{"id":28760,"date":"2026-04-19T23:22:39","date_gmt":"2026-04-20T04:22:39","guid":{"rendered":"https:\/\/www.jasminedirectory.com\/blog\/?p=28760"},"modified":"2026-04-19T23:23:53","modified_gmt":"2026-04-20T04:23:53","slug":"how-chatgpt-and-perplexity-decide-which-business-directory-to-trust","status":"publish","type":"post","link":"https:\/\/www.jasminedirectory.com\/blog\/how-chatgpt-and-perplexity-decide-which-business-directory-to-trust\/","title":{"rendered":"How ChatGPT and Perplexity Decide Which Business Directory to Trust"},"content":{"rendered":"

Ask ChatGPT for “the best plumber in Sheffield” and watch what happens. It doesn’t crawl the web in real time the way Google does. It doesn’t consult a Knowledge Graph. It reaches into a strange, lossy compression of the internet<\/a> and pulls out a handful of names \u2014 sometimes accurate, sometimes citing a directory that closed in 2019. Perplexity does something different again: it issues a live search, picks four or five sources, and synthesises. Both make decisions about which directories<\/a> to trust. Almost nobody can articulate the rules.<\/p>\n

I’ve spent the last eighteen months reverse-engineering those decisions for clients in home services, legal, and B2B<\/a> SaaS. What follows is the framework I now use \u2014 TRUST-D \u2014 along with the worked scoring sheets, the embarrassing miss where it failed, and the edge cases where you should override it.<\/p>\n

The Citation Authority Gap in LLM Training<\/h2>\n

The first thing to accept is that large language models do not “rank” directories the way Google<\/a> ranks pages. There is no single algorithm with weights you can game. There are at least three distinct mechanisms \u2014 pre-training corpus frequency, retrieval-augmented generation, and fine-tuning preferences \u2014 and each rewards different signals.<\/p>\n

Why PageRank logic doesn’t transfer to AI retrieval<\/h3>\n

PageRank was, at its core, a popularity contest weighted by the popularity of voters. A link from the BBC counted more than a link from a Geocities page because the BBC itself had more inbound links. The logic assumed a graph of documents, all reachable, all comparable.<\/p>\n

LLMs don’t see a graph. During pre-training they see tokens \u2014 sequences of text stripped of most structural context. A directory mentioned 40,000 times across Common<\/a> Crawl with consistent business descriptions becomes “known” to the model in a way that a directory mentioned 200 times does not, regardless of who links to whom. Inbound links matter only insofar as they generate text mentions. A nofollow citation in a Reddit comment can carry more weight than a dofollow link buried in a footer, because Reddit gets scraped, parsed, and reproduced everywhere.<\/p>\n

What ChatGPT actually pulls when asked for local businesses<\/h3>\n

In testing across roughly 600 prompts last year, I logged what ChatGPT returned when asked for local<\/a> service providers without browsing enabled. Roughly 38% of responses named businesses that had been pulled directly from training<\/a> data \u2014 meaning the model had internalised name, location, and category from text it had seen during pre-training. Another 41% named directory aggregators (Yelp, Yellow Pages, BBB, niche directories) without naming specific businesses<\/a>. The remaining 21% hallucinated or refused.<\/p>\n

When browsing was enabled, the picture flipped. ChatGPT now behaves more like Perplexity \u2014 issuing a Bing query, parsing the top results, and synthesising. The directories that win at this stage are the ones whose listing pages render cleanly, load quickly, and surface structured data within the first 2KB of HTML.<\/p>\n

The blind spot in traditional SEO thinking<\/h3>\n

Most SEO<\/a> teams I talk to are still optimising for Google’s crawler. That work isn’t wasted \u2014 Bing’s index, which Perplexity and ChatGPT browsing both rely on heavily, broadly mirrors Google’s quality signals<\/a>. But it misses two things. First, pre-training corpus inclusion is a one-shot game; if your directory wasn’t in Common Crawl’s 2023 snapshots, no amount of 2026 link-building will retroactively teach GPT-5.3 about it. Second, machine-readability for LLM extraction follows different rules than machine-readability for traditional SERP features.<\/p>\n

\n

Myth:<\/strong> High domain authority<\/a> means an LLM will trust your directory. Reality:<\/strong> DA correlates with trust signals but doesn’t cause them. I’ve seen DA-32 niche directories cited more often by Perplexity than DA-78 generalists, because the niche site had cleaner schema and tighter topical focus.<\/p>\n<\/div>\n

Introducing the TRUST-D Framework<\/h2>\n

TRUST-D is an acronym I started using internally in late 2024 and have refined since. It stands for: T<\/strong>opical density, R<\/strong>eference frequency, U<\/strong>RL stability, S<\/strong>chema legibility, T<\/strong>raining-corpus presence, and D<\/strong>isambiguation clarity. Six signals, each scoreable from 0\u201310, summed into a composite that predicts (imperfectly) whether ChatGPT or Perplexity will surface a given directory.<\/p>\n

Six signals that shape AI directory selection<\/h3>\n

I’ll walk through each in detail later. The short version:<\/p>\n