AI Visibility · 1 August 2026

By Eddie Eastwood, Founder

How 20 Identical Queries Expose AI Search's Hidden Sources and Blind Spots

We ran the same 20 local-professional searches on ChatGPT and Google AI Mode. When customers ask for a dentist, plumber, lawyer, or accountant, which businesses do these engines recommend? Which ones do they read about but never surface? And how much do they actually agree?

The answers reveal structural differences in how AI engines surface and cite information — with direct implications for any business trying to show up in AI search.

What We Tested

We searched for local professionals in Austin, Texas across four categories, five prompts per vertical:

Dentists (5 queries) — emergency care, walk-in availability, implants, family-friendly, weekend hours Plumbers (5 queries) — emergency 24-hour service, water heater repair, burst pipes, drain cleaning, licensed providers Lawyers (5 queries) — family law, personal injury, divorce, free consultations, small business Accountants (5 queries) — tax, bookkeeping, business accounting [not detailed here; see full methodology log]

For each search, we ran the EXACT SAME prompt on both platforms: - ChatGPT (web version; geolocation was NOT overridden and remained set to the operator's UK location) - Google AI Mode (browser geolocation manually set to downtown Austin, TX)

Note: Google showed AI Overviews automatically in 6 of 20 queries (30%). The other 14 required manually switching to the separate "AI Mode" tab. This distinction between automatic AI Overview and manually-selected AI Mode is significant for visibility.

We captured every business named, every source cited in visible source panels, and observable recommendation decisions.

Screenshots from the 20 queries:

The Headline Finding: These Engines Pull From Different Sources

For subjective "best X" queries, the two engines recommended almost entirely different businesses. For objective factual queries, they converged more.

Google AI Mode required manual selection for 14 of 20 queries. Only 6 searches auto-triggered a standard AI Overview on Google's default results tab; the other 14 needed users to manually click into the separate "AI Mode" tab. This is a practical finding: if your business isn't appearing in the automatic AI Overview on Google's default results, many customers won't discover that an AI-powered answer exists.

For subjective queries vs. factual queries: - Subjective "best X" queries showed low overlap. When both engines answered the same subjective question ("best family dentist," "best plumber for water heater repair"), they recommended almost entirely different businesses. Across the subjective prompts, we saw only 1-2 shared business names out of 9-13 recommendations combined per query. - Factual queries showed higher overlap. For the objectively verifiable "dentist open on weekends" query, 4 of Google's 5 recommendations also appeared in ChatGPT's 7-name list — the highest observed overlap rate in this 20-query sample (4/5 Google and 4/7 ChatGPT).

What this suggests: The engines drew from different sources for subjective queries but converged for objective, factual queries in this sample.

Where Google Cites Sources: A Structural Problem

Google's AI Mode cited businesses from locations that had nothing to do with the Austin query:

Out-of-State & Out-of-Area Citations

Google's AI Mode cited sources from locations that had nothing to do with Austin:

Confirmed examples from this 20-query sample: - A Pennsylvania HVAC company's page cited to support an Austin water-heater recommendation - A Brownsville, TX plumbing company (280 miles away) displayed as a recommendation for an Austin emergency plumber query - Supporting sources for plumbing advice drawn from Florida, California, Missouri, Arizona, and New York - A Gibsonia, Pennsylvania business page cited 5 separate times to support local Austin business recommendations in a single answer

Most serious case: For the "plumber in Austin Texas for burst pipe" query, Google recommended Lone Star Plumbing with its location listed as Slaton, TX (320 miles away, in the Lubbock area). A customer in an active emergency could follow this recommendation to a business in a different region entirely.

Category Mismatches

Google's own business-profile cards sometimes contradicted the AI recommendation: - ABC Home & Commercial: Google's AI Mode recommended this business for plumbing work in the "emergency plumber" query, but Google's own business-profile category label showed "Pest control service." (The underlying business does offer plumbing services, but the category mismatch is visible directly in Google's UI.) - Goettl Air Conditioning: Recommended for an Austin plumbing emergency at a 512 area code number, but its map card displayed "Las Vegas, NV" as the primary listed location

Paid/Sponsored Sources

Google cited "sponsored selectee" listings from Super Lawyers and other paid directories. ChatGPT explicitly flagged these as "= Sponsored super lawyers selectees" in its sources — but both engines still read and considered them.

Source Quality Issues in Both Engines

Both engines cited sources with freshness and attribution problems:

Reddit: Aged Sources (Both Engines)

Both engines cited Reddit threads to support local-service recommendations, but with a freshness problem. The oldest Reddit citation found across both engines was August 2021 — nearly 5 years old at the time of this study. For queries about "best plumber" or "good family dentist," both ChatGPT and Google pulled from Reddit discussions spanning years old (August 2021) to recent (July 2026), without visible weighting for recency.

Lead-Gen and Directory Sites

For competitive "best X" queries, ChatGPT's sources were dominated by third-party ranking aggregators (Expertise.com, HomeAdvisor, Angi's List, SiftPros) rather than businesses' own sites. Some of these featured lead-gen hallmarks:

• Sponsored listings flagged explicitly ("100% Pre-Screened Dentist," national toll-free numbers)
• Conflicting rankings — ChatGPT read SiftPros' "#1 pick" as a source but didn't recommend that business; instead recommended a different firm it ranked higher

Staging/Development Sites

For a divorce lawyer query, ChatGPT explicitly cited "Batrice Law Firm Staging" — a pre-production staging/development environment, not the live business website — as a source to support the business's qualifications in its recommendation.

Self-Inflating Claims in Sources

Businesses in both engines' sources displayed large, specific dollar figures ("$300M+ recovered," "$400M+ won," specific settlement amounts) in their own listings and directory entries. Both engines read these sources but notably did NOT promote the businesses with the highest claims into their top recommendations — a possible indicator that both engines may weight inflated claims with caution, though the sample size is too small to confirm intent.

The "Read But Never Named" Gap

Both engines have a consistent pattern: they read far more sources than they name as recommendations.

Confirmed examples from this sample: - ChatGPT on "licensed plumber in Austin": Cited 15 total sources; named 5 businesses in the recommendation. Of the 14 unnamed sources, one was an individual attorney with an out-of-state (Florida) trade license — a potential licensing-verification gap hidden in the background sourcing. - ChatGPT on "drain clog": Cited 16 total sources; named 5 businesses. The unnamed 11 included 4 businesses with specific recovery-dollar claims, national lifestyle-media DIY articles (Southern Living, Tom's Guide), and Reddit threads as old as August 2021. - Google AI Mode on "small business lawyer in Austin": Cited 24 total sources; named only 6 businesses. Of the 18 unnamed sources, 9 were competing law firms' own pages cited to support different firms' recommendations.

What this means: If your business isn't in an engine's top recommendations, you may still be influencing the answer through background research. Your pages are being read, even if not surfaced to the user.

Citation Attribution Problems

Cross-Company Citations

Google's AI Mode cited competitors' website pages as supporting evidence for a business's own claims. For divorce lawyers, six different competing law firms' own pages were cited to support a SINGLE firm's recommendation — turning a competitor's marketing copy into evidence for another firm's credibility. This is structurally unusual and worth noting: AI engines are treating generic corroborating-topic content (e.g., "how to handle uncontested divorce") as equivalent to third-party validation, even when it's from a direct competitor.

"Quoted" Claims That Don't Actually Quote Anything

• A Yelp search-results page was cited as the source for two specific businesses' specific "10-minute response time" claims — but the Yelp page was just a search listing, not a verified claim for either business.
• A San Antonio Statesman award article about a dentist's nomination was cited by ChatGPT, read as a source for local-dentist recommendations, but was geographically irrelevant to an Austin query.

Domain-Based Trust Signals

An "americanbarassociations.org" domain (plural "associations") was cited by ChatGPT. The domain name resembles the real American Bar Association (americanbar.org), but is not affiliated with the actual ABA. This is the sample's clearest example of a domain name that mimics a trusted institution.

Highest Overlap Observed: The Weekend-Hours Query

In this sample, one query produced notably higher agreement between engines than others.

The "dentist open on weekends" query: - ChatGPT named 7 businesses - Google named 5 businesses - 4 of Google's 5 also appeared in ChatGPT's 7-name list - Overlap rate: 4/5 Google and 4/7 ChatGPT

Subjective queries like "best family dentist" or "which lawyer is best for complex divorce" showed only 1–2 shared names out of 9–13 combined.

How AI Engines Actually Choose What to Recommend

Subjective Queries = Structured Recommendations

For questions asking "best X" or "who's good with kids," both engines independently chose to tier their recommendations: - ChatGPT: Created decision matrices ("best for X situation → recommend Y firm") - Google: Organized answers into color-coded map categories and 3-tier groupings

Both engines also self-acknowledged uncertainty. Google opened one "best lawyer" answer with: "Finding the 'best' family lawyer depends on your specific needs..." — then immediately gave specific ranked recommendations, a framing that says "this is subjective" while simultaneously presenting an ordered list.

"Expertise" Rankings Vary Per Source

When ChatGPT read competing "best plumber" rankings from multiple sources (SiftPros, Expertise.com, HomeAdvisor), each named a different #1. ChatGPT's own final recommendation matched NONE of the three sources' #1 picks — it synthesized across them rather than adopting any single ranking wholesale.

Negative Information Is Sometimes Softened

When a Reddit thread's source snippet explicitly stated "whatever you do, don't call Radiant Plumbing," Google's AI Mode still recommended Radiant Plumbing as a top-5 pick, softening the warning to "some community members advise shopping around for minor repairs." The negative source was visible in Google's own expanded sources panel for anyone who dug into the details, but the main answer text downplayed it.

What Successful AI-Visible Businesses Share

Across all 20 prompts, certain businesses appeared repeatedly:

• Roto-Rooter appeared in nearly every ChatGPT plumber query
• Radiant Plumbing showed up in multiple Google and ChatGPT answers despite mixed reviews
• Evans Family Law Group appeared across multiple lawyer queries on both engines
• Emergency Dentist of Austin repeated in dental queries

Shared characteristics among these repeatedly-recommended businesses (in this sample): 1. Appeared on multiple directories (Google Business Profile, Yelp, Avvo, Expertise.com) 2. Displayed credentials, ratings, and reviews on their sites and directories 3. Had service-specific pages matching searched queries ("emergency plumber," "walk-in dentist," "divorce lawyer free consultation") 4. Were mentioned in local news or appeared in sources the engines cited 5. Stated hours, services, and availability clearly on their own sites

What This Sample Suggests

Based on 20 prompts across 2 engines and 4 verticals:

1. Be Indexed and Crawlable

Recommended businesses in this sample had indexed, navigable web content. We did not test blocked or outdated sites.

2. Build Credentials and Third-Party Validation

Businesses appearing repeatedly in recommendations across both engines displayed: - Author/professional bio with stated credentials - Third-party mentions (awards, professional recognitions, local media) - Years in business visible on their website - Specific, documented case results or portfolio examples

We cannot confirm from this sample whether these factors caused the recommendations or coincided with them.

3. Directory Presence

Recommended businesses in this sample appeared on multiple directories — Google Business Profile, Yelp, Avvo, Expertise.com, SuperLawyers, or industry-specific listings. We did not systematically compare recommended businesses against equivalently-credentialled businesses with fewer directory listings.

4. Service-Specific Content

Businesses with explicit service-specific pages ("emergency plumbing available 24/7," "walk-in dentist available Saturdays," "free divorce consultations") appeared in this sample's recommendations. We did not compare identical businesses with and without service-specific pages.

5. Specific, Verifiable Facts vs. Superlatives

Claims like "#1 plumber" or "recovered $100M" appeared in sources both engines read. Specific, verifiable facts ("available 24/7 with same-day service," "board certified since 2004," "flat-fee consultations starting at $X") were displayed by recommended businesses in this sample.

6. Local Media Mentions

Businesses mentioned in local news (Austin American-Statesman appeared in multiple answer sets) were present in recommendations. We did not compare recommended businesses against unmentioned businesses with equivalent credentials.

Critical Limitations of This Sample

Methodology constraints: - This is a 20-query spot-check, not a statistically representative sample of thousands of searches - Google AI Mode required manual selection for 14 of 20 queries; auto-trigger rates may differ by device, location, personalization setting, or date - Key asymmetry: ChatGPT's queries were run from the operator's UK location (device location not overridden), while Google's were set to Austin, TX. This geographic difference affects how location-dependent business recommendations appear; results are not fully equivalent - Both engines' citation behaviors, business rankings, and source weighting change daily and may differ by user account, geography, and search timing - The specific sources, directory versions, and business pages cited in this July 2026 sample may no longer appear identically

Run your own tests: Search for your own professional category in your own location on both platforms. Check which businesses appear, what sources they cite, and whether you're visible at all — even if not named in the recommendation, you may be influencing the answer through background research.

What We Observed

Both AI engines read far more sources than they surface in recommendations. Both cited sources with geographic mismatches, freshness problems, and category inconsistencies. For subjective queries in this sample, they pulled from different sources and recommended different businesses. For factual queries, they converged more.

Patterns observed in this 20-query sample: 1. Recommended businesses appeared on multiple directories (Google Business Profile, Yelp, Avvo, Expertise.com, industry-specific listings). 2. Recommended businesses displayed credentials, awards, and third-party validation on their own sites. 3. Businesses with explicit service-specific pages ("open Saturdays," "flat-fee consultations," "24-hour emergency service") appeared in recommendations. 4. Claims like "#1 plumber" or "$100M recovered" appeared in sources both engines read. 5. Businesses mentioned in local news appeared in recommendation sets.

Both engines expose their sources in visible citations and source panels. You can see which sources each engine cited. Audit where you're appearing, check your directory presence, and ensure your own website contains specific, verifiable details about your services and availability.

This research analyzed 20 identical queries across ChatGPT (web version, UK geolocation) and Google AI Mode (Austin, TX geolocation), with detailed source attribution and citation logging. All findings are based on direct observation of what sources each engine cited in its visible sources panel and what it recommended. Specific quotes and citations are preserved exactly as they appeared in each engine's answer and sources panel. This is a 20-query sample, not a statistically representative survey. The geographic asymmetry (UK for ChatGPT, Austin for Google) affects location-dependent recommendations and means results are not fully equivalent.

Get your AI Visibility audit →