Every number we have published, in one place. This is the complete statistical record of Steady Demand's research into AI search, Google Business Profile and Local Services Ads – 175 figures drawn from 21 studies, tests and client case studies published since December 2024.
The work spans large-scale AI citation analysis (1,487 local-service queries run through Gemini and ChatGPT, 14,472 citations classified), controlled GBP tests, LSA performance benchmarks, and results from client engagements. Every stat links back to the study that produced it, where you will find the methodology and sample size. Figures are recorded exactly as published – nothing rounded, nothing restated.
AI Citation Ledger — What AI Actually Cites
Published Aug 6, 2026
1,487 local-service queries run through Gemini and ChatGPT across 50 US metros and 10 verticals, with 14,472 citations classified. The largest look yet at where AI answers about local businesses actually come from.
Study scale
1,487queries completed (of 1,500 targeted)↳ The wider field: Profound classified 27 million citations across ChatGPT, Gemini and AI Overviews
14,472citations collected and classified50US metros × 10 service verticals3phrasings per metro/vertical combination3,611unique domains cited4,410distinct businesses named by GeminiJul 27–Aug 20, 2026collection windowGemini API + ChatGPT comparison passesWhere citations point (Gemini)
59.9%citations to the business's own website14.4%citations to social / forum sources13.7%of all citations go to reddit.com10.4%citations to directories10.3%combined local-service directory share
Ratings & reviews of AI-recommended businesses
4.75★average rating of AI-recommended businessesvs 4.84★ for the plain-search baseline90%of AI picks rated 4.5★ or higher97%of AI picks rated 4.0★ or highervs 99.6% of baseline businesses919average review count for AI picksvs 846 for baseline; medians 295 vs 32235%of AI picks have 500+ reviews2.0%of businesses got recommended in more than one metro87 businesses total93%Google Places match rate for named businesses
- Main report: What AI Actually CitesAug 6, 2026 (upd. Aug 20)
- Methodology & study designAug 6, 2026
- Interactive dashboardAug 16, 2026
- Raw dataAug 16, 2026
Grounding Drift — AI Citations Change Every Time
Published Aug 6, 2026
Byte-identical queries repeated back-to-back, hours later and the next day. AI answers are not a ranking you can hold; the classic local pack is roughly ten times more stable.
Repeat-query consistency
46%best-case domain overlap on immediate repeatJaccard 0.463↳ Others found the same instability: half of cited content is under 13 weeks old
0.056Jaccard overlap of the model's internal search strings0.410domain overlap after ~3.5 hours0.380domain overlap next day0.265three-repeat agreement rate within one day72repeat-test calls across 4 rounds over 2 daysAI vs the classic local pack
7.9%Gemini returns the same top business on repeat90.2%Google local pack returns the same top listing0.177Gemini loose full-set overlapvs 0.884 for the local pack5,000pairwise Gemini round comparisons4,466 local-pack comparisons
Phrasing sensitivity
40%average source agreement across three phrasings of the same question3% → 87%word-overlap range between phrasingsworst: HVAC/Birmingham · best: Locksmith/Cleveland12.7%overlap of Gemini fan-out queries with Google's People Also Ask80%of fan-out queries score below 0.2 lexical overlap with PAA
External benchmarks cited
Gemini vs ChatGPT — Two Different Internets
Published Aug 6, 2026
The same 1,487 queries through both engines. They almost never agree on sources, and ChatGPT's source mix changed overnight on August 20, 2026.
Agreement between engines
8.3%average cited-domain overlap between Gemini and ChatGPT↳ Semrush found the same split – Gemini names brands 83.7% of the time, ChatGPT 20.7%
4.2%of queries where both name the same top business0.037overlap across the full recommended-business lists0.5% – 14.9%domain overlap range by verticallowest: personal injury · highest: auto repairChatGPT's source mix before → after Aug 20, 2026
41.7% → 0.0%reddit.com citation share for local queries15.9% → 42.8%citations to business websites34.6% → 40.8%citations to general directories3.0% → 11.6%citations to local-service directories (Angi, HomeAdvisor)86.4%decline in ChatGPT's Reddit citations overallPromptwatch tracking
Reddit reliance by vertical (July 2026, pre-change)
72.2%dentist queries60.1%auto repair queries12.3%personal injury lawyer queries60%of July queries mentioned Reddit in the answer text899 of 1,500
Verification pass
334citations checked post-change — zero from Reddit150 queries across 5 metros6.7%Reddit share in a same-day 40-query Gemini sample19 of 284 citations1,650total post-drop ChatGPT queries checked
- Gemini vs ChatGPT local citations comparedAug 6, 2026
AI Overviews vs AI Mode — Opposite Citation Behavior
Published Aug 22, 2026
8,000 local-service searches across 50 metros and 10 home-services verticals, 86,645 citation references analyzed. The two Google surfaces reward almost opposite things.
Study scale
7,990local-service searches (≈4,000 per surface)↳ Ahrefs measured what that costs: 58% lower click-through at position one
86,645citation references analyzed~3,995usable queries per surface3,992 AI Overviews · 3,998 AI ModeHow often each surface appears
48.8%AI Overviews surface rate overallas observed by third-party collection — a manual browser check found AI Overviews substantially more often100%AI Mode surface rate overallnear-definitional — a dedicated AI Mode endpoint answers by design21.8%AI Overviews on local-intent queriesAI Mode: 100%. Collection-method rate; see study Finding 0175.8%AI Overviews on informational queriescollection-method rate; see study Finding 01
What each surface cites (local intent)
70.7%AI Overviews → business websites12.4%AI Overviews → social sources79.8%AI Mode → Google Maps12.8%AI Mode → business websites42.3% / 22.7%AI Mode informational: business sites / Google's own properties95.5%cited domain matches the GBP-listed website6,006 citations — 19% of all business-site citations88.0% / 43.4%AI Mode → Google Maps: highest vertical (dentist) / lowest (personal injury lawyer)varies far more by vertical than by metro71.2%–84.8%AI Mode → Google Maps: range across all 50 metrosevery metro above 70%
Traits of cited pages — AI Overviews vs AI Mode
79.3% / 74.0%schema markup presentbusiness-site citations only, one count per page94.8% / 95.1%viewport meta tag present — no difference between surfacescorrected 2026-09-09; the earlier 87.2%/81.9% gap was a page-composition artifact72.8% / 69.1%page contains statisticsbusiness-site citations only, one count per page78.9% / 74.4%visible phone numberbusiness-site citations only, one count per page39.2% / 34.3%linked social profiles4.77★ / 4.75★average star rating of cited businesses412 / 450median review countmeans were 1,129 / 1,266, skewed by national brands21.6 yrsmedian domain age — identical on both surfaces
Published Aug 21, 2026
Intent is the driver
58.3%ad presence on local-intent queries1.44× the 40.5% general-commercial control23.8% vs 58.0%naturalistic vs templated phrasing≈2.4× difference54.3% vs 30.8%normal markets vs small towns
Ad presence by category
75.8%dentists70.8%hotels68.3%lawyers68.3%restaurants67.5%plumbers63.3%florists53.3%museums52.5%bookstores44.2%furniture stores18.3%auto repair
Ad quality
ChatGPT's Local Results Aren't Coming From Foursquare
Published Aug 21, 2026
2,880 prompts across 12 verticals and 12 US markets, plus a 1,728-prompt cross-check on a second API surface, to test the widely repeated claim that ChatGPT's local data is Foursquare's.
Foursquare's actual share
0.00%Foursquare citation share on the primary surface↳ Local Falcon found 83% of restaurants absent from ChatGPT entirely
0.06%Foursquare share on the independent cross-check surface7.5% vs 5.7%ChatGPT business resemblance to Google vs Foursquare87.5%of downstream repetitions of the original claim stripped all caveatsoriginal May 2025 study: 60–70%, five citiesWhat is actually in the grounding layer
95.83%Yelp present in business-card groundingbut only 1.04% cited, 2.08% retrieved-not-cited28.90%TripAdvisor in business-card grounding13.55% cited · 20.70% retrieved16.88%OpenTable in business-card grounding1.70% cited · 3.58% retrieved4,608total prompts across both surfaces72-hour window, US-English
LSA Performance Benchmarks
Published Apr 18, 2026
What separates accounts generating leads from accounts that don't, from managed-account performance data.
Levers that move lead volume
+43%more leads when message leads are enabledand +19% impressions↳ Sterling Sky tracked LSA visibility rising from 11% to 31% of queries in a year
53.7average leads for accounts at 85%+ responsiveness+77% lead volume, +52% impression rate141.8average leads for accounts hitting their budget capover 4× non-constrained accounts0leads for accounts with fewer than 4 photosCustom Services on a Google Business Profile
Published Dec 15, 2024
A single-profile test of adding custom services, tracked from March 2023 through December 2024.
Result
Local Services Ads — Client Results
Managed and whitelabel LSA accounts, with the numbers each case study reports.
Cleaning company — 2 years
5,586total leads, Sept 2023 – Nov 2025$174,109total spend~$31blended cost per lead$29.29 at launch → $29.85 at end349,182total ad impressions99.5% / 54.3%top impression share / absolute top rank at peak99.03% / 49.21% in Nov 2025190 → 218monthly leads, first month to last
National HVAC franchise — 4 years
12 → 68active LSA accounts0leads, calls or billable activity at handoff$5K–$46Kmonthly spend per brand3verticals: HVAC, plumbing, electrical
Moving company — 9 months
#12 → #1rank placement37 → 147monthly calls at peak season300%+ call volume increase$38–$39cost per lead, stabilized97–98%top impression share272 → 290reviews
Garage door client — 12 months
22 → 56–62monthly leads, baseline to peak60+ at peak$40.97–$49.63cost per lead range2,921 → 6,340ad impressions407 → 574reviews; rating 4.1 → 4.398–99%top impression shareholding ~#4 rank
- Cleaning company — launch to 5,586 leads in 2 yearsSep 2023 – Nov 2025
- National HVAC franchise — 12 to 68 LSA accounts4 years, ongoing
- Moving company in a saturated market9 months, ongoing
- Garage door client — escaping LSA verification limbo12 months, ongoing
GBP Suspensions, Spam & Recovery
Reinstatements, spam removal and ranking recovery work, with the reported numbers.
Spam removal — Dallas–Fort Worth
~500fake listings in the initial sweep100% success rate within 30 days1,200+spam listings removed in under a year~100/month sustained~3×increase in calls from the client's listingheld top three in the map pack
Account-level suspension recovery
48 hoursto reinstate all but four locations4 days total−50%revenue vs the previous week during suspension−70%revenue vs the same period the prior year
Repeat suspensions — garage door
17 daysfirst reinstatement (Jan 7–24, 2020)$2,000/daylost revenue while suspended$34,000 total from the first suspension3suspensions in one yearlater reinstatements took 1 and 2 days4 of 16audit points out of compliance
Ranking recovery & franchise fixes
#15 → top 4ranking restored after a review transfer went wrong~1,800 reviews moved between locations72 hoursto visible recovery17%call-in orders at a mis-chained franchise locationvs 40–55% at sibling locations
Local SEO boost — Barbara Oliver Jewelry
AI Citation Ledger — September 2026 Re-run
The identical design re-run six weeks later: 1,500 queries, 10,229 citations. Reddit fell out of both engines, the directories took the share, and one of the original findings reversed outright. July 2026 stays as the frozen baseline; these are the September figures, not a correction to it.
Study scale
1,500queries completed (of 1,500 targeted)July wave: 1,48710,229citations collected and classifiedJuly wave: 14,4726.8sources cited per query, on averagedown from 9.7 in July2,474unique domains citedJuly wave: 3,6113,738distinct businesses named by GeminiJuly wave: 4,410Sept 6–9, 2026collection windowGemini via Vertex AI + ChatGPT comparison passes
Where citations point (Gemini)
49.6%citations to the business's own websiteJuly: 58.8%19.1%citations to local-service directoriesJuly: 10.4%17.7%citations to general business directoriesJuly: 10.5%3.9%citations to social / forum sourcesJuly: 14.4%1.7%of all citations go to reddit.comJuly: 13.7% — the single most-cited domain in the July study
Where citations point (ChatGPT)
46.0%citations to general business directoriesJuly: 39.1%27.2%citations to local-service directoriesJuly: 6.7%10.5%citations to the business's own websiteJuly, restated after the Sept classifier re-audit: 2.9%0.0%of ChatGPT citations go to reddit.comJuly: 41.7% — a complete removal, first measured Aug 20
Ratings & reviews of AI-recommended businesses
4.76★average rating of AI-recommended businesses vs 4.82★ for the plain-search baselineJuly: 4.75★ vs 4.84★91%of AI picks rated 4.5★ or higherJuly: 90%98%of AI picks rated 4.0★ or higher vs 99% of baseline businessesJuly: 97% vs 100%860average review count for AI picks vs 850 for baseline; medians 307 vs 311July, restated after a sampling fix: 906 vs 84636%of AI picks have 500+ reviewsJuly: 35%92%Google Places match rate for named businessesJuly: 93%
Repeatability and drift
43.0%citation overlap when the identical question is asked twice the same day497 queries, 10 round pairs42.7%citation overlap a day later — statistically identical to the same-day figuredifference +0.2 pts, 95% CI -0.9 to +1.2. July's next-day decay did not reproduce41.2%agreement when the same question is reworded three waysJuly: 40.3% — rewording costs about as much as simply asking twice83.3%of the time Google's local pack returns the same top business on a repeatJuly: 90.2% — the control moved too5.1%of the time Gemini names the same top business on a repeatJuly: 7.9%
Gemini vs ChatGPT
8.3%of cited domains match when both engines get the identical questionJuly: 8.3% — unchanged to the decimal4.9%of the time the two engines name the same top businessJuly: 4.2%783queries where both engines named a recognisable businessJuly: 1,099. Gemini now returns no business name at all on 47.4% of queries, up from 25.8%
Concentration
Download all figures as CSV – free to cite with attribution to Steady Demand and a link to the source study.
For third-party research on the same questions, see the Local Search & AI Industry Research Index.
Data last updated September 14, 2026
Download all figures as CSV — free to cite with attribution to Steady Demand and a link to the source study.
