How we measure what AI recommends
AI Syndicate Local exists to make one invisible channel measurable: when a customer asks an AI assistant who to hire, who gets named. Everything published in the index comes from the monthly pipeline below — recorded, not curated.
The pipeline, step by step
8 rules, applied identically to every city and category, every month.
Fix a question set per metro and category
Each measurement cell — one metro crossed with one service category — gets a small, fixed set of consumer-phrased questions (for example, "Who are the best plumbers in Tampa, FL?"), with phrasing variants for recommendation and price intent. The set is frozen: movement over time reflects the assistants changing their answers, never us changing the questions. Every ranking page lists its own question set verbatim.
Put identical questions to all 12 assistants
The same wording goes to every assistant we can reach: ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Microsoft Copilot, Siri, Alexa, Grok, Meta AI, DeepSeek, Mistral. ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, Microsoft Copilot run in live, web-connected mode — the configuration a real customer reaches. Siri and Alexa answer through the voice stack. Grok, Meta AI, DeepSeek, Mistral answer from model knowledge only, because no consumer live-retrieval API is available for them; their answers are recorded and labeled as such, never presented as evidence of what the web says today. An assistant that gives no usable answer for a city and category in an edition is left out of that edition's count rather than counted as not naming anyone, which is why a ranking can read "named by 5 of the 10 that answered".
Record every answer, append-only
Every run is stored as raw data — the full answer text and the sources the assistant cited. Runs are never deleted, and each edition is derived from the accumulated history, so any figure on the site can be traced back to the runs that produced it.
Resolve names to real businesses
Assistants name the same business in different ways — with and without an LLC suffix, with a location qualifier, with a franchise prefix. Name variants are folded into one canonical entity so counts never double-credit a business for spelling differences.
Drop what can't be verified
A name that appears in only a single run and resolves to no verifiable business is treated as a probable hallucination and excluded from rankings. A business must be named in at least two independent runs to be listed. Ranking pages disclose how many names were dropped this way.
Link each business to its own site, or to nothing
A ranked business's website is found in one of three ways: an assistant cited or wrote a domain that matches the business name; a business-listing lookup returned that exact business by name, in the right state, with its site; or a domain spelled from the business's full name. None of these is trusted on its own. Before a link is published we load the page and check that it names the business — and a site found from the name alone must also place the business in the right city or state, so a same-name company elsewhere is never linked. A link that stops passing is removed. A wrong link would send a company's customers to a competitor, so a business we cannot confirm is shown with no link at all. Every page the assistants cited is published alongside the ranking too, whether or not it belongs to a ranked business.
Rank by measured consensus
A business's rank in a cell reflects how often it was named across the edition's runs and how many distinct assistants named it. A name four engines repeat is a different signal from one that surfaced once, and the per-row detail shows exactly which assistants named each business.
Enforce a sample floor
A cell publishes a ranking only after at least 10 usable answers accumulate in an edition (one answer is one assistant answering one question once), so the ordering rests on more than a couple of replies. Below the floor the page says "recording" — an honest gap rather than a confident-looking ranking built on noise. Each ranking page states its own sample: how many answers, from how many assistants, on how many question phrasings.
Editions, files and how to cite them
The terms every page, file and download uses, defined once.
Edition
Rankings publish monthly. Each edition is named by the ISO date it opens (for example 2026-10-01) and covers the answers collected that month.
Assistants asked and assistants that answered
All 12 assistants are asked every question. "Named by 5 of 10" means 10 gave a usable answer for that city and category and 5 of them named the business; when every assistant answers it reads "of 12". The data files carry both numbers (assistantsAsked, assistantsAnswered).
Revisions
The latest edition can still be revised until the next one publishes, as late answers arrive or a correction is made. The site and the data files always show the latest edition.
Files
The latest edition is at https://www.aisyndicate.com/ai-local/data.json, data.jsonl (one ranking row per line), data.csv and all.md. Every page also has a Markdown twin, and https://www.aisyndicate.com/ai-local/llms.txt is a short guide for AI assistants.
Licence and citation
The data is CC BY 4.0: reuse it with credit. Cite as: AI Syndicate Local. AI Syndicate Collective LLC. https://www.aisyndicate.com/ai-local/ (CC BY 4.0), and name the edition date with any figure you quote.
What we don't do
We don't sell placement
No business can pay to appear, rank higher, or have a bad month removed. Rankings are derived from assistant outputs alone, and the derivation is described on this page in full.
We don't cherry-pick runs
Every recorded run enters the history, including editions where the assistants disagree or reverse themselves. Movement you see is movement that happened.
We don't pad thin data
Cells that haven't cleared the sample floor say "recording." An invented ranking would be easier to ship and worthless to trust.
Corrections and removal
If a listing misidentifies your business, or you want your business excluded from the index, email support@aisyndicate.com and we'll resolve it — the record keeps the raw runs, but the published pages will reflect the correction.
Want this measurement on your own business?
The index measures whole markets monthly. Our platform runs the same measurement on a single business daily — with the levers to move it.
AI Pulse: all 12 AI assistants, every feature, cancel anytime.