Generative AI is transforming purchasing journeys. For a brand, tracking its AI share of voice is not enough: it is necessary to analyze omissions and weak signals to be recommended with confidence
For several months, marketing departments have had to respond to a new problem: “how to be cited by ChatGPT?” » and other generative search engines. These tools have changed users’ search habits and purchasing journeys: most of the purchasing decision now takes place before the prospect arrives on the company’s website. AIs are discovery, comparison and recommendation interfaces.
In this context, it has become crucial to know if your brand is present in the conversation when a prospect asks about generative AI in their market. The “AI share of voice” (Share of Voice / Share of Model) metric has become one of the key metrics to know if your brand is present or absent in the responses. But this indicator has an important limitation: it measures whether the brand is cited, but not how it is cited. This qualitative information is, however, more strategic and reveals more operational lessons for a marketing team.
AI share of voice is a vanity metric that is not very actionable
AI share of voice has established itself as the first GEO KPI (i.e. optimization of visibility in responses generated by AI). However, it is often used as a vanity metric which does not really improve the visibility or perception of a brand in AI.
The share of voice misses two important pieces of information (i) the importance of the quote: being cited on a question with strong purchasing intention vs. on a question that explicitly cites the brand does not have the same value, (ii) the quality of the quote: the brand can be cited before or after its competitors, be presented as a reference or as a solution to avoid. This qualitative richness is what matters most when evaluating a brand’s performance in generative AI, and it is not captured by share of voice.
Furthermore, the AI share of voice measurement is strongly correlated to the set of prompts on which it is evaluated. If the brand name is in the question, it will necessarily be in the answer. A question that explicitly asks for business proposals as solutions should not be evaluated in the same way as an informational question.
Example :
• Query 1: “What is brand X worth?” – the brand is cited, but because it is in the question.
• Query 2: “Which solution should I choose to solve Y?” – the brand is absent, although this is where the recommendation comes into play.
The share of voice is a relevant indicator only on a fixed set of prompts for which we observe the variation over time and which we compare to that of its competitors, but it remains insufficient.
AI sentiment scores also have limitations
Classic brand sentiment indicators aim to fill in the gaps in “AI share of voice” by providing this qualitative analysis. But certain precautions are necessary to prevent their interpretation from leading to errors.
First, generative models are trained to produce balanced, careful, and non-defamatory responses. It is therefore very rare for them to directly say bad things about a brand. They will rather tend to formulate reservations, operate by omission, or avoid mentioning the brand.
Thus, an AI brand sentiment perceived as neutral may be a false positive. Neutrality in an AI is in reality often akin to a lack of available evidence, low authority, mixed sentiment or opinions. Having a “neutral sentiment” because the AI makes no criticism of the brand is problematic when competitors are described as “benchmarks” or “offering the best ROI”, then neutrality is a competitive disadvantage.
An AI visibility strategy should be driven by omissions and weak signals
The best way to build an action plan for your AI visibility is not to track generic metrics, but to focus on omissions and weak signals to detect what the brand is missing in order to be favored by the models.
What a model doesn’t say about a brand is almost as important as what it does. This makes it possible to identify themes and messages where the brand has a deficit of authority. It is necessary to audit: key queries on which the brand is absent, expertise or strong evidence of the brand which is not included by the AI.
Then, we must note the weak signals or nuances introduced by AI which reduce the persuasive force of the recommendation. Certain formulations such as “more oriented towards…”, “may be relevant if…”, “to be verified according to…” following a recommendation are in reality to be interpreted as a reservation. If they are not “negative” in the classic sense of the term, they indicate a lower level of confidence of the generative model in the brand.
For example, a brand that explicitly targets large companies could be evoked by an AI as “X is a recognized reference for large companies” or “X may be an interesting option depending on the needs”. Both mentions are positive on the surface, but the second is actually a signal of a perceived lack of authority for AI.
Likewise, one way to detect more weak signals is to identify processing differences with competitors. The way in which a model highlights a competitor (adjectives used, precision of the description, associated use cases, social proof mentioned, position in the response, etc.) makes it possible to identify the gaps that the brand must fill in order not to be left behind in the response.
The topic is not to replace the AI share of voice with a new sentiment score, but to analyze the way in which AIs qualify, prioritize or exclude a brand from the conversation.
What marketing teams can actually do
The omissions and weak signals identified provide specific topics to reinforce so that the authority perceived by the Generative AI is improving. For brands, it is a question of densifying their semantic footprint and their proof of authority on these subjects in order to become essential for generative AI.
Concretely, each of the key messages must be present in a sufficiently dense manner both in the brand’s discourse but also in third-party discourse. Generative AI implicitly compares several layers of discourse: what the brand says on its site and social networks, but also what third parties say about it.
They penalize inconsistencies between the different layers of discourse. A promotional distinction present in the brand’s discourse, but which is not taken up by third-party sources, or even contradicted, has very little chance of being identified as reliable by a generative AI and therefore of being taken up.
Brands must succeed in strengthening their evidence and authoritative content supporting their communication objectives on the different sources of influence of AI. Each inconsistency or misalignment between sources leads the model to have to arbitrate and therefore to risk losing control of its narrative.
Ultimately, brands that only manage their AI visibility with generic indicators risk missing out on real opportunities to improve their image in generative AI. The question is no longer just whether a brand is cited, but whether it is confidently recommended. On this new battlefield, the ability to identify weak signals will be the key to focusing efforts on decisive actions.
