AI can quote you. Big deal. Can it vouch for you?

14 min read

Editor’s note: The figures here come from named studies, Ahrefs, Semrush, Seer Interactive, Adobe, Gartner, 6sense, and the peer-reviewed Princeton GEO paper, with vendor data flagged as vendor data throughout. This is a marketing analysis, not investment or legal advice.

TL;DR

• Gartner projected traditional search volume would fall 25% by 2026. It overshot, since Google held 90%+ share with AI Overviews, but the direction was right. AI answers are now a primary discovery surface.
• Adobe recorded AI-referred traffic to US retail up 693% year on year over the 2025 holidays, converting 31% better than non-AI traffic. 94% of B2B buyers now use generative AI somewhere in the purchase (6sense).
• Ahrefs studied 75,000 brands: branded web mentions correlate with AI Overview visibility at 0.664. Backlinks correlate at 0.218. Brand strength beats link-building by roughly 3 to 1.
• Being cited as a source and being named as the recommendation are different things. Semrush calls the gap the Mention-Source Divide, and only about a quarter of brands manage both.
• Seer Interactive’s tests suggest AI picks which brand to recommend from training-data memory first, then finds a citation to justify it. The citation is the bibliography, not the brainstorm.
• We call the machine-visible version of brand equity AIBE. It is measurable across five dimensions: comprehension, citation, recommendation, consistency, and substrate.

Your Brand Has a Second Reputation Now

For twenty-five years the job was to rank. A query went into a search box, ten blue links came back, and marketing’s work was to be one of them. That surface is shrinking. Gartner projected in early 2024 that traditional search volume would drop 25% by 2026. The prediction overshot, because Google absorbed much of the shift with AI Overviews and kept more than 90% of the market, but the underlying move was real. People increasingly get an answer instead of a list.

The commercial weight behind that shift is measurable. Adobe’s holiday 2025 data put AI-referred traffic to US retail sites up 693% year on year, converting 31% better than everyone else, a pattern we traced in our piece on agentic checkout and AI discovery. On the B2B side, 6sense’s buyer research found 94% of buyers using a generative AI tool somewhere in the purchase. Many of those interactions never touch a website at all: Semrush puts 93% of Google AI Mode sessions ending with no click.

Here is the part that should worry a brand team. When the answer arrives pre-assembled, your brand is either inside it or invisible for that buyer. There is no page two to climb. Visibility becomes something a model decides, from what it already learned about you, in the moment it writes the sentence.

Mentioned, Cited, or Recommended

Not all AI appearances are equal, and conflating them is the most expensive mistake in this space. A model can do three different things with your brand:

• Mention it, naming you somewhere in an answer.
• Cite it, footnoting your page as a source.
• Recommend it, telling the buyer to choose you.

These come apart in practice. Semrush documented a pattern it calls the Mention-Source Divide, where AI systems pull facts from your content while naming a competitor as the answer. Only about a quarter of brands in its analysis achieved both citation and recommendation. Seer Interactive found the same shape across hundreds of thousands of AI responses, describing “ghost citations” where a model cites your URL and recommends someone else in the same breath.

The strategic reading is uncomfortable. A rising citation count can mean your content is getting more useful to the model while a competitor collects the recommendations it earns you. Being quoted is worth more than being mentioned, and being recommended is worth more than either.

The Number That Should Move Your Budget

So what makes a model reach for one brand over another? The most cited data point comes from Ahrefs, which analysed 75,000 brands to see which signals predict inclusion in Google’s AI Overviews. Branded web mentions came out on top at 0.664. Backlinks, the metric the SEO industry organised itself around for two decades, sat at 0.218.

Ahrefs is careful, and so should we be. The company states plainly that correlation is not causation. Large, well-known brands naturally have both more mentions and more AI visibility, so the mentions may be a proxy for brand strength rather than a lever on it. That caveat is the honest centre of this whole topic, and it points the same direction either way: what predicts AI visibility is how much the wider web talks about you, not how many links you bought.

Seer Interactive pushed the point further. Across a large body of behavioural tests, its researchers found evidence that a model decides which brand to recommend from what it already knows, its parametric memory, and only then goes looking for a citation to back the choice. If that holds, and it has not been independently replicated at that scale, citation optimisation alone cannot rescue a weak brand. The recommendation was settled before the search ran.

There is a content lever that does work, and it is measured rather than asserted. The peer-reviewed Princeton GEO study (Aggarwal et al., ACM KDD 2024) ran roughly 10,000 queries and found that adding statistics, quotations and citations to a page lifted its visibility in AI answers by 25 to 41%. Keyword stuffing, the old reflex, did nothing.

AI Brand Equity, Defined

Brand equity is not a new idea. Kevin Lane Keller formalised customer-based brand equity in the Journal of Marketing in 1993, building on David Aaker’s 1991 work: the value that accrues when people recognise a brand, associate the right things with it, judge its quality highly, and stay loyal. For thirty years that equity lived in human memory, measured through surveys and brand-tracking studies.

A second copy now lives somewhere else, in the weights and retrieval systems of the models people ask for advice. Call it AI Brand Equity, or AIBE: the degree to which AI systems understand your brand, cite it, and recommend it. It is built from the same raw material as human brand equity, a broad and credible and consistent presence, but it is read by a machine and expressed as a recommendation rather than a memory.

What makes AIBE structurally different from an SEO score is where it comes from. A ranking is earned on your own page. AIBE is earned mostly off it, in the reviews, community threads, podcasts, press and video transcripts a model absorbs during training and retrieval. It behaves less like a search result and more like a reputation.

The AIBE Framework: Five Dimensions

Five dimensions, each measurable, each with a job. Score a brand across all five and you can see which part of its machine reputation is failing.

1. Comprehension

Action: check whether the model describes you accurately. Why: a model that misstates what you sell cannot recommend you correctly, and Seer’s work suggests it recommends from this stored understanding before it cites anything. How: ask the major models to describe your brand, category and differentiators cold, and log every error. Misconceptions here cap everything above them.

2. Citation

Action: measure how often your pages are used as sources. Why: citation is the machine-readable proof that your content is useful, and the Princeton data shows it is the thing on-page work can actually move. How: track your share of citations across a fixed prompt set, then add statistics, quotations and clear sourcing to the pages you most want pulled from.

3. Recommendation

Action: measure how often you are named as the answer, not just the source. Why: this is the dimension that converts, and it is the one most decoupled from citation. How: run category prompts like “best X for Y” and record whether you are recommended, ignored, or used to justify a competitor. This is your share of voice inside the answer.

4. Consistency

Action: check whether you appear across models, not just one. Why: the citation layer is fragmented. Kevin Indig’s analysis of 3.7 million citations found only 2.37% of cited URLs appeared across all three major LLMs, and 91% appeared in just one. How: measure the same prompt set on ChatGPT, Gemini, AI Overviews and Perplexity separately, and treat any single-model number as a floor rather than a total.

5. Substrate

Action: build the off-site signals the other four rest on. Why: Ahrefs’ correlations put branded web mentions (0.664), branded anchors (0.527) and brand search volume (0.392) as the top predictors, all earned off your own domain. How: invest in editorial coverage, credible community presence, review platforms and video, the places models actually read. Publishing more pages on your own site does not move this number.

Where This Thesis Could Be Wrong

This is a young field, measured mostly by companies selling tools into it, and the honest version keeps the contradictions in view.

The first problem is that most “AI optimisation” tactics do not survive testing. C-SEO Bench (Puerto et al., 2025), the first systematic benchmark of conversational-SEO methods, found that most of them do not help and several actively hurt, while plain source relevance keeps working. When several brands adopt the same trick at once, the edge cancels out.

The second is causation. Ahrefs’ own caveat means the 0.664 figure may simply be telling us that strong brands are strong. If so there is no shortcut, only the slow work of becoming a brand worth mentioning.

The third is instability. Because the numbers fragment across models and shift week to week, any single AIBE score is soft. It is a direction of travel, not a share price. And Seer’s post-hoc finding, if it holds, means the biggest lever is brand authority itself, which no content calendar builds in a quarter.

None of this makes the measurement pointless. A brand that scores itself across the five dimensions at least knows which one is failing, which is more than most teams can say today.

Where to Start

Pick ten prompts a real buyer in your category would type, the “best tool for,” “alternatives to,” “is X worth it” questions. Run them across ChatGPT, Gemini and Google’s AI answers. For each, write down three things: did the model describe you correctly, did it cite you, did it recommend you. Most teams find they are cited and not recommended, or described wrong and cited anyway.

That hour tells you which of the five dimensions is broken, and it costs nothing but attention. For adjacent reading, we have covered what happens to paid media as AI agents take over discovery, how AI is reshaping the buy button itself, and keeping a brand safe when machines mediate what people see.

FAQ