Research

ChatGPT and Perplexity agreed on only 1 in 4 local business recommendations

By Ravinder Cheema · September 2026 · Raw data and method at the end

I run a small SEO agency in Bakersfield, California. Over the past year, the most common question I get from clients has changed from "why aren't we ranking on Google?" to "does ChatGPT recommend us?" I didn't have a good answer, so I tested it.

I asked ChatGPT and Perplexity the same simple question, "best {industry} in {city}, CA," for four industries in five California cities, and recorded the top five businesses each tool named. I expected the two tools to mostly agree. They didn't.

The short version

• Across 20 city-and-industry searches, Perplexity named 99 businesses in its top fives. Only 24 of those (24%) also appeared in ChatGPT's top five for the same search.
• The two tools picked the same #1 business only once in 20 searches (plumbers in Fresno).
• They appear to read different parts of the web. About half of Perplexity's picks were attributed to the businesses' own websites, while ChatGPT leaned heavily on third-party "best of" directories and review aggregators.
• The same tool gave different answers depending on whether I was signed in.
• A vague query ("best lawyer") produced directories, not businesses.

What I did

On September 26, 2026, in one session, I ran the prompt best {industry} in {city}, CA in both tools, using a new chat for every prompt.

• Cities: Bakersfield, Fresno, Sacramento, Los Angeles, San Diego
• Industries: dentist, plumber, personal injury lawyer, landscaping
• Tools: ChatGPT (signed in, web search on) and Perplexity (signed in, default search)
• Recorded: the first five businesses each tool named, in order, plus the source each answer credited, and any ratings the tool quoted

That's 40 top-five lists and 198 business mentions. Two lists had only four businesses, which is why the totals are 99 per tool rather than 100. A business counts as "named by both" when the same business appears in both tools' top fives for the same city and industry.

Overlap by city and industry

Each cell shows how many of Perplexity's top five also appeared in ChatGPT's top five.

Two things stand out. First, smaller markets agreed more. Bakersfield, the smallest city in the set, had 10 overlaps; San Diego had 1. My guess is that a smaller market simply has fewer plausible candidates, so any two lists are more likely to collide. Second, dentists and personal injury lawyers, the two most competitive and heavily marketed categories, had the least agreement.

The #1 spot almost never matched

If you only look at the first business named, which is the one most people will act on, the tools agreed once in 20 searches. For a business owner, that means "being recommended by AI" isn't one thing. You can be the top answer in one tool and absent from the other.

They seem to read different parts of the web

Both tools show the sources behind their answers, so I recorded which source each recommendation was credited to. The pattern was consistent:

• Perplexity most often credited the business's own website (roughly half of its 99 picks), followed by local listings and Yelp.
• ChatGPT most often credited third-party lists: "best of" directories, review-widget pages and aggregators such as ConsumerAffairs, Expertise, Avvo, Birdeye and Yably. Around a quarter of its picks came with no clear single source.

That would explain a lot of the disagreement. If one system is effectively asking "which businesses describe themselves well?" and the other is asking "which businesses appear on lists about this city?", they'll return different names even when both are working as designed.

Some of this is consistent with what the companies say publicly. OpenAI's help documentation says ChatGPT search sometimes sends rewritten queries to third-party search providers, and can use approximate location and saved memories when rewriting a query. Perplexity describes itself as searching the web in real time and citing its sources. Neither publishes how local businesses are ranked, so the rest is inference from the citations.

Reviews decided some categories almost completely

For dentists and plumbers, the order of recommendations tracked star ratings and review counts closely. In one Los Angeles plumbing answer, Perplexity's first four businesses were in descending order of review count: 1,832, 1,097, 780 and 773. Note that these ratings were quoted by the AI tools; I didn't verify them against the review platforms.

Directories decided lawyers

For personal injury lawyers, most picks came from legal directory profiles rather than law firm websites. In Fresno, Perplexity's first four picks were all credited to the same legal directory. For that category, a firm's own website seemed to matter less than its directory presence.

Signing in changed the answer

I also ran "best dentist in Bakersfield, CA" in Perplexity while signed out, and compared it with the signed-in answer. The signed-out answer named three practices; the signed-in answer named three completely different ones. This matches OpenAI's statement that ChatGPT can use memory and location when searching, and suggests there's no single "AI ranking" to measure. Two people asking the same question can get different businesses.

Vague questions returned no businesses

"Best lawyer in Bakersfield, CA" returned directories and advice about how to choose a lawyer, not firm names. The tools only named businesses when the question included a specific service. For the local businesses I work with, that argues for pages about specific services rather than general "we do everything" pages.

Small sources had outsized influence

Reddit threads, a Yahoo Local roundup, Thumbtack, and review widgets all placed businesses into answers. In San Diego, a single roundup article appeared to decide Perplexity's #1 plumber. That's a low bar, and an easy one to game, which I expect the tools will try to fix.

Limitations

This is a small, single-session snapshot, not a controlled study. Please read it that way.

• One run per prompt. LLM answers vary between runs. I didn't repeat prompts to measure that variance, so some disagreement between tools is likely just noise within each tool.
• Signed-in accounts. Both tools were used signed in, from one location, by one person. Personalization and location may have shaped the results.
• Name matching. I matched businesses by name. Slight naming differences could hide a few true matches, so the overlap figure is more likely too low than too high.
• Source attribution. I recorded the source each tool displayed. That shows what the tool credited, not necessarily what drove its choice.
• One prompt wording. Different phrasing ("top rated", "near me", "affordable") would likely change the results.
• Small sample. 5 cities, 4 industries, 2 tools, one day. Treat the percentages as directional.

What I'd like to test next

• Run each prompt 5 to 10 times per tool to separate within-tool variance from between-tool disagreement
• Compare signed-in and signed-out results systematically, and from different locations
• Add Google AI Mode and Gemini
• Repeat monthly to see how stable the recommendations are over time

If you've run similar tests, especially on run-to-run variance, I'd like to compare notes.

Data

The full dataset (all 198 mentions, with rank, tool, city, industry, quoted rating and credited source) is available as a spreadsheet download (.xlsx). You can also explore the answers side by side, by city and industry, in the interactive version of the index.

Ravinder Cheema runs Clicks Dynasty, a small SEO and digital marketing agency in Bakersfield, California. ChatGPT is a product of OpenAI; Perplexity is a product of Perplexity AI. This study is independent and not affiliated with either company.