The directories AI cites: 201K citations ranked by brand attributions
201K directory-classified citations and a host ranking by brand attributions. Inspect the vertical split, engine shares, and method.
Originally published October 7, 2026
Directory-classified sources and selected review hosts supplied 2.7% of measured citations. Use the attribution ranking to inspect relevant pages, not to buy a bulk listing count.
July 9–October 6, 2026 (UTC) · 200,999 directory-classified citations out of 7,429,354 citations
- Classified citation share
- 2.7%Broad source taxonomy plus host overlay
- Citation occurrences
- 200,999Not listing success rate
- G2 attributions
- 9,773A separate source-brand unit
Directory selection needs a market-specific shortlist, not a bulk submission target. Our classified directory and review-host group supplied 200,999 of 7,429,354 measured citations, or 2.7%. The attribution ranking below shows which observed hosts were associated with brands named in answers, rather than merely listing websites that accept submissions.
The measurements come from Parse, the AI answer index Dimitry Apollonsky also runs; the analysis and interpretation are Signals'. This is an observational study, not a test of a Signals service.
The window is July 9 through October 6, 2026 on ChatGPT Search and Google AI Mode. Signals runs an aged Reddit account marketplace plus an editorial network for AI brand mentions across Reddit, Quora, Product Hunt, and Threads.
G2 led the observed attribution ranking
The main g2.com host had 9,773 brand attributions in the measured window. An attribution links a cited page to a brand named in the answer. It is a different unit from a citation occurrence: a source associated with several named brands can contribute several attributions within the same answer observation.
G2 led the observed directory-host attribution ranking
Source relationships to brands named in answers · attributions · July 9–October 6, 2026 (UTC)
View data table
| g2.com | 9,773 |
|---|---|
| learn.g2.com | 6,429 |
| capterra.com | 2,991 |
| apps.shopify.com | 2,782 |
| attorneys.superlawyers.com | 2,286 |
| indeed.com | 1,934 |
| expertise.com | 1,921 |
| openvc.app | 1,705 |
| napfa.org | 1,468 |
| trustpilot.com | 1,311 |
Leading ten hosts by attributions in eligible answers. Source: Parse index, analyzed by Signals.
Attributions can count several brands per source. They are not citation occurrences or paid-listing results.
The table ranks selected leading hosts and retains separate hostnames, including learn.g2.com. That distinction matters because a host's editorial articles and product profiles can serve different purposes. A buyer considering a profile should not assume the host's entire attribution total came from pages like the proposed listing.
The ranking is an observed source map, not a list of paid inventory or a recommendation to buy from the largest host. It lacks the number of eligible listings, market coverage, and prices needed for an outcome-per-listing comparison. Inspect the specific page and buyer question before using the ranking in a brief.
Vertical hosts also appear in the source map
Selected vertical hosts include professional-services and specialty directories, while the horizontal comparison group contains broad review platforms. The figure reports host-level attributions using the same observation window. Its purpose is to show that the candidate set extends beyond a familiar software-review shortlist, not to establish that every vertical directory beats every horizontal platform.
Vertical hosts appear alongside broad review platforms
Selected host brand attributions · attributions · July 9–October 6, 2026 (UTC)
| Horizontal: g2.com | 9,773 |
|---|---|
| Horizontal: capterra.com | 2,991 |
| Horizontal: trustpilot.com | 1,311 |
| Vertical: attorneys.superlawyers.com | 2,286 |
| Vertical: expertise.com | 1,921 |
| Vertical: psychologytoday.com | 488 |
Selected horizontal and vertical hosts; not normalized by listing inventory. Source: Parse index, analyzed by Signals.
The selection illustrates host types, not an exhaustive vertical-versus-horizontal market comparison.
The groups are chosen by host purpose rather than inferred from a business's purchase history. They are not normalized for the number of categories, profiles, or brands each host contains. A larger count can reflect broader coverage or a better match to the tracked questions rather than a more effective listing product.
For an operator, the next action is to check eligibility and category fit. A specialist directory can be relevant to one profession and irrelevant to another. A global table helps find candidates, but the proposed business profile still needs a reason to exist on the host and accurate facts that buyers can use.
Directory shares differ between engines
The engine split divides classified directory citation occurrences by all measured citation occurrences for each engine. That gives ChatGPT Search and Google AI Mode their own denominators and avoids confusing different citation volumes with preference. These are calendar-window source sets, not a matched experiment showing both engines the same directory pages.
Across the combined sample, the directory-classified group contributes 200,999 occurrences. The source taxonomy is broad and includes marketplaces and other directory-like hosts, with an explicit overlay for selected review and vertical sites. That definition is why the headline should not be quoted as a census of traditional business-directory listings alone.
Use the split to preserve engine-specific results when monitoring a listing. A profile appearing on one engine does not establish the same exposure on the other. Save the exact question, answer, date, and URL so subsequent checks can distinguish a genuine repeated observation from a changed source panel or a different query.
Directory classification is broader than a verified listing census
The cohort combines the index's directory labels with explicitly selected review and vertical-directory domains. The underlying labels are coarse. Some classified hosts publish several kinds of content, and a domain can function as a marketplace, editorial resource, or directory depending on the page. We preserve that uncertainty instead of treating the taxonomy as a manual audit.
The study excludes parse.gl and soar.sh, including their subdomains. This is particularly important because a self-referential index can otherwise enter a directory group and inflate the result. The host patterns and exclusion rule appear in the SQL receipts so a reader can inspect the actual definition.
The ranked figures focus on identifiable hosts, but the broad citation share still uses the stated classification. A narrower future study could manually verify listing pages and eligibility. It should publish a new denominator rather than silently relabeling this broad group as a fully verified population of business profiles.
A citation can point to a category or an article
A directory's host total does not identify which page type delivered the evidence. The cited URL may be a company profile, category comparison, review page, marketplace listing, or editorial article. Those surfaces provide different information and may require different work from an operator seeking accurate representation of a business or product.
This is especially visible when related hosts appear separately, such as g2.com and learn.g2.com. The observation invites page inspection; it does not prove that purchasing a profile would reproduce the editorial host's source use. Treat the final destination as part of the evidence, not a detail to omit from a campaign report.
A useful vendor brief should identify the deliverable at that same level. Submission, publication, verification, and editorial inclusion are different milestones. The directory-listings guide explains how to specify them and how to reject a proposal that replaces a relevant public page with a large count of completed forms.
No bulk-listing service effect was measured
The Signals ledger available for this batch does not contain a complete delivered-URL cohort for directory listings. We therefore cannot connect a listing package to a subsequent change in citing domains or brand naming. The directory share and attribution ranking are Parse observations, not outcomes measured from purchased Signals listing campaigns.
This is not a zero-effect result. It is an unmeasured effect, and the distinction matters. A zero would require an adequately observed treatment cohort and a defined outcome that remained unchanged under the chosen comparison. Here, the delivered-listing and publication evidence needed to construct that cohort is missing.
The study also lacks a paid-versus-free listing comparison. Host visibility cannot resolve that question because the cited material may not be a paid profile at all. Use the source map to improve selection and verification, while keeping claims about the return from a package outside the conclusions this analysis can support.
The study cannot rank Google Maps against directory work
This analysis is built around cited web sources, not a complete model of local discovery. Maps, business profiles, and other interfaces are not consistently represented as comparable citations in the panel. Their absence from the directory table therefore cannot establish that they have less value for a local business or for AI-assisted discovery.
A local operator should define the specific location and service questions being checked, then inspect the sources backing the named businesses. Specialty directories may contain credentials or service details, while other sources provide address, availability, or customer experience. Those information needs are not interchangeable and should not be reduced to a single global host rank.
Parse's directory research offers related category-level context. Its result has its own denominator and should be read alongside this attribution cut rather than merged into it. The practical decision remains which relevant public records need correction and which evidence is missing for the particular buyer question.
The checks preserve distinct counting units
The citation share is computed from the classified citation set and reconciled through the engine totals. The attribution ranking separately joins brand-source relationships to eligible answers in the same window. These outputs cannot be added together because one counts source occurrences and the other counts relationships between sources and named brands.
The receipts expose the raw host rows and the number trace used for the figures. The leading-host and selected-vertical figures retain host labels and exact attribution counts. The percentage figure uses only citation-occurrence denominators, with rounding applied for display. No mixed-unit chart asks the reader to compare unlike quantities on one scale.
These checks validate the arithmetic and join scope, not the accuracy of every taxonomy label. A quarterly refresh should rerun both the source classification and the answer eligibility rules, then disclose any definition changes. A corrected taxonomy can move totals even when no underlying engine behavior changes, which is another reason to preserve the method.
The operator should inspect the candidate page before buying
The useful output is a shortlist of pages and hosts to examine, followed by a factual listing brief. Start with the relevant category, confirm eligibility, and compare the public business information with the official record. A directory count becomes actionable only when it points to a concrete correction or a relevant source the business can legitimately maintain.
The directory guide covers the vendor specification and monitoring record. Other Signals research articles describe editorial, Reddit, and listicle sources. Their findings can inform a broader plan, but they do not form a common service-return ranking because the measured populations and units differ.
For a new listing, save publication and correction receipts before monitoring AI answers. Keep the profile's existence, citation presence, and brand naming separate. That makes a later report more informative even when the result is no observed citation in the selected panel. The absence should stay visible rather than being replaced by a delivery count.
Questions about the directory ranking
The ranking describes observed hosts, so it is best used to decide what to inspect next. It does not guarantee access to an editorial page or show the effect of purchasing a profile. The questions below address the common ways a host-level source map can be mistaken for a placement recommendation or a measured service outcome.
Are these the directories every brand should join?
No. Eligibility, category, geography, and the cited page's purpose still determine whether a host is relevant to the business.
Does an attribution mean the directory caused the recommendation?
No. It records a source-brand relationship in an answer, not a causal effect of a listing or payment.
Does this measure a bulk-listing package?
No delivered-listing cohort was available for this analysis. The service effect is unmeasured, not a measured zero.
How we measured
Read-only production queries frozen at October 7, 2026. Answers are ChatGPT Search and Google AI Mode in the stated UTC window, excluding gpt-5-mini and gpt-5-3-mini records while retaining blank model labels. Citation IDs are expanded from eligible answers. Parse and Soar source domains and subdomains are excluded. Latest content snapshots are selected strictly before October 7. SQL, raw output, and number trace are committed under docs/research-receipts for this slug. Directory source-type labels are combined with explicit review and vertical-host patterns. Attribution rows must match an eligible answer and source in the window. Selected vertical hosts are contrasted with selected broad review platforms.
What this cannot tell us
- Directory labels are coarse and include marketplaces or multi-purpose publishers.
- The cited page may be editorial rather than a listing.
- No delivered-listing cohort, paid/free comparison, or local-platform census is available.
- The tracked question panel is not a random sample of all AI search.
- This observational analysis does not establish a causal effect of a Signals service.
