How we measured what AI cites for local recommendations
The full method behind the figure that 73.8% of AI local-recommendation citations go to the business’s own website. Written so a journalist or researcher can judge it, repeat it, or disagree with it.
The sample
We ran 176 buyer-intent queries across 22 US local-service markets and recorded every source the AI answer cited. That produced 1,189 individual citations, which is the denominator for every percentage on this page.
Date of the run: 20 August 2026. AI answers change. Treat every number here as a snapshot of that date, and quote the date with the number.
How the queries were built
For each market we generated diverse buyer-intent prompts for one service in one city, phrased the way a person actually asks: who is the best, can you recommend one, what are my options near me. We did not use keyword strings, because that is not how people talk to an answer engine.
Every prompt was then run through two engines that expose their sources:
- OpenAI Responses API with web search enabled, which is the ChatGPT-equivalent answer path.
- Google Gemini with google_search grounding, which is the Google AI-equivalent answer path.
Every cited domain in every answer was extracted and counted. A domain cited by two engines counts twice, because it was cited twice.
The result
| Source type | Citations | Share |
|---|---|---|
| Individual business websites | 877 | 73.8% |
| Google Maps | 143 | 12.0% |
| Directories | 94 | 7.9% |
| Editorial or aggregator | 60 | 5.0% |
| Social | 10 | 0.8% |
| News | 5 | 0.4% |
The directory share is the number most people find surprising. Across the four largest home-service directories the counts were Angi 28, HomeAdvisor 27, BBB 18 and Thumbtack 11, which is 84 citations combined. Google Maps alone was cited 143 times, more than those four put together.
The correction we made against ourselves
Our first pass put every non-directory commercial domain in the business-website bucket, which produced a headline of 78.6%. That was too generous to our own argument.
Eight domains were reclassified out of the business bucket because they are aggregators or review platforms rather than an individual business’s own site: consumeraffairs.com, expertise.com, forbes.com, homeguide.com, ifoam.com, toprateddentist.com, yelp.com and yourdentistryguide.com.
That correction lowered our own headline from 78.6% to 73.8%. The lower figure is the one we publish, and the one to cite.
What this study does not show
- It counts citations, not clicks or customers. Being cited is not the same as being chosen, and we did not measure what anyone did next.
- Two engines, not all of them. Perplexity, Copilot and Claude were not included. Do not read this as a claim about AI search in general.
- Local service verticals only. The markets were service businesses such as dentists and contractors. It says nothing about retail, software or national brands.
- One snapshot, not a trend. A single run on one date. We have not re-run it often enough to claim a direction of travel.
- 22 markets is a sample. It is enough to show a strong pattern and not enough to describe every US metro.
Citing it, and checking it
The study is ours and you are free to cite it. Please include the sample size and the date, for example: “176 AI answers across 22 US markets, Honey Badger SEO, August 2026”.
If you want to interrogate it rather than quote it, write to [email protected]. We will send the market list, the category rules, and the per-market answer files we hold. The measurement tool is ours and the study can be re-run on request, including on markets you choose rather than markets we chose.