Fractl tests what makes AI recommend a brand
Fractl's studies show how retrieval, skeptical prompting, and third-party sources shape which brands AI recommends.
Fractl has published two connected studies that challenge a comfortable assumption about AI visibility: getting into an answer is not the same as being trusted by the model producing it. Its controlled experiment shows how strongly recommendation systems follow retrieved sources, while its real-world analysis maps the much broader field of brands and publishers that those systems actually use.
Together, the findings suggest that AI recommendation strategy has two separate jobs. Brands need to become retrievable across the sources an assistant consults, then give newer models enough independent evidence to believe what those sources say. Visibility creates candidacy. Corroboration increasingly determines credibility.
Table of contents
Jump to each section:
- The fake-company test exposes the retrieval problem
- AI recommendations do not mirror Google rankings
- A fragmented field changes the competitive playbook
- What marketers should know about AI recommendation visibility
The fake-company test exposes the retrieval problem
In Where AI Recommendations Actually Come From, Fractl placed a fabricated company beside real competitors in synthetic search-result sets. The setup tested whether models would treat retrieval as evidence or apply their own checks before making a recommendation.
6,048 runs across nine AI models tested how recommendation behavior changed with brand reality, source position, content strength, and skeptical prompting.
The default behavior was permissive. Several systems treated presence in the retrieved set as enough reason to include a brand, even when that brand had no website, customers, or prior web footprint.
Six of nine models recommended the fabricated company in 94% to 100% of runs when its page appeared beside five real competitors under the default framing.
Skeptical instructions changed the outcome. When models were told to recommend only providers they were confident existed, nearly every system began filtering the fake while continuing to include genuine smaller companies.
Eight of nine models cut fake-company recommendations to 4% or less under skeptical instructions, while gpt-oss remained the exception at 78%.
Real small companies were still recommended in 85% to 100% of default runs under comparable conditions, showing that models were capable of distinguishing unfamiliar businesses from invented ones.
That distinction matters because it reframes the risk. The problem is not simply that AI favors famous brands. It is that some systems accept retrieved inclusion as a proxy for legitimacy, while newer or more cautious systems look for corroboration.
Retrieval gets a brand into the room. Verification decides whether it gets to speak.
Lena Marlowe
AI recommendations do not mirror Google rankings
The companion study, What AI Actually Recommends, moves from controlled conditions to live shopping and buying questions. It analyzed brand recommendations and citations across a wide range of commercial categories, then compared those sources with ordinary Google results.
11,573 AI answers across 15 industries were used to measure which brands appeared and which sources were cited in real recommendation responses.
The overlap with search rankings was much smaller than an SEO-led planning model would imply.
73% of AI citations came from outside Google’s first page across the queries studied, putting much of the recommendation source pool beyond the traditional top search results.
The common assumption is that stronger Google rankings will naturally translate into stronger AI recommendation visibility. The contrasting reality is that assistants often retrieve Reddit threads, YouTube videos, niche reviews, and specialist publishers that do not occupy those rankings. The strategic implication is not that SEO has stopped mattering, but that page-one performance measures only part of the evidence environment shaping AI answers.
AI visibility is not a new ranking layer placed neatly on top of search. It is a separate distribution problem with overlapping inputs.
A fragmented field changes the competitive playbook
The real-world data also pushes back on the idea that AI recommendations will consolidate around a small set of category leaders. Instead, assistants produced a broad and variable field of brands, leaving meaningful room for challengers with current, retrievable evidence.
Category leaders captured only 2% to 8% of recommendations in the industries studied, indicating that no single brand held a dominant share of AI-generated shortlists.
It took 17 to 76 brands to account for half of recommendations within a category, showing how widely recommendation share was distributed.
This is strategically important for smaller brands. Historical fame can help when sources provide weak or conflicting evidence, but the retrieved set remains the primary candidate pool. A challenger can therefore earn recommendation share through credible coverage even when it lacks an incumbent’s training-data footprint.
The opportunity is broad, but it is not random. Fractl found that Reddit and YouTube repeatedly surfaced alongside industry-specific publishers, review sites, and comparison resources.
Reddit and YouTube ranked among the top three cited sources in nearly every industry tested across the study’s 15-category sample.
That pattern makes source strategy more specific, not less. Brands need to understand which independent venues shape recommendations in their category, then build evidence that belongs naturally in those venues. A generic content-volume target cannot answer that question.
What marketers should know about AI recommendation visibility
The combined research points toward a more demanding model of AI visibility, one that separates presence, persuasion, and proof.
Map the retrieved set. Identify the publishers, comparison pages, communities, and videos that assistants cite for real buyer questions in your category. Those sources define the practical candidate pool.
Measure repeatedly across models. Recommendation behavior varies by system and prompting. A single answer can reveal a citation, but it cannot establish stable visibility.
Build independent corroboration. One favorable page may persuade a permissive model, yet cautious systems can discount isolated promotional evidence. Reviews, editorial mentions, and consistent third-party descriptions make a claim easier to verify.
Keep SEO and AI visibility separate in reporting. Search rankings remain valuable, but they should not stand in for recommendation share or citation presence. Each measures a different route into discovery.
The deeper shift is that distribution and credibility are becoming inseparable. Marketers once treated third-party coverage as support for a message largely controlled on owned channels. AI assistants can now use that external evidence to assemble the shortlist itself.
This also raises the standard for earned visibility. If weakly vetted systems can repeat a fabricated claim, simply appearing in an answer is not a sufficient success metric. The more durable objective is to be recommended for reasons that survive a skeptical prompt and a buyer’s own verification.
AI recommendation strategy is therefore moving beyond mention counting. The advantage will belong to brands that can make themselves easy to retrieve, useful to cite, and difficult to doubt.
