Ahrefs studied 75,000 brands to find what predicts appearance in Google's AI Overviews. The strongest signal was not backlinks, which correlated at 0.218, nor a brand's own domain content. It was branded web mentions across the open web, correlating with AI visibility at 0.664. Ahrefs measured AI Overviews specifically, but the mechanism generalises to any engine drawing on the same record: the model reads the open web to decide whether your brand appears, and reads the same material to decide how to describe you when it does.
The description is the part most teams never measure. When a model calls a brand "widely recommended," or "a smaller regional player," or "solid for mid-market but less proven at enterprise scale," it is drawing that from third-party reviews, editorial coverage, forum threads, and competitor comparisons. The brand's own site is not where that judgement is formed.
For a UAE challenger competing against a well-covered global incumbent, this is structural. A thin third-party footprint gives the model less to corroborate, so it describes the challenger more cautiously by default. The effect compounds: Ahrefs found the most-mentioned brands earned up to ten times the AI presence of the quartile below. Presence and framing move together, because both are read from the same record.
Every time a model names your brand in a buyer-intent prompt, the mention lands as one of three outcomes, each handing the buyer a different starting position before they have spoken to anyone or visited your site.
| Framing outcome | Signal language the model uses | Buyer's starting position |
|---|---|---|
| Endorsement | "Widely used by enterprises in this space," "consistently recommended for this use case," "strong track record in regulated industries" | Enters with positive prior. Evaluates to confirm. |
| Neutral listing | "One option to consider," "also offers this capability," "available in this region" | No prior. Evaluates from scratch, with the framed competitors already holding relative advantage. |
| Cautious hedge | "May be better suited to smaller deployments," "less established than alternatives," "limited independent reviews available" | Enters with doubt. Evaluates to disprove, which most buyers will not bother doing. |
Forrester's 2026 State of Business Buying report found that 19% of B2B buyers felt less confident in a recent purchase decision after encountering unreliable AI information, rising to 28% among procurement professionals. That finding is specifically about factual error. Cautious framing operates in the same register by lowering confidence early, even when the model has made no factual mistake. It has simply read a thin third-party record and responded accordingly.
The practical reason teams focus on visibility is that it has always been easier to count. A brand either appears or it does not. But modern AI visibility tooling has made sentiment just as measurable. HubSpot's AEO tools score sentiment on a scale from -100 to +100 across ChatGPT, Gemini, and Perplexity, alongside visibility score and share of voice. A brand can hold a strong visibility score and a weak sentiment score simultaneously, and the dashboard will show both if you know to look. HubSpot's AEO Grader weights sentiment as its highest-scoring dimension, above presence, recognition, and share of voice. That weighting is a design decision about what matters most to the buyer, not a fixed number since the product continues to evolve, but the signal is clear: the people building this tooling believe framing outranks frequency.
Within that sentiment score, there are three layers worth separating. General sentiment is the overall tone the model uses when describing the brand across a prompt set. Contextual sentiment is more precise: tone by topic, where a brand might read as strong on implementation support but thin on enterprise scalability within the same answer. Source-based sentiment is the credibility and character of the third-party material the model draws on when forming that description. The source-based layer closes the loop with the Ahrefs finding directly: the same off-site footprint that predicts whether you appear also sets the tone when you do. Contextual sentiment is the layer teams almost never audit. A brand can be aware its general sentiment score has slipped and still have no visibility into which specific topics are pulling it down.
Running sentiment tracking correctly requires a few deliberate choices that most teams skip. First, log sentiment per engine rather than as a blended average. Perplexity runs live retrieval and reacts to new third-party sources faster than ChatGPT or Gemini, whose training and update cycles are longer. A widening Perplexity gap relative to the other two engines is usually the earliest signal that a new review, article, or comparison piece is reshaping how the brand is described. Catching it there, before it propagates, gives you a real window to act.
When a mention hedges, trace it to source before touching anything on your own site. Pull the citations the engine listed. Find the asset feeding the caution, whether it is a lukewarm G2 review cluster, a trade publication comparison that positions you as the secondary option, or a forum thread with unanswered criticism. Fix that asset first. The model built its description from that material, not from your domain. Editing the homepage in response to a third-party-fed hedge is like correcting a transcript by changing your notes.
Set a sentiment floor per priority prompt and review it on the same cadence as your visibility metrics. A mention that slides from endorsement to neutral will not trigger any alert on a standard visibility dashboard because you are still appearing. The brand is present. The framing has deteriorated. Without a classification system that tracks which tier each prompt lands in, per engine, per week, that drift goes unnoticed until it has compounded.
The sentiment delta is the distance between how a brand describes itself and how an AI model describes it. For most established global brands, the delta is small because the third-party record is dense enough to corroborate the brand's own claims. For a UAE B2B challenger with limited analyst coverage, few published case studies in international outlets, and a review profile that is either thin or geographically narrow, the model defaults to caution because corroboration is sparse. It is not making a judgement. It is reflecting the footprint it can read.
Closing the delta is a third-party signal problem. The inputs that move it are analyst mentions and firm categorisations, review coverage on platforms the model's retrieval layer indexes, consistent language in trade publications that matches the positioning the brand wants to hold, and technical comparisons where the brand is present and credibly represented. Where coverage is thin, the model hedges. Where coverage is corroborated across multiple independent sources, the model endorses. This is not a content marketing argument dressed up as an AI one. It is the same correlation Ahrefs found for visibility, applied one layer deeper to tone.
For brands in competitive verticals like SaaS, healthcare technology, or professional services, the stakes compound further. A buyer researching two vendors and receiving one endorsement and one cautious hedge from the same AI answer will not necessarily investigate further. The prior is set. The evaluation begins unequal.
Stop reporting mention rate as the primary signal. Start classifying the language of each mention: endorsement, neutral, or cautious hedge, per engine, per prompt, per week. When a priority prompt drops a tier, that is the signal to act, and the work begins with identifying which third-party source fed the shift.
This is where Oxygen's AI visibility auditing and optimisation sits: finding where a model frames a brand cautiously, tracing the sources feeding it, and changing what the model reads around the brand. The work is rarely on the website. It is in the third-party record the model trusts.
A model's answer is now often the first substantive impression a buyer forms of a brand, assembled from sources the brand does not own, cannot directly edit, and may not know are being read. That impression is a marketing responsibility, and it is now measurable at the prompt level. Most teams are still watching whether they appear. The number that matters is how they are described when they do.