AI Search Brand Citations: Engines Pick You, Then Skip Your Site

A Babel42 study published on 4 November 2026 found that five AI assistants often overlap on which brands they recommend but lean on very different websites. Only 15% of 11,147 brand recommendations cited the brand's own site. For enterprise teams tracking AI search brand citations, a recommendation and a citation are separate outcomes that must be measured engine by engine. The data covers US English only.

Key facts: Babel42 asked five assistants 257 US English questions across 10 industries from 1 to 3 November 2026; only 15% of 11,147 brand recommendations cited the brand's own site, falling to 2% in beauty; when all five named brands, they shared at least one 77% of the time but the same first pick only 27% of the time; in 286 cases where all five cited sources, just 23 had a single site in common.

What did the Babel42 study find?

The assistants overlap on brands far more than they overlap on sources. Babel42's 2026 citation study put the same 257 questions to ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode once a day for three days. It recorded 3,776 answers, every cited site and every recommended brand.

Each question began as a real, high-volume Google search, rewritten as a chat question. The 10 industries ranged from beauty and personal finance to legal services and sportswear.

Babel42 found the source gap is wide:

• Google AI Overviews cited YouTube in 76.2% of answers and Reddit in 39.2%, according to Babel42.
• Perplexity's top source was forbes.com at 42.4%. Babel42 found it cited YouTube in 0.4% of answers and Reddit in none.
• ChatGPT cited neither YouTube nor Reddit. Babel42 reports its top source was nih.gov at 7.3%.
• NerdWallet, in personal finance, was the only site in the top three for all five assistants in any industry, per Babel42.

Babel42 also found that ChatGPT and Google AI Mode shared at least one cited site for only 23% of questions, the lowest pairing. AI Overviews and Gemini shared one 83.1% of the time.

Do AI engines really agree on brands?

They partly agree, and how much depends on how agreement is measured. Babel42's 77% figure means at least one brand in common. The same first pick appeared only 27% of the time.

Sitefire's analysis of 650,545 AI answers over roughly seven weeks measured brand overlap for the same prompt on the same day at 30% to 36%, depending on the pair. Cited-page overlap was just 6% to 10%. Between ChatGPT and Gemini, there was no web page in common 60% of the time.

So "broad agreement" on brands is real but loose. Agreement on sources is weak by any measure.

Why is your own site so rarely the cited source?

The studies suggest engines lean heavily on third-party content when they recommend a brand, though none proves that third-party pages cause recommendations. All the research is observational.

The AirOps report on offsite signals, from October 2025, found that 85% of 21,311 brand mentions came from external domains and 13.2% from the brand's own domain. Nearly 90% of those third-party mentions came from listicles, comparisons or reviews. AirOps says third-party validation amplifies credibility but adds that owned content remains the baseline.

Sitefire offers a supply explanation. There is one Wikipedia and one Reddit, but tens of thousands of company sites could answer a given prompt, so models converge on reference sources and scatter on company sites. On an ultra-niche prompt about tennis-court domes, both models cited the same five vendor sites.

How much do the studies disagree on owned-site citations?

They disagree, and their metrics are not like for like. Treat each figure within its own method.

• Study - Data window - Scope - Owned-site measure
• Babel42 - 1 to 3 Nov 2026 - 5 assistants, US English - 15% of brand recommendations cited the brand's site
• Dankrank - Benchmark dated Sep 2026 - 4 engines, agency-selection prompts - 9.7% (ChatGPT) to 40.2% (Perplexity) of brand appearances
• Sitefire - About 7 weeks - 3 models - 46% to 57% of citations were any company website
• AirOps - Oct 2025 - 3 models - 13.2% of mentions from the brand's domain
• Yext - Jul to Aug 2025 - 6.8M citations - 44% first-party websites

Yext pushes back hardest. Its release on 6.8 million AI citations says 86% come from sources brands control, but it counts listings and reviews as brand-managed, which AirOps and Babel42 treat as third-party. Yext also sells a product that distributes brand data, so its framing is a vendor's claim. Yext and AirOps data predate current model versions and work best as background.

Why does ranking in one engine tell you little about another?

Each engine draws on a different source pool and changes at a different pace. Strong visibility in Perplexity says little about ChatGPT or AI Mode.

Citation behaviour varies sharply. Babel42 found ChatGPT cited 1.8 sites per answer and Perplexity 8.6. Babel42 also found Gemini cited any source in only 46% of answers, and that Perplexity cited 91% of the same sites the next day while Gemini cited 40%.

Sitefire reports that AI Mode's source mix moved less than 1.5 percentage points in any category over seven weeks, while ChatGPT's moved up to 8.5 points. Sitefire concludes that work aimed at AI Mode has a longer shelf life, though it cannot tie ChatGPT's shift to a specific model release.

This is why single-engine AI visibility tracking can mislead. Results need to be kept separate by platform.

How can teams map which third-party pages each engine cites?

No source offers a full mapping method, but the studies point to a disciplined, per-engine record of cited links. A practical approach:

1. Build a fixed query set from real searches. Babel42's method of rewriting high-volume searches as chat questions is a sound template.
2. Run each engine separately and repeat over several days. Stability differs too much to trust one snapshot.
3. Record every cited link and sort outcomes into four buckets. The Dankrank citation study separates the brand's own site as source, another publisher as source, a mention with no brand domain cited, and discussion without recommendation. It warns against merging these into one number.
4. Classify third-party pages by type and owner. Dankrank found 42 of 67 source owners in its agency market were agencies, often competing in the category they wrote about. That calls for ongoing competitor monitoring.
5. Separate search results from citations. Babel42 found ChatGPT's searches returned a Reddit thread for 84.8% of answers, yet no answer linked to or named Reddit. Forum activity still matters for other engines, which makes community listening worth covering.
6. Check what the cited pages say about you. Third-party pages can carry outdated or wrong details, so AI brand-perception monitoring belongs in the workflow.

Sitefire adds a caution: spreading content across more source categories will not make models agree, because they diverge most on company sites.

What does this news mean for enterprise marketing teams?

The practical shift is from asking "are we recommended?" to asking "which pages earned that recommendation, in which engine?" Babel42 found that only 15% of its 11,147 brand recommendations cited the brand's own site. If that reflects your category, most recommendations may rest on pages you do not own. Leadership reporting should reflect that.

Yext says 64% of marketing leaders surveyed are unsure how to measure success in AI search. MarkGrid builds that measurement layer. Its AI marketing platform covers the workflow above:

• Model Share tracks how often ChatGPT, Gemini, Perplexity, Claude and Microsoft Copilot mention or recommend your brand against competitors, with query-level analysis, cross-model comparison and AI citation analysis. It flags visibility gaps and the factors behind competitors' citations.
• Competitive Intel watches competitors' SEO, content, messaging and AI-search visibility, which matters when rivals publish the comparison pages engines cite.
• Brand Research tracks how models describe your brand by product, service and region. MarkGrid highlights it for accuracy monitoring in fintech and service-line monitoring in healthcare.
• Community Signals covers Reddit, Discord, Quora and niche forums.
• SEO Intelligence connects Google rank tracking with AI citation optimization to guide owned content, the baseline AirOps says still counts.

Teams working outside the US should run their own market queries rather than assume these US English findings transfer.

FAQ

What does the 15% figure actually measure?

It is Babel42's share of 11,147 brand recommendations that cited the recommended brand's own site. It covers US English questions asked from 1 to 3 November 2026. Babel42's exact definition and per-assistant rates are not fully published, so treat it as a category signal, not a universal rate.

Does the finding apply in India or other markets?

No source in this research covers India. Babel42 tested only US English, and Sitefire briefly reports little difference between English and German prompts. Teams in other markets need their own query data.

Do third-party pages cause AI recommendations?

The studies show correlation, not causation. AirOps says third-party validation amplifies credibility, but all the research is observational. Treat cited pages as evidence of what engines read, not proof of why they recommend.

Which engine is most stable to optimise for?

Sitefire found Google AI Mode's source mix barely moved over seven weeks while ChatGPT's shifted noticeably. Babel42 found Perplexity repeated 91% of its cited sites the next day versus 40% for Gemini. Stability varies by engine, so retest regularly.

Why do the studies report such different owned-site rates?

They measure different things. Some count the recommended brand's site, others count any company website, and Yext counts listings as brand-controlled. Sample sizes, prompts and dates also differ.

How reliable is the Babel42 data?

It is the most recent of these studies, but it has limits. It used API versions of three assistants, which can cite differently from their apps, and it ran for three days on high-volume questions rather than a random sample.

Last updated 12 November 2026.