How Perplexity Picks Local Business Recommendations (Reverse-Engineered)

Perplexity is not a single model with an opinion. It is a retrieval pipeline wrapped around a router, and understanding that pipeline explains most of what follows in this article. When you ask a question, Perplexity does not just query one search index and summarize the top results the way a wrapper around ChatGPT Search might. It runs a query-planning step (frequently branching one question into several sub-searches), fetches a wider set of candidate pages than it will end up citing, scores them, and hands a filtered set to a language model — commonly its own Sonar models, tuned specifically for answering from retrieved sources rather than from parametric memory.

The default experience most people use is Copilot mode (called Search on the free tier), which asks clarifying questions and shows its source list openly, right next to the answer, rather than burying citations in footnotes. That transparency is useful for our purposes: it means you can audit exactly which sources Perplexity leaned on for a given query, which is a big part of how the testing behind this article was possible.

The other structural piece worth knowing is Comet, Perplexity's browser, which changes retrieval behavior further for users who have it. Comet can browse live pages directly rather than relying solely on a cached index, which in our testing produced citations of pages so recent that a standard search-index crawl could not plausibly have picked them up yet. If your business publishes something today, a Comet session can theoretically surface it today — a much tighter loop than the days-to-weeks lag typical of index-based retrieval.

To move past guesswork, we ran 50 local-intent queries through Perplexity in Copilot mode across three verticals with different content maturity levels: home services (HVAC, plumbing, roofing), medical and dental (dentists, med spas), and legal (personal injury and family law attorneys), spread across the mid-size metros we work in most. Queries ranged from broad ("best HVAC company in Tampa") to specific ("emergency plumber that answers calls after 10pm in Austin") to comparison-style ("difference between two dental implant clinics in Naples").

For every query we logged the full citation list Perplexity displayed, not just the businesses named in the prose answer, and recorded seven attributes for each cited source: publish date, last-modified date if visible, presence and density of schema markup, whether the page named a single clear business entity or covered several, how many other domains were cited alongside it, how deep the page went on the specific topic asked (a full page versus a passing mention), and whether the source was cited elsewhere in our dataset (a rough proxy for third-party corroboration).

This is a directional study, not a controlled experiment against Perplexity's actual scoring code, which is not public. But running the same structured logging across 50 queries and three fairly different verticals is enough to separate consistent patterns from noise, and it is the same method we use to validate GEO recommendations before we put them in front of clients.

A methodology note worth flagging: verticals behaved differently in scale, not in kind. Legal queries pulled from a noticeably smaller pool of qualifying sources than home services, since fewer attorney sites publish deep, frequently updated content, which meant a single well-optimized page could swing a legal query's citation list more than an equivalent change could in the more crowded home-services set. The six signals below held across all three verticals, but if you operate in a thinner content category, expect individual fixes to move your citation rate faster than the averages here suggest.

Across the dataset, six attributes correlated most consistently with a source being cited, in roughly descending order of how reliably they predicted inclusion. None of these is a confirmed weight from Perplexity's own documentation — they are patterns, but they were consistent enough across all three verticals that we now treat them as our working model.

If you have already optimized for ChatGPT citations, most of that work transfers, but the emphasis shifts in two directions that matter enough to change how you prioritize.

Perplexity rewards freshness more aggressively. In side-by-side testing on overlapping queries, ChatGPT would still cite a well-established, authoritative source from two or three years ago if nothing newer covered the topic as well. Perplexity was noticeably more willing to drop an older, more authoritative page in favor of a thinner but more recently published one, especially on queries with any time-sensitive framing ("best," "top," "2026"). If your content strategy has been "publish once, let authority accumulate," Perplexity is the engine that punishes that approach first.

ChatGPT, by contrast, weights authority and third-party trust more heavily relative to recency. A well-established local business with strong directory presence and press mentions but a stale website tended to still surface in ChatGPT answers; the same business scored worse in our Perplexity log unless something on the page had a recent date attached. The practical read: ChatGPT forgives a dated page if the entity behind it is well-established; Perplexity does not extend the same grace period.

The other consistent difference is citation count and diversity. ChatGPT answers in our broader tracking tend to name fewer sources per answer and will sometimes concentrate on one. Perplexity spreads citations across more domains by default, which is good news for smaller or newer businesses: you do not need to be the single best answer, you need to be a credible, current, on-topic option among several.

Given those signals, five tactics moved the needle most in our follow-up testing, where we made specific changes to client and test pages and re-ran the same query set 30 to 45 days later.

Run this monthly with Copilot mode, using your own site's exact sources so you can tell whether you are cited or just mentioned in the answer text without a source link.

One more distinction worth understanding: Perplexity Pages lets any user (including you) publish a shareable, AI-generated research page on a topic, which itself can rank in Google and get indexed. This creates two separate ways to show up in the Perplexity ecosystem, and they are not the same play.

Being cited as a source means your own website page is one of the domains Perplexity pulled from when answering a user's query — this is what the rest of this article is about, and it is the higher-value outcome because it puts your business directly in front of someone actively deciding. Being the subject of, or publishing, a Perplexity Page means a piece of AI-generated content about your topic exists on Perplexity's own domain, which can itself accumulate visibility in Google search results over time, functioning more like a distribution channel than a citation.

For most local businesses, the source-citation path matters far more day to day, since it captures people mid-decision inside Perplexity itself. Perplexity Pages are worth a small amount of experimentation (publishing a well-sourced page on a topic where you are the best available source can itself become something later Perplexity queries cite), but should not distract from the core work: making your own site the clear, current, well-structured answer that the retrieval pipeline finds first.

One caution from the legal-vertical portion of our test set: a competitor's Perplexity Page summarizing "top personal injury attorneys" in a metro we tested was itself getting cited as a secondary source in follow-up queries, ahead of some of the individual firm websites it referenced. That is a reminder that a well-built Perplexity Page on your topic, even one you did not publish, can insert itself into the citation chain between a user and your own site. Checking whether one already exists for your core service and city terms takes five minutes and tells you whether you are competing against a page as well as against other businesses.