57% of US local markets had their #1 AI recommendation change in a single week

Every Tuesday findloc.ai asks ChatGPT, Perplexity, Claude, and Gemini 15 real customer-search questions per market — "best hotels in Chicago", "top-rated dentists in Austin", "recommended HVAC company in Phoenix", etc. Across 100 US MSAs and 7 industries, that's 690 markets and about 41,400 queries per week.

This past week (2026-W27) we compared each market's #1-cited business against the prior week (2026-W26). The AI's top recommendation had changed in 391 of the 690 markets — 57%.

Volatility varies sharply by vertical. Higher-consideration, higher-price categories are more stable — probably because there is more written content per business, so the models have thicker evidence to keep naming the same candidates. Impulsive or price-competitive categories flip more:

One takeaway that surprised us: not a single industry stayed below 40% volatility. Even wedding venues — a decision people research for months — showed the AI shuffling its #1 recommendation in 43 of 100 markets in one week.

Aggregate percentages are abstract. Here are specific businesses that were the top-cited answer across all four AI engines in their market last week and are no longer in the top 20 this week. Nothing about these businesses changed in the physical world — the AI just started naming other names.

• InterContinental Miami — Miami hotels — was #1 with 20 citations across the four engines last week. Zero citations this week.
• Hotel Nikko San Francisco — SF hotels — dropped from #1 to #15.
• Moonrise Hotel (St. Louis) — St. Louis hotels — from #1 to unranked.
• BEST WESTERN Oxnard Inn — Oxnard hotels — from #1 to unranked.
• The Tinsmith (Madison, WI) — wedding venues — from #1 to #27.
• Cooper Hurley Injury Lawyers (Virginia Beach) — personal-injury lawyers — from #1 to #18.
• POSH Medical Spa (Reno) — med spas — from #1 to #15.

Three overlapping mechanics, in decreasing order of impact:

1. Every answer is a fresh sample. Google shows a ranked list from an index that changes at hours or days of latency. ChatGPT and friends synthesise a new answer per query from a mix of live-search results and training-data recall. The sampling step is not deterministic — even the same query 20 minutes apart can produce different candidates.
2. Small query wording changes pull different candidates. "Best hotels in New York" and "top-rated hotels in Manhattan" hit different indexed passages. The 15 curated queries we ask per market include wording variations on purpose — a business that only surfaces for one wording will show up in fewer of the 15 queries and score lower.
3. Search-tool re-ranking is opaque and non-deterministic at the top. The four engines all use some form of retrieve-then-rerank. Small changes in the retrieved-passage set can flip the top 3 businesses without changing the top 10. From the customer's point of view, the top 3 is all that matters.

The cumulative effect is that "being #1 last week" is not a defensive position. It is a lucky sample of a distribution.

Some businesses in our data ARE stable — the ones cited by 3 or 4 engines rarely drop out entirely. Comparing them against the volatile ones, three patterns keep showing up:

1. Optimise for multi-engine coverage over peak position. A business at rank #5 in three engines is more durable than a #1 in only one engine. When the AI reweights, coverage-breadth survives; single-engine peaks disappear.
2. Maintain persistent citations on the directories the engines actually crawl. Yelp, Yellow Pages, BBB, and TripAdvisor for hospitality; Justia and Avvo for lawyers; OpenTable for restaurants; industry-specific directories for HVAC, medspa, wedding venues. Every engine crawls a subset of these; the more directories you're listed on, the more engines will have you in their retrieval set.
3. Publish structured content — schema.org markup, an /llms.txt file, FAQ pages. This is not about "SEO" — it's about giving each engine unambiguous facts. LocalBusiness schema tells the engine your address without parsing HTML; FAQPage schema surfaces answers to specific questions; /llms.txt gives an AI a one-fetch summary of who you are.

We published a live report for every business the AI has named in the last three weeks — 5,395 businesses in total across our 700 tracked markets. Each report shows:

• Which of the 4 AI engines currently name your business, and how often.
• Which engines do not name you at all — the coverage gap.
• The top competitors in your market and how many engines name them.
• Historical rank across the weeks we have tracked.

Reports are free and public. Start from findloc.ai/directory to browse tracked businesses by industry, or open findloc.ai/ai-visibility to browse by market.

If you run a US local business we do not track yet, findloc.ai/ai-visibility/us/[your-city]/[your-industry] shows the market-level rankings. Reply to any of our tweets or email support@findloc.ai and we will add your industry to next week's probe.

We are publishing this openly so anyone can verify. If you spot a bug in the count, tell us — we will fix and reissue.

1. Query set: 15 hand-curated customer-search phrasings per (city, industry). Phrasings deliberately vary intent ("cheapest", "best", "family friendly", "near me") to sample the full retrieval distribution.
2. Engines: openai/gpt-4o-mini-search-preview, perplexity/sonar, anthropic/claude-3.5-haiku (via web plugin), google/gemini-2.0-flash. All four are the search-augmented variants — this is intentional. Non-search variants would test the training snapshot only.
3. Extraction: response text is passed through a business-name extractor (also an LLM) with per-industry filters that reject service names ("nitrous oxide", "contingency fee"), directory brands (CareCredit, HealthGrades, HomeAdvisor), and generic terms.
4. Ranking: each business gets a score of (citation count) x (engines seen in). Ties broken by raw citation count. #1 in this post refers to the highest-scoring business for the market that week.
5. Volatility metric: for each (city, industry), we compare the #1 name in W26 vs W27. The 57% figure is the fraction of markets where those two names differ. We only counted markets that had a valid #1 in both weeks — a handful had W26 data but not W27 or vice versa.

The raw week-by-week data is available at findloc.ai/data as CSV. We run the whole four-engine pipeline as a lean batched-probe architecture — a technical writeup of that is on our roadmap.

A few open questions we are still figuring out. If you have a strong view, DM us:

• Are 15 queries per market enough signal? We tried 25 in an earlier version — the extra 10 barely changed the top-10 rankings but doubled cost. Push back if that intuition is wrong.
• Which industries should we add next? We are looking at auto repair, plumbing, chiropractic, veterinary. Tell us which one your customers ask AI about.
• Should we track brand-name queries too ("what do people say about [specific hotel]") in addition to category queries ("best hotels in [city]")? Brand queries measure reputation; category queries measure discoverability. Both matter for different reasons.

And if you run a small business and want to see this data for your own name — findloc.ai/directory has the searchable index. Everything on that page is free forever.