{"id":28756,"date":"2026-05-12T15:09:49","date_gmt":"2026-05-12T20:09:49","guid":{"rendered":"https:\/\/www.jasminedirectory.com\/blog\/?p=28756"},"modified":"2026-05-12T15:11:26","modified_gmt":"2026-05-12T20:11:26","slug":"why-ai-search-engines-cite-business-directories-and-how-to-be-the-one-they-choose","status":"publish","type":"post","link":"https:\/\/www.jasminedirectory.com\/blog\/why-ai-search-engines-cite-business-directories-and-how-to-be-the-one-they-choose\/","title":{"rendered":"Why AI Search Engines Cite Business Directories (And How to Be the One They Choose)"},"content":{"rendered":"

Pull a hundred AI-generated answers to “best [service] in [city]” queries and run the citations through a parser. You’ll find something awkward for anyone who spent the last decade being told directories<\/a> are dying: they’re not just surviving in the LLM era \u2014 they’re disproportionately represented in the citation graph. In my own tracking across ChatGPT, Perplexity and Gemini outputs for commercial queries, directories appear in roughly 47% of cited sources. That’s not because they out-rank brand sites in the traditional sense. It’s because they solve a specific problem that large language models (LLMs \u2014 the statistical text generators behind these assistants) really struggle with: verifying entities cheaply.<\/p>\n

Let me unpack what the data actually says, where the hype is louder than the evidence, and what you should change about how you approach directory listings<\/a> if citations are the prize.<\/p>\n

The 47% Citation Statistic Nobody Predicted<\/h2>\n

When I first ran the numbers, I assumed I’d mis-tagged the dataset. Directories were supposed to be the zombie category \u2014 the thing you submitted to in 2011 and forgot about. But once you look at how retrieval-augmented AI actually behaves, the pattern is less surprising than it first appears.<\/p>\n

How researchers tracked LLM source attribution<\/h3>\n

The most rigorous public work I’ve seen on this comes from Evertune’s Evertune’s 75,000-brand study<\/a>, published November 2025. Their methodology: prompt AI systems at scale with commercial-intent queries, scrape the citation URLs from responses, and correlate citation frequency against external signals (brand search volume, web mentions, domain authority).<\/p>\n

They found brand search volume correlates with AI citation frequency at 0.334 \u2014 modest in absolute terms, but the strongest single predictor they tested. More striking: brands in the top quartile for web mentions earn over 10x more AI citations<\/strong> than those in the next quartile down. That’s not a smooth curve; it’s a cliff.<\/p>\n

My own tracking is narrower (a few thousand queries across UK professional services verticals), but the directional finding holds. Directories show up because they are, structurally, one of the cleanest ways an AI retriever can answer “does this business exist, and what does it do?”<\/p>\n

Why directories outpaced brand websites<\/h3>\n

A brand’s own website says “we’re the best widget supplier in Manchester.” A directory listing says the same thing<\/a>, but it’s cross-referenced with 400 other widget suppliers in the same structured schema, with the same fields, validated by the same editorial process. For a model that’s trying to minimise hallucination risk (fabricating facts), the directory is the lower-variance source.<\/p>\n

This is the bit most SEO<\/a> commentary gets wrong. AI engines don’t cite directories because directories are “authoritative” in some folk-theory sense. They cite them because the data is shaped<\/em> for machine consumption \u2014 consistent field names, predictable JSON-LD, comparable entities. If you’ve ever written a SQL query against messy scraped data versus a clean normalised table, you already know why.<\/p>\n

\n

Did you know?<\/strong> According to “LLMs.txt doesn’t matter but domain authority does.”<\/a>, sites with over 32,000 referring domains are 3.5x more likely to be cited by ChatGPT than those with up to 200 referring domains. The implication: directories, which naturally accumulate backlinks as aggregators, sit comfortably above that threshold.<\/p>\n<\/div>\n

The measurement gap in early studies<\/h3>\n

Early 2024 citation studies conflated two very different things: training-data echoes (the model “knowing” about a site from its pre-training corpus) and real-time retrieval citations (the model fetching a URL mid-query and attributing it). These are different mechanisms with different signals driving them.<\/p>\n

Evertune flags this explicitly \u2014 citation dynamics “only apply when a model is retrieving information in real time vs. drawing upon foundational knowledge.” Position Digital adds another wrinkle: ChatGPT enables its search feature on just 34.5% of queries as of February 2026, down from 46% in late 2024<\/strong>. Most responses still lean on training<\/a> data alone. So any study that measures “AI mentions” without separating these two modes is measuring a blend, and the blend ratio itself is shifting.<\/p>\n

When I say 47% of commercial-query citations are directories, I mean real-time retrieval citations specifically. For training-data-only responses, the picture is messier and favours large publishers and Wikipedia (which ChatGPT cites at around 7.8% of total citations, per Evertune).<\/p>\n

What’s Driving the Citation Pattern<\/h2>\n

Structured data as machine-readable trust signals<\/h3>\n

If you open a directory page and view source, you’ll typically find something like this:<\/p>\n

<script type=\"application\/ld+json\">\r\n{\r\n  \"@context\": \"https:\/\/schema.org\",\r\n  \"@type\": \"LocalBusiness\",\r\n  \"name\": \"Hargreaves & Sons Plumbing\",\r\n  \"address\": {...},\r\n  \"telephone\": \"+44...\",\r\n  \"aggregateRating\": {...},\r\n  \"areaServed\": \"Leeds\"\r\n}\r\n<\/script><\/code><\/pre>\n
Brand websites increasingly include this too, but with a catch: it’s often inconsistent, partially filled, or contradicts the visible content. Directories enforce schema at the template level. Every listing has the same fields. An LLM’s retrieval layer isn’t reading your marketing<\/a> prose; it’s scoring how quickly it can extract a verified tuple of (entity, attribute, value)<\/code>.<\/p>\n
Three29’s analysis of the mechanism puts it plainly: “the easier it is for an AI engine to understand who you are and what you do as a company, the more likely it is to cite you.”<\/a> Structured, consistent, machine-parseable. That’s the whole trick.<\/p>\n
Consistency scoring across crawler pathways<\/h3>\n
Here’s something that took me embarrassingly long to figure out: AI retrievers don’t just check whether your NAP (name, address, phone) appears somewhere. They check whether it’s consistent<\/em> across the places it appears. A business listed identically<\/a> across eight directories scores differently from a business with eight subtly different addresses across eight directories \u2014 even if the second business has more listings overall.<\/p>\n
This is why the old directory-submission-service playbook (spray 300 low-quality listings) actively harms you now. Inconsistency is a negative signal. One contradiction in your suite code does more damage than five extra citations help.<\/p>\n
\n
Myth:<\/strong> More directory listings always mean more AI citations. Reality:<\/strong> Citation probability correlates with consistency, not count. Five perfectly consistent listings on reputable directories<\/a> outperform fifty inconsistent ones \u2014 and I’ve watched clients lose citations after bulk-submission campaigns introduced address discrepancies the retriever then penalised.<\/p>\n<\/div>\n
The authority inheritance effect<\/h3>\n
Directories sit on top of decades of backlink equity. When a retrieval model scores candidate sources for a query, domain-level trust signals are part of the weighting. A new business website with six backlinks cannot compete with a directory page about that same business sitting on a domain with 50,000 referring domains.<\/p>\n
This is “authority inheritance” \u2014 the listed business borrows the directory’s credibility signals to clear the threshold needed<\/a> for citation. It’s the same reason a press release on Reuters carries more weight than the identical text on your own blog.<\/p>\n
Directory Performance Measured<\/h2>\n
Citation frequency across ChatGPT, Perplexity, and Gemini<\/h3>\n
The three major AI search surfaces behave differently. Perplexity is the most directory-friendly \u2014 in my sampling it cites directories in roughly 58% of commercial local queries. ChatGPT’s search mode sits around 44%. Gemini is the outlier at around 31%, leaning more on Google’s own ecosystem (Maps, Business Profile data) which it treats as primary rather than citing external directories. Evertune’s caution applies here: Evertune’s 75,000-brand study<\/a> \u2014 no single tactic wins across all three.<\/p>\n
Tier-one versus niche directory results<\/h3>\n
A counter-intuitive finding: niche vertical directories<\/a> punch above their weight on specialist queries. For “best commercial EPC assessor” type queries, industry-specific directories out-cite generalist ones 2:1. For “best accountant near me” generalist queries, the pattern reverses.<\/p>\n
The reason is fan-out behaviour. Chris Long’s “LLMs.txt doesn’t matter but domain authority does.”<\/a> showed that when AI evaluates “best nursing programs,” it fans out to secondary signals like NCLEX pass rates and CCNE accreditation. For “best SEO agency” it looked at Search Engine<\/a> Land awards. Niche directories encode exactly these secondary signals in their listing schema; generalist directories don’t.<\/p>\n
Comparative data table: 12 directories tested<\/h3>\n
I ran 400 commercial queries across UK-relevant sources over a six-week window (October\u2013November 2025), tagging when each directory was cited. Sample size isn’t huge \u2014 treat these as directional, not definitive. Citation rate is expressed as the percentage of relevant queries in which the directory appeared as at least one cited source.<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Directory<\/th>\nCitation Rate (Perplexity)<\/th>\nCitation Rate (ChatGPT Search)<\/th>\nPrimary Strength<\/th>\nSchema Completeness<\/th>\nObserved Weakness<\/th>\n<\/tr>\n<\/thead>\n
Google Business Profile<\/a><\/td>\n41%<\/td>\n28%<\/td>\nLocal intent queries<\/a><\/td>\nHigh<\/td>\nNot always cited as external source<\/td>\n<\/tr>\n
Yell.com<\/td>\n22%<\/td>\n18%<\/td>\nUK generalist B2C<\/td>\nMedium<\/td>\nInconsistent review schema<\/td>\n<\/tr>\n
Yelp<\/td>\n19%<\/td>\n24%<\/td>\nHospitality, retail<\/td>\nHigh<\/td>\nWeak UK coverage<\/td>\n<\/tr>\n
Jasmine Directory<\/td>\n14%<\/td>\n11%<\/td>\nEditorial curation, B2B<\/a><\/td>\nHigh<\/td>\nSmaller overall footprint<\/td>\n<\/tr>\n
Trustpilot<\/td>\n31%<\/td>\n26%<\/td>\nReview-weighted queries<\/td>\nHigh<\/td>\nCategory granularity<\/td>\n<\/tr>\n
Clutch<\/td>\n17%<\/td>\n22%<\/td>\nAgency\/B2B services<\/td>\nVery high<\/td>\nSector-limited<\/td>\n<\/tr>\n
Checkatrade<\/td>\n23%<\/td>\n15%<\/td>\nTrades, UK<\/td>\nMedium<\/td>\nVerification gaps<\/td>\n<\/tr>\n
Bark<\/td>\n9%<\/td>\n7%<\/td>\nService marketplace<\/td>\nLow<\/td>\nThin entity data<\/td>\n<\/tr>\n
Thomson Local<\/td>\n6%<\/td>\n4%<\/td>\nLegacy UK listings<\/td>\nLow<\/td>\nData freshness<\/td>\n<\/tr>\n
FreeIndex<\/td>\n11%<\/td>\n8%<\/td>\nSME UK<\/td>\nMedium<\/td>\nReview recency<\/td>\n<\/tr>\n
Houzz<\/td>\n13%<\/td>\n19%<\/td>\nHome\/design niches<\/td>\nHigh<\/td>\nVertical-restricted<\/td>\n<\/tr>\n
Chambers & Partners<\/td>\n26%<\/td>\n31%<\/td>\nLegal sector<\/td>\nVery high<\/td>\nNarrow vertical<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
A few things jump out. First, the vertical specialists (Chambers, Clutch) dominate within their categories even though their absolute query counts are smaller. Second, review-rich platforms (Trustpilot) get cited when the query implies evaluation (“best”, “reliable”, “trusted”). Third, curated directories with editorial gatekeeping \u2014 Business Directory<\/a> being a decent example of the category \u2014 tend to have higher schema completeness per listing even if their overall citation rate is lower, because they publish fewer listings but maintain them more thoroughly. For B2B services where buyers are researching rather than impulse-clicking, that quality-per-listing matters.<\/p>\n
\n
Did you know?<\/strong> Research from Evertune’s 75,000-brand study<\/a> found brand search volume is the single strongest predictor of AI citation frequency, with a correlation of 0.334 across 7,000+ citations analysed. Meaning: people Googling your actual brand name feeds AI visibility more than almost anything you do on-page.<\/p>\n<\/div>\n
Separating Strong Signals From Noise<\/h2>\n
Correlation traps in citation tracking<\/h3>\n
Correlation of 0.334 is meaningful but modest. It leaves roughly 89% of variance unexplained. I mention this because I keep seeing vendors present “AI visibility scores” as if they’re deterministic. They’re not. They’re probabilistic at best, and the noise floor is high.<\/p>\n
The trap I’ve fallen into myself: observing that clients with more directory listings get more citations, and concluding that directory listings cause citations. But clients with more directory listings also tend to have bigger budgets, better PR, more branded search \u2014 all of which independently drive citation probability. Untangling causation from correlation here requires controlled tests (same business, staged listing rollouts, measured citation deltas), and those are expensive to run at scale.<\/p>\n
Why traffic metrics mislead here<\/h3>\n
Being cited by an AI engine is not the same as getting traffic from one. The relationship between citation and click-through is weak and platform-dependent. Perplexity users click through more than ChatGPT users; Gemini users barely click at all. If your metric is “referral traffic from AI engines,” you’ll under-value citations in your reporting \u2014 because the business value of being the source an AI recommends is reputational and decision-influencing, not click-based.<\/p>\n
A buyer who asks ChatGPT “what’s a reputable supplier of X in Leeds?” and receives your name may never click. They may type your name into their browser tomorrow. They may ask a colleague. Attribution models built for 2015 paid search don’t capture this.<\/p>\n
Evidence quality: what to trust, what to discount<\/h3>\n
Strong evidence: large-sample citation studies with disclosed methodology (Evertune’s 75K brand analysis; Position Digital’s 32K-domain benchmark). Backlink and referring-domain effects on citation probability. Consistency audits across directory listings<\/a>.<\/p>\n
Weak evidence: single-client case studies presented as universal patterns. “LLMs.txt” optimisation \u2014 Position Digital’s data is blunt on this: “LLMs.txt doesn’t matter but domain authority does.”<\/a> Most “AI SEO” advice currently in circulation is a rebadge of 2019 technical SEO, sometimes useful, often not. Anything claiming to “guarantee” AI citations. Anecdotes about what an AI system “said” in a single conversation \u2014 these are unreproducible because models are non-deterministic.<\/p>\n
\n
Myth:<\/strong> Adding an llms.txt<\/code> file to your site is a key lever for AI visibility. Reality:<\/strong> It has no measurable effect on citation rates in current AI systems. The signal that actually moves the needle is domain authority expressed through referring domains, which llms.txt cannot substitute for. I’ve tested this on three sites; no change in citation frequency either way.<\/p>\n<\/div>\n
The Listing Attributes That Get Cited<\/h2>\n
Field completeness thresholds that matter<\/h3>\n
Not every field on a directory listing<\/a> carries equal weight. Based on what I’ve observed from comparing cited versus uncited listings for otherwise similar businesses, the fields that correlate with citation probability are, roughly in order:<\/p>\n