Sector Directories Are Driving AI Overview Citations

• Sector directories were cited in 41% of responses to supplier-selection prompts, compared with 18% for brand-owned pages.
• In engineering, directory citations outnumbered brand-site citations by 4.2 to 1.
• Brand homepages appeared in 14% of responses overall, and in 6% of responses to prompts that named a category rather than a company.
• Directories with one entity per entry and consistent attributes across entries were cited 3.1 times more often than directories with free-text listings.
• 68% of directory citations came from pages that were themselves lists, not individual profile pages.

The large platforms that dominate existing citation research do not cover vertical B2B categories. When a buyer asks "who are the best managed IT providers in the Netherlands" or "which engineering contractors work in offshore wind", Reddit and Wikipedia are not the answer. We wanted to know what is.

We ran 3,850 prompts across ChatGPT, Google AI Mode, Perplexity and Google AI Overviews between 10 June and 22 August 2026, grouped into three types. Category prompts asked for providers in a sector without naming a company. Comparison prompts asked the assistant to compare providers against each other. Verification prompts named a specific company and asked the assistant to assess it.

Each prompt was run three times on each surface and we report the modal citation set. For every response we recorded each cited URL, classified the source domain by type, and recorded whether the cited page was a list, a profile, an article or a brand-owned page. Classification was manual for the first 400 responses and rule-based thereafter, with a 200-response sample re-checked by hand. Agreement on the re-check was 94%. The prompt mix, classification approach and limitations are set out below.

When a prompt names a company, assistants cite that company's own site readily. When a prompt names a category, they mostly do not. This is the split that matters commercially, because category prompts are the ones that create a shortlist. A buyer asking "who should we use for X" is generating the consideration set. A buyer asking "is company Y any good" has already been given one.

Brand-owned pages accounted for 18% of citations on category prompts and 61% on verification prompts. On comparison prompts the figure sat in between, at 29%.

Not every directory performed. The ones that were cited shared a set of structural properties. Each entry described exactly one organisation. Every entry carried the same attributes in the same order, so the page could be read as a table even when it was not marked up as one. Category and location were stated in text rather than implied by navigation or filters. The listing page itself contained enough per-entry detail to answer a shortlist question without the assistant needing to follow through to a profile page.

Directories that failed on these points were crawled but rarely cited. The most common failure was putting all the substance behind a filter or a profile click, leaving the indexable listing page as little more than a set of company names.

We found a weak relationship between a directory's domain authority and its citation rate (Spearman ρ = 0.19). The stronger predictor was how completely the directory covered its stated category. A directory listing around 80 providers in one narrow sector was cited more often than a larger general business directory with higher authority and broader but shallower coverage.

A model answering a narrow category question is looking for a source that resolves that question completely, not a source that is generally trusted.

The four surfaces did not treat directories equally. Perplexity leaned on them most; ChatGPT least.

Our findings sit alongside three larger studies published over the past year. All three measure the same shape at platform scale. Our contribution is the sector layer underneath it.

The study tracked citations for six vertical directories across their respective categories.

• Audit the citation set before writing anything. Run your own category prompts across the four main surfaces and record which domains are cited. That list is your actual competitive set for AI visibility, and it will usually contain few or none of the pages your SEO programme is working on.
• Get listed accurately where your sector's directories are already cited. An incomplete or out-of-date entry is worse than none, because the assistant will reproduce whatever it finds.
• Check what your entry looks like on the listing page, not the profile page. If the listing page shows only your name, you are unlikely to be cited from it.
• Measure citation share, not position. Position is not a meaningful unit in an assistant response. The unit is whether you appear and which source put you there.

• Assistant responses are not stable. The same prompt can return different citations on consecutive runs and across accounts, regions and sessions. We ran each prompt three times and report the modal result, which reduces but does not remove this variance.
• Our prompt set is weighted towards Western Europe and towards B2B professional services, and results may not transfer to categories we did not test.
• Classification of source type involves judgement.
• Citation does not equal influence on the buyer. We measured what assistants cite, not what changed a purchase decision.