AI Search · GEO / AEO

AI Assistants Cite Directories Twice as Often as Brand Sites When Recommending Providers. Four Datasets Point the Same Way.

On-Page.ai Blog · September 21, 2026 · Data study analysis

Sector directories supplied 41% of the citations behind AI-generated provider recommendations last summer, against 18% for the providers’ own sites. When a buyer asks an assistant who to hire or buy from, the answer is more than twice as likely to be sourced from a directory as from the business being recommended.

That finding comes from a study of 3,850 commercial prompts published September 15 by Citations.press, run across ChatGPT, Google AI Mode, Perplexity, and Google AI Overviews between June 10 and August 22, 2026. There is one important caveat. The publisher operates the six directories its researchers tested. We will get to that, and to what the publisher’s own raw-data promise looks like when you check it.

The result has company. Three other datasets, produced by three organizations with no directory business, measured different samples over different windows with different methods, and point in the same direction. Across the four datasets and fourteen months of measurement, AI assistants answering recommendation queries consistently resolve to pages that hold many providers rather than to the providers’ own pages.

The headline numbers

The breakdown of the prompts helps explain the result. Of the 3,850 prompts, 2,140 (56%) named a category but no company (“who does industrial automation in the Midlands”), 890 (23%) compared options, and 820 (21%) named a specific company to verify. The directory advantage concentrates where the buyer has not yet chosen: the brand homepage appeared in just 6% of responses to category prompts, the moment when a provider most needs to be discovered.

Four datasets, one direction

The Citations.press study is small relative to what has come before it, and it says so itself: the figures “have not been replicated by an independent party.” Its own release lists three larger studies as context. We pulled all three, plus the new one, into a single comparison. They measure different things at different scales, which makes the agreement worth attention.

Why the engines resolve to aggregators

The study offers a mechanism worth quoting: “A model answering a narrow category question appears to favour a source that resolves that question completely over a source that is generally trusted.”

Complete resolution is a retrieval property. When the prompt asks “who are the options,” a directory page answers the whole question in one document: ten providers, their specialties, locations, contact routes. A brand site answers a different question, “tell me about this one company,” one-tenth of what the buyer asked. The assistant that cites the directory still has a complete answer; the one that cites a single vendor does not.

DeltaV’s dataset shows the same mechanism from the other side. Comparison pages posted the highest citation rate of any page type they measured, 1.87 citations per retrieval against a portfolio average of 1.29, a 45% premium. Yet comparison pages earned only 4.1% of total citations, because so few exist. Pages shaped to resolve multi-provider questions get cited at above-average rates; the market just under-produces them. That is a gap a content team can fill.

Winning page types differ by industry

DeltaV split its 25,337 citations by industry and found that the winning page type changes with the sector. The engines and the 90-day window stayed the same, but the answer to “what should we publish” changed with the industry.

Read across the four datasets, the working rule is: the more a purchase looks like procurement (compare providers, shortlist, verify), the more citations flow to aggregators. Where the purchase looks like choosing one obvious institution or a nearby business, citations stay with brand-owned pages.

At 30-million-source scale, aggregators sit in the top ten

Peec AI’s March 2026 analysis of 30 million cited sources across ChatGPT, Google AI Mode, Gemini, Perplexity, and AI Overviews ranked the most-cited domains in US AI search. Two review/listing platforms sit inside the top ten, ahead of every individual brand or publisher except the giants.

The conflict of interest, checked firsthand

The study’s publisher has a direct financial interest in its headline.

The release also says the full prompt set, raw citation records, classification rules, and the 200-response manual audit are “published in full alongside the study,” and that they are “available to any editor or researcher on request.” Those two statements are not the same thing, so we checked what is public. Citations.press runs a searchable JSON API for its citation index. On September 21, 2026:

That leaves the finding unaudited, and it is why the convergence with DeltaV, Semrush, and Peec matters more than the exact percentages.

Domain authority barely correlates with AI citations

The release also contains a result that matters more than the headline if it survives replication. Inside the dataset, the correlation between a directory’s domain authority and how often it was cited was a Spearman rank correlation of 0.19. Coverage depth within a stated category predicted citations better than authority did.

For classic Google rankings, authority metrics correlate with outcomes strongly enough that the industry built an entire tooling category around them. This dataset says the engines behind AI recommendations are asking a different question, one about how completely a page answers “who provides X.” That is consistent with Semrush’s observation that citation weight pools in a few aggregating platforms regardless of any single page’s links, and with Peec’s finding that G2 outranks most publishers despite no editorial operation.

Moves that match each citation pattern

The actionable read depends on where your category’s citations resolve today. Run the prompts a buyer would run, record which domains come back, and branch from there.

The limitations that come with these numbers

The study measured citation frequency, and being cited more often is not proven to win more deals, as the release itself says. The researchers also flag response instability. The same prompt can draw different citations across repeated runs, accounts, and regions, so any snapshot, including theirs, is one draw of a moving system. The prompt set was also weighted toward the six sectors tested, and classification of source types involved judgment, checked to 94% agreement on a 200-response sample but not independently audited.

The direction survives those limitations. Four organizations with different incentives, from a 3,850-prompt study to a 100-million-citation dataset, independently found that AI assistants answer “who should I use” by citing pages that aggregate providers rather than pages that own them. Directories stopped driving web traffic twenty years ago.

Sources

• Citations.press study release, September 15, 2026: markets.businessinsider.com (Marketers Media wire; also at news.marketersmedia.com). Publisher site: citations.press. API verification performed by On-Page.ai, September 21, 2026.
• DeltaV Digital, “AI search citations study: what 25,000+ citations reveal,” July 13, 2026: deltavdigital.com.
• Semrush, “The Most-Cited Domains in AI: A 3-Month Study,” November 10, 2025: semrush.com.
• Peec AI, “Top domains cited by AI search: Analysis based on 30M sources,” March 31, 2026: peec.ai.