different World
Executive Summary
Tonight we found the thing that changes everything about how we should advise our client base, and it came from an experiment someone ran with four AI platforms and a single question: "best AC repair company near me in Arlington TX." Zero contractors appeared in more than one tool's recommendations. Not low overlap. Zero.
We've been building frameworks through extensive research: the Trust Prism (how each platform evaluates differently), the Entity Truth Layer (how to structure identity), the Root System (how to surface proprietary knowledge). Tonight those frameworks crashed into hard data, and the collision produced something I'm calling The Shattered Mirror -- the discovery that AI platforms don't just rank the same businesses differently, they see entirely different businesses. ChatGPT favors review platforms and franchises. Perplexity rewards well-built local websites. Gemini pulls from Facebook and Nextdoor. Four platforms, four completely separate realities, zero intersection.
This matters urgently because of the second finding: 45% of consumers now use AI for local business recommendations (BrightLocal 2026), up from 6% in 2025. A 750% increase in one year. AI is now the third most-used discovery platform, behind only Google and Facebook. And yet individual contractors are recommended less than 1% of the time. The chasm between consumer adoption and business visibility is the widest gap I've documented through our research.
We also dug deep into what actually predicts AI citations, using SE Ranking's landmark 2.3-million-page SHAP analysis. The answer challenges much of the GEO advice being sold right now: domain traffic dominates (SHAP 0.63), content formatting accounts for roughly 5% of citation variance, and domain authority is nearly irrelevant (r=0.18). The industry's "format your content for AI" playbook is the equivalent of polishing the hull while ignoring the engine.
Topic 1: The Citation Hierarchy -- What Actually Predicts AI Citations (And What Doesn't)
The SE Ranking SHAP Analysis
SE Ranking published the most rigorous AI citation prediction study I've seen: 2,328,533 pages from 295,485 domains across 20 niches, analyzed with XGBoost regression and SHAP (SHapley Additive exPlanations) values. SHAP is borrowed from game theory -- it measures the marginal contribution of each feature to the prediction. Higher SHAP = more influence on whether a page gets cited.
The hierarchy is stark:
| Factor | SHAP Value (AI Mode) | What It Means |
|---|---|---|
| Domain traffic | 0.63 | Sites with >1.16M monthly visitors get 6.4 citations avg vs 2.4 for <2,700 visitors |
| Referring domains | 0.56 | <300 referring domains: 2.5 citations. >24,000: 6.8 citations |
| Content length | 0.20 | Sweet spot around 1,500 words. Over 2,300 shows diminishing returns |
| Page Trust score | 0.12 | Threshold at 19-23 points. Below 7: 2.7 citations. Above 24: 6.2 |
| Quora mentions | 0.087 | Optimal range: 3,800-93,000 mentions. Beyond that, slight decline |
| INP (Interaction to Next Paint) | 0.085 | Technical performance matters -- faster interaction = more citations |
| Content freshness | 0.071 | Updated within 2 months: 5.0 citations vs 3.9 for >2 years old |
| Semantic URL relevance | 0.06-0.07 | URL structure should reflect content topic |
| LCP (Largest Contentful Paint) | 0.051 | Keep under 1.85s. Above 2.65s: measurable citation drop |
| FAQ sections | 0.049 | Actual FAQ content in body: 4.9 citations. No FAQ: 4.4 |
| Reddit mentions | 0.048 | Optimal: 35,000-718,000 mentions |
The shocking finding: Domain authority (DA), the metric the SEO industry has been selling for 15 years, shows a correlation of only r = 0.18 with AI citations. It's nearly irrelevant. What matters is actual traffic -- are real humans visiting this site? -- not link-based authority scores.
The Formatting Myth
The Sill Blog published an independent analysis of 1,238 AI-cited pages and found that structural formatting (headings, bullet lists, statistics density, tables, images) accounts for only ~5% of citation variance (adjusted R-squared = 0.045).
This is devastating for the entire "GEO content optimization" industry. The playbook being sold -- "add statistics, use clear headings, include quotation marks" -- isn't wrong, but it's like optimizing your business card when the real question is whether anyone knows your company exists.
The data reveals a clear hierarchy:
- Domain-level signals (~70% of prediction): Traffic, referring domains, brand mentions across the web
- Semantic relevance (~20%): Does the content actually address the query with substantive answers?
- Content formatting (~5%): Headers, statistics, structure
- Everything else (~5%): Schema, social signals, miscellaneous
ChatGPT vs AI Mode: Different Engines, Different Weights
The SE Ranking data reveals meaningful differences between how Google AI Mode and ChatGPT weight factors:
| Factor | AI Mode SHAP | ChatGPT SHAP | Implication |
|---|---|---|---|
| Domain traffic | 0.63 | 0.62 | Similar weight |
| Referring domains | 0.56 | 1.21 | ChatGPT values backlinks 2.2x more |
| Content length | 0.20 | 0.06 | AI Mode cares about length; ChatGPT doesn't |
| Content freshness | 0.071 | 0.018 | AI Mode rewards freshness 4x more |
This means the optimization strategy differs by platform. For Google AI Mode: keep content fresh, aim for 1,500+ words, optimize Core Web Vitals. For ChatGPT: build referring domains and brand mentions -- the content itself matters less than who links to you.
The FAQ Schema Surprise
FAQ schema markup showed minimal impact on AI Mode citations -- roughly 3x weaker than actual FAQ content in the main body text. AI Mode prioritizes substantive Q&A sections over metadata. This directly challenges my previous research finding that FAQPage schema had a 67% citation rate. The resolution: schema helps with AI Overviews (Google's older feature) but matters much less for AI Mode (Google's newer feature). The citation landscape is platform-specific all the way down.
Listicles Are Declining
Seer Interactive documented a 30% month-over-month decline in listicle citation strength in early 2026. Listicles still dominate volume (38% of cited pages, 43.8% of ChatGPT citations), but their effectiveness is fading. This aligns with the March 2026 core update's emphasis on Information Gain -- listicles that merely aggregate existing information without adding original insight are being devalued across both traditional search and AI citations simultaneously.
Platform Fragmentation: The Numbers
The fragmentation between AI surfaces is more severe than we previously documented:
- AI Overviews vs AI Mode: Only 13.7% URL overlap
- ChatGPT vs Google Top 10: Only 6.82% overlap
- 28.3% of ChatGPT's most-cited pages have zero Google visibility
- 77% of cited domains appear in only ONE AI surface
- Citation rates vary 615x between platforms (Grok: 27.01% vs Claude: 0.04%)
This isn't fragmentation. This is balkanization. Each AI surface is essentially a separate search engine with its own corpus, its own ranking logic, and its own winners.
Topic 2: The Local AI Visibility Crisis -- And the Shattered Mirror
The 750% Adoption Surge
BrightLocal's 2026 Local Consumer Review Survey produced the most important number I've seen this year: 45% of consumers now use AI tools for local business recommendations, up from 6% in 2025. A 750% increase in twelve months.
The breakdown:
| Metric | Value |
|---|---|
| AI adoption for local discovery | 45% (up from 6%) |
| AI rank in discovery channels | 3rd (behind Google 71%, Facebook) |
| Surpassed | Yelp, TripAdvisor |
| ChatGPT for local recs | 31% |
| Google AI Mode for local | 23% |
| Trust AI recommendations | 63% of active users |
| Trust AI equally with reviews | 42% |
| Highest adoption age group | 30-44 (64%) |
| Lowest adoption | 60+ (24%) |
| Verify AI recommendations | 88% fact-check sources |
| Always cross-reference reviews | 42% |
The trust numbers are particularly striking: 42% of consumers now trust AI recommendations as much as traditional reviews. For millennials (the 30-44 cohort), AI is becoming the primary discovery layer.
The Visibility Crisis: Less Than 1%
Against this adoption surge, the visibility data is brutal:
- Individual contractors are recommended by AI less than 1% of the time (Metricus)
- AI platforms overwhelmingly recommend lead-gen platforms, not actual service providers:
- Angi: mentioned in 90%+ of AI responses about home services
- HomeAdvisor: ~75%
- Thumbtack: ~65%
- Individual contractors: <1%
The structural reason: Angi generates 50-70 million monthly visits with millions of indexed pages. The average local contractor gets 100-1,500 monthly visits. That's a 10,000x to 100,000x web presence disparity. And as the SE Ranking SHAP analysis just confirmed, domain traffic is the #1 predictor of AI citations.
This creates a vicious cycle: AI recommends platforms with high traffic -> consumers use those platforms -> platforms get more traffic -> AI recommends them even more. The individual contractor is structurally invisible.
The Shattered Mirror: Zero Overlap
MarketingCode ran the experiment that crystallized everything. They asked four AI platforms the identical question: "best AC repair company near me in Arlington TX."
The result: Zero contractors appeared in more than one tool's recommendations. Not low overlap. Zero.
Each platform reconstructs local reality from completely different signals:
| Platform | Primary Sources | Selection Bias |
|---|---|---|
| ChatGPT | Review platforms (Angi, Yelp, BBB, Google Reviews) | Favors national franchises, emergency service brands |
| Perplexity | Local business websites directly; cited 17 sources for one query | Rewards specific, well-built websites more than any other tool |
| Gemini | Facebook, Nextdoor, community platforms | Completely ignores data sources other AIs use |
| GPT-4.1 | Mixed, but different selections than ChatGPT | Different model = different recommendations even from same company |
This is what I'm calling The Shattered Mirror. The Trust Prism described how each platform evaluates trust differently. The Shattered Mirror goes further: each platform literally perceives a different business landscape. They don't just rank the same businesses differently -- they see different businesses entirely.
For a local service business, being visible to ChatGPT does nothing for your Perplexity visibility, which does nothing for your Gemini visibility. You exist in fragments, and most businesses exist in no fragments.
The Conversion Paradox
Despite the visibility crisis, when AI does recommend a local business, the results are extraordinary:
- AI-referred leads convert at 73% vs 31% for Google organic leads (for contractors)
- AI visitors convert 4.4x higher than standard organic across all verticals
- AI visitors bounce 27% less and spend 70% longer on sessions
- But AI traffic represents only 3-7% of local service traffic (vs 18-25% for tech/SaaS)
This creates the paradox: AI recommendations are the highest-converting channel in existence for local services, but almost no local businesses are receiving them. The prize is enormous; the door is almost completely shut.
Google Auto-Editing GBP: The Silent Rewrite
Google's AI is now actively scanning contractor websites, reviews, and competitor profiles to auto-populate Google Business Profile services sections. If your profile has gaps, Google fills them using its best guess.
The risks: - Inaccurate services listed without owner approval (Google may add services you don't offer) - Missing specialties when websites lack specific service pages - Competitor-influenced listings -- Google compares your profile against competitors and fills gaps based on what they list
This isn't theoretical. It's happening now, to our clients' profiles, without notification. Every client needs a GBP audit.
March 2026 Core Update: Local Impacts at Day 12
With the core update completing April 10-11 (2-3 days from now), the local-specific changes are significant:
| Change | Impact |
|---|---|
| GBP completeness | Now a direct ranking input (not just rewarded signal). 18% average rank drop for incomplete profiles |
| Review recency | Weight increased 2.3x. Past 90 days outperform accumulated volume |
| Owner response rate | Now correlates more strongly with pack position than raw review counts |
| Proximity | Tightened in competitive categories. Non-proximate businesses dropping |
| LSA presence | Now in 40% of local searches (up from 11% in early 2025 -- 260% increase) |
| Sponsored ads in 3-pack | 22% of queries (up from <3%) |
| AI Overview compression | Shows 32% fewer businesses than traditional 3-packs. 88% of markets affected |
| NAP inconsistencies | More consequential due to sharpened entity resolution |
The HVAC case study is instructive: one firm that optimized their site structure and shifted to buyer-intent keywords grew from 198 to 609 high-value keywords and achieved top-3 rankings in primary service areas immediately post-update, generating a spike in direct service calls.
What This Means for Our 169 Clients
Combining the citation hierarchy data with the local visibility crisis, we can now articulate a clearer strategy than we could previously:
The uncomfortable truth: For most of our local service clients, content formatting optimization (the most commonly sold GEO service) will have minimal impact because their domain traffic is too low to trigger AI citation in the first place. Polishing content for AI when you have 500 monthly visitors is like designing a beautiful storefront on a street with no foot traffic.
The real priority stack (revised from previous research):
-
GBP completeness and freshness -- now a direct ranking factor. Audit all profiles for auto-edited services. Respond to every review within 48 hours. Post weekly. This affects both traditional local pack AND AI surfaces.
-
Review velocity and recency -- the update weighted recency 2.3x higher. Client review campaigns should target fresh, detailed reviews mentioning specific services, not just star ratings.
-
Root System content -- surface the proprietary knowledge inside each business. This is the only content strategy that can overcome the domain traffic gap, because it creates pages that AI cannot find replicated elsewhere on the web.
-
Multi-platform presence -- given the Shattered Mirror, businesses need to exist across ChatGPT's sources (review platforms), Perplexity's sources (well-built websites), and Gemini's sources (community platforms). This means: strong Angi/Yelp profiles, a well-structured website with answer-first content, AND community engagement (Nextdoor, local Facebook groups, Reddit).
-
Schema and structured data -- attribute-rich LocalBusiness schema with specific subtypes remains important, but as a force multiplier for the above signals, not as a standalone strategy.
The Adoption Plateau: A Counterpoint
One finding tempers the urgency somewhat: Orbit Media's AI-Search Adoption Survey shows that overall AI search adoption has plateaued. ChatGPT usage is flat at 36% year-over-year. The percentage who prefer AI for local business search actually decreased from 26% to 24%.
The reconciliation with BrightLocal's 45% number: BrightLocal measures "have used AI for local recommendations" (ever/occasionally), while Orbit measures "prefer AI over Google for local search" (habitual default). AI is being tried massively but not yet preferred for local discovery. Google still dominates local search through Chrome integration and GBP infrastructure.
This suggests a window -- maybe 12-18 months -- before AI local search shifts from experimental to habitual. Businesses that build multi-platform visibility now will be positioned when that shift completes.
The GEO Market Signal
The economic signal supports urgency regardless of adoption plateau: the GEO market is projected to grow from $848M (2025) to $33.7B (2034) at a 50.5% CAGR. AI search ad spending is projected to leap from $1B (2025) to $26B (2029). Capital is pouring into this space. The infrastructure is being built. The adoption will follow the infrastructure.
Sources
- AI Search Statistics 2026: 60+ Data Points -- Superlines
- The AI-Search Adoption Survey (Orbit Media) -- 6 charts on where and how people search
- How to Optimize for AI Mode (SE Ranking) -- 2.3M page SHAP analysis
- What 1,238 AI-Cited Pages Reveal (Sill Blog) -- formatting variance study
- 2026 AI Citation Position & Revenue Report (Digital Bloom)
- Nearly Half of Consumers Ask AI for Business Recommendations (BrightLocal)
- Why Doesn't AI Mention My Contracting Business? (Metricus)
- Every AI Recommends Different Contractors (MarketingCode)
- Google Is Auto-Editing Contractor Business Profiles (MarketingCode)
- Local SEO After March 2026 Core Update (DigitalApplied)
- Great Visibility Reset (FTI Consulting)
- Google March 2026 Core Update Complete Guide (ClickRank)
- Google Core Update April 2026 (Quantifi Media)
- 45% of Consumers Use AI Search for Local Services (MarketingCode)
