The Listicle Layer: Google AI Overviews cites third-party blogs 4x more than the products it recommends

Named on stage. Missing from the source list.

The clearest pattern in the dataset is the distance between what Google AI Overviews recommends and what it cites. Across 100 queries, the AI Overview named a total of 445 products as recommendations. In only 125 cases (28.1%) did that product’s own website appear anywhere in the source list. The remaining 320 recommendations, roughly seven out of every ten, were named without a single vendor-side citation. This is the pattern we call a ghost recommendation.

The distribution is not uniform. In 26 of the 100 searches, Google AI Overviews cited none of the products it recommended. In just 3 searches did every recommended product get its own site cited. The rest sit in the middle, with one or two vendor sites appearing alongside a much larger set of third-party listicle URLs.

Ghost recommendations in the wild

Real examples of products Google AI Overviews named without a single citation to their own site

“best crm for small business” · 4 products recommended 0 cited

HubSpot Monday Pipedrive Zoho

“best hr software” · 7 products recommended 0 cited

BambooHR Deel Gusto HiBob Remote Rippling Workday

“top crm tools 2026” · 5 products recommended 0 cited

HubSpot Microsoft Dynamics 365 Pipedrive Salesforce Zoho

“best marketing automation for b2b” · 4 products recommended 0 cited

ActiveCampaign HubSpot Marketo Pardot

The citation layer is a listicle economy

If seven out of ten recommendations are ghost picks, where are all the citations going? Pooled across every source Google AI Overviews cited in these 100 queries, third-party “best-of” blogs account for the majority. The vendor’s own site is a distant second. Review sites like G2 and Gartner, together with Reddit, YouTube, and news outlets, split the remainder. This is the source hierarchy that LLMs are quietly optimising for.

The reading is straightforward. Buyers who assume Google AI Overviews is triangulating the products it recommends against those products’ own documentation are almost always wrong. The engine is triangulating against a small ring of category listicles that already exist on the open web. If a company is not named in those listicles, it is very unlikely to be recommended, no matter how much first-party content it publishes on its own site. This is a signal-level distinction we cover in depth in the signals LLMs use to select sources beyond E-E-A-T.

The category divide, from decoupled to coupled

The 71.9% headline is an average. Underneath it, categories divide sharply into four groups. In seven categories, the AI Overview cites vendor sites for one in ten recommendations or fewer. In five categories, the coupling is moderate. In seven, it becomes meaningful. And in exactly one, it flips: the citation and recommendation layers align. The pattern is inversely related to how deep the listicle ecosystem is in that category.

The seven “strongly decoupled” categories are the largest, most mature B2B SaaS markets: accounting, CRM, HR, marketing automation, application monitoring, business intelligence, and iPaaS. In all seven, Google AI Overviews is essentially reading a listicle to build its recommendation set. The category leaders (HubSpot, Salesforce, Workday, Datadog, Tableau) get named. Their own sites are almost never the source.

The one category where the pattern breaks

QuickBooks hosting is the exception. Across five queries in the category, the AI Overview cited 70% of its recommended providers via their own site. The list of “coupled” companies reads like the vendor roster itself. This is the only category in the dataset where a buyer can trust that the source list underneath the answer maps cleanly onto the products the answer names.

QB Hosting self-citation

70%

2.4x the category average of 29%
The only category that inverts the listicle pattern

Why this category behaves differently.

QuickBooks hosting is a narrow, high-consideration niche. The dominant content type on the open web is vendor content itself: comparison pages published by the hosting providers, feature grids, and support documentation. The listicle economy in this space is small and thinly authoritative, so Google AI Overviews defaults to what it can retrieve directly from the vendors.

The takeaway generalizes. The listicle layer’s power over recommendations grows with the size of the listicle ecosystem in a category. In small, niche categories, first-party vendor content still shapes the answer. In large, mature categories, the vendor is invisible to the source list even when it is the product being recommended. We’ve seen this play out in practice with our client Verito, one of the hosting providers cited most often in this space.

Providers cited via their own site

Verito Rightworks Ace Cloud Hosting Summit Hosting Sagenext

How many sources back an answer

The volume of citations per answer varies as much as their composition. Some AI Overview answers cite just 3 sources for 4 products (barely one per recommendation). Others cite 46 for 5 products (more than nine per recommendation). Averaged across the dataset, an answer cites 12.6 sources to back 4.5 products, roughly 2.8 sources per recommendation. But the average hides three different behaviors, and the time it takes to get citation coverage varies by engine.

A weak positive relationship exists between the length of an answer and the number of sources cited (r ≈ 0.49). But the number of products recommended has essentially no effect on the number of sources cited. A longer answer often means more citations pointing at the same short list, not more products backed up individually.