Why qualifying is not winning in AI Visibility.

You've successfully optimised your website's schema, organised your content for better large language model (LLM) retrieval, and now, the AI is mentioning you more in its responses.

Congratulations, you've qualified to appear in the AI conversation. But did you survive contact with the buyer?

Marketing teams are rightly investing in first-generation visibility tools to monitor how, and how often their brand is mentioned by LLMs like ChatGPT, Claude and Gemini, using a battery of engineered prompts. These Generative Engine Optimisation (GEO) tools are the crucial first step for investing in AI commerce because they allow teams to see their baseline visibility, and measure change through technical optimisation, structured data and digital PR efforts.

But qualifying is not winning. In a real AI purchase conversation a lot happens after the brand makes its appearance. It's a bit like judging an athlete on qualifying for a competition, ignoring the competition itself.

Meet Nova

To measure the competition, we built Nova. Nova puts real people in front of an AI, with real purchase decisions to make, and tracks every layer of the complete, multi-turn conversation that follows.

In Nova's pilot study, we screened UK participants across a range of tech purchases, recruited the 75 who told us they were in the market for a laptop, and used Claude Sonnet 4.6, with live web search, to power the conversation.

The findings that follow are taken from this study, for this product type, in this market, with this LLM model. Whether they transfer to other markets, models and product types is a question we'll continue to explore with further analysis and more studies.

This is the first post in a series. It describes how an AI purchase conversation takes shape. Later posts showcasing Nova's insights will build on this foundation.

The Rules of the Asymmetric Shelf

If the AI only ever mentioned one brand per conversation, and the shopper chose that brand 100% of the time, we could count our mentions and call it a day.

The reality is much messier.

• In 229 of 259 cases (88.4%), a brand first appeared in a Nova message, not a shopper message (counting each brand once per conversation).

• A typical conversation put about ten different names on the table (10.1 on average, counting retailers, platforms, services and components), 3.5 of them laptop manufacturer brands.

• 2,132 mentions happened across the 75 conversations, 1,208 of them of laptop manufacturer brands, spread over 20 logos. 

The AI sets the terms of the conversation, and it does so early. Faced with that stream of names, specs, caveats and sources, a shopper cannot produce a verdict on every aspect of the AI's response. In some cases, dense AI responses can elicit pushback due to overwhelm: “Wow thats a lot condense into a table please”.

To keep the exchange moving, the shopper follows the option or enquiry that fits their immediate needs. The conversation progresses, and the brands in the race quickly narrow down, filtered by the shoppers' guidance.

The Shifting Shelf

A typical AI purchase conversation has a shape. To follow brands through it, we track three numbers at each of Nova's responses, across the sessions still active at that point. These three measures count laptop manufacturer brands, and exclude components, retailers, platforms and services: 

• The shelf: The different manufacturer brands named so far by both Nova and the shopper.

• The lobby: The number of manufacturer brand mentions in a given Nova message.

• The entrants: The manufacturer brands appearing in the conversation for the first time in a Nova reply.

Shelf, lobby and entrants

Brands per conversation by Nova message

Average number of brands per conversation at each Nova message, in three views: the different brands named so far (shelf), the brands named in that message (lobby) and the brands appearing for the first time (entrants).

ManufacturersAll groups

Nova message 1Shelf 0.8463 brands across 75 conversationsLobby 0.8463 brand mentions across 75 conversationsEntrants 0.8463 new brands across 75 conversationsNova message 2Shelf 1.95144 brands across 74 conversationsLobby 1.59118 brand mentions across 74 conversationsEntrants 1.0074 new brands across 74 conversationsNova message 3Shelf 2.89214 brands across 74 conversationsLobby 1.93143 brand mentions across 74 conversationsEntrants 0.8865 new brands across 74 conversationsNova message 4Shelf 3.17225 brands across 71 conversationsLobby 1.69120 brand mentions across 71 conversationsEntrants 0.2417 new brands across 71 conversationsNova message 5Shelf 3.30211 brands across 64 conversationsLobby 1.3385 brand mentions across 64 conversationsEntrants 0.149 new brands across 64 conversationsNova message 6Shelf 3.27180 brands across 55 conversationsLobby 1.0558 brand mentions across 55 conversationsEntrants 0.053 new brands across 55 conversationsNova message 7Shelf 3.36151 brands across 45 conversationsLobby 1.0949 brand mentions across 45 conversationsEntrants 0.042 new brands across 45 conversationsFewer than 50 conversations runningNova message 8Shelf 3.52109 brands across 31 conversationsLobby 0.9429 brand mentions across 31 conversationsEntrants 0.031 new brands across 31 conversationsFewer than 50 conversations runningNova message 9Shelf 3.5485 brands across 24 conversationsLobby 0.8320 brand mentions across 24 conversationsEntrants 0.000 new brands across 24 conversationsFewer than 50 conversations runningNova message 10Shelf 3.4569 brands across 20 conversationsLobby 0.8016 brand mentions across 20 conversationsEntrants 0.000 new brands across 20 conversationsFewer than 50 conversations runningNova message 11Shelf 3.0042 brands across 14 conversationsLobby 0.9313 brand mentions across 14 conversationsEntrants 0.071 new brands across 14 conversationsFewer than 50 conversations running

Nova message 1Shelf 1.68126 names across 75 conversationsLobby 1.68126 name mentions across 75 conversationsEntrants 1.68126 new names across 75 conversationsNova message 2Shelf 3.84284 names across 74 conversationsLobby 2.76204 name mentions across 74 conversationsEntrants 1.91141 new names across 74 conversationsNova message 3Shelf 6.24462 names across 74 conversationsLobby 3.85285 name mentions across 74 conversationsEntrants 2.23165 new names across 74 conversationsNova message 4Shelf 7.42527 names across 71 conversationsLobby 3.59255 name mentions across 71 conversationsEntrants 1.3092 new names across 71 conversationsNova message 5Shelf 8.09518 names across 64 conversationsLobby 2.92187 name mentions across 64 conversationsEntrants 1.0265 new names across 64 conversationsNova message 6Shelf 8.18450 names across 55 conversationsLobby 2.16119 name mentions across 55 conversationsEntrants 0.4726 new names across 55 conversationsNova message 7Shelf 9.13411 names across 45 conversationsLobby 2.58116 name mentions across 45 conversationsEntrants 0.8237 new names across 45 conversationsFewer than 50 conversations runningNova message 8Shelf 9.81304 names across 31 conversationsLobby 2.1065 name mentions across 31 conversationsEntrants 0.4213 new names across 31 conversationsFewer than 50 conversations runningNova message 9Shelf 10.46251 names across 24 conversationsLobby 2.2955 name mentions across 24 conversationsEntrants 0.5012 new names across 24 conversationsFewer than 50 conversations runningNova message 10Shelf 11.40228 names across 20 conversationsLobby 2.4048 name mentions across 20 conversationsEntrants 0.5010 new names across 20 conversationsFewer than 50 conversations runningNova message 11Shelf 10.64149 names across 14 conversationsLobby 2.0729 name mentions across 14 conversationsEntrants 0.649 new names across 14 conversationsFewer than 50 conversations running

Dashed: fewer than 50 conversations running at that turn. Shaded bands are our grouping of messages. Numbers under the axis show the number of conversations still running at each turn.
Source: Nova Laptop Pilot Study, UK, Aug-Sept 2026. 75 conversations, Claude Sonnet 4.6 with live web search.

In this study, the average "shelf" size consisted of 3.5 manufacturer brands per session. By Nova's third reply (typically message 6), the average conversation still running had already named 2.9 of them (214 mentions across 74 sessions). The shelf tallies everything named so far by either side, so it includes brands the shopper named. 

The size of the lobby peaks in Nova's third reply too (at 1.9 brands per conversation), and 218 of the 259 brands that ever appear in a conversation (84%) had first appeared. 

Between Nova's third and fourth response, new entrants in an average conversation fall from 0.88 to 0.24. From there, the lobby size begins a steady downward trajectory, to about one brand by its sixth reply (1.05). This is the narrowing phase of the exchange. 

You might assume the size of the shelf hits a ceiling as Nova's behaviour changes. Yet it keeps climbing, to 3.3 by its fifth reply. Nova's new entrants fall to 0.14 a conversation by then and to 0.05 by its sixth reply, but do not stop completely, and the shopper adds a few brands of their own.

The types of names that appear change too. Components and software platforms follow a similar arc to laptop brands, because they’re heavily tied up in the use cases and requirements which frame how and why the laptop brands are discussed. Retailer appearances trail behind, and their growing share of appearances in the room is a clear indication of a conversation narrowing towards a purchase.

Arrival by entity group

Brand arrival by Nova turn, by entity group

Cumulative share of first appearances that had occurred by each Nova turn.

ManufacturersComponentsPlatformsRetailers

Nova turn 1Manufacturers 24.3% · 63 of 259Components 27.6% · 29 of 105Platforms 14.2% · 20 of 141Retailers 5.8% · 14 of 24175 sessions runningNova turn 2Manufacturers 57.1% · 148 of 259Components 45.7% · 48 of 105Platforms 41.1% · 58 of 141Retailers 12.9% · 31 of 24174 sessions runningNova turn 3Manufacturers 84.2% · 218 of 259Components 71.4% · 75 of 105Platforms 65.2% · 92 of 141Retailers 32.0% · 77 of 24174 sessions runningNova turn 4Manufacturers 92.7% · 240 of 259Components 84.8% · 89 of 105Platforms 74.5% · 105 of 141Retailers 53.5% · 129 of 24171 sessions runningNova turn 5Manufacturers 96.5% · 250 of 259Components 89.5% · 94 of 105Platforms 80.9% · 114 of 141Retailers 70.1% · 169 of 24164 sessions runningNova turn 6Manufacturers 97.7% · 253 of 259Components 91.4% · 96 of 105Platforms 83.7% · 118 of 141Retailers 77.2% · 186 of 24155 sessions runningNova turn 7Manufacturers 98.5% · 255 of 259Components 97.1% · 102 of 105Platforms 85.8% · 121 of 141Retailers 88.0% · 212 of 24145 sessions runningFewer than 50 sessions runningNova turn 8Manufacturers 98.8% · 256 of 259Components 98.1% · 103 of 105Platforms 89.4% · 126 of 141Retailers 90.5% · 218 of 24131 sessions runningFewer than 50 sessions runningNova turn 9Manufacturers 98.8% · 256 of 259Components 99.0% · 104 of 105Platforms 94.3% · 133 of 141Retailers 92.9% · 224 of 24124 sessions runningFewer than 50 sessions runningNova turn 10Manufacturers 98.8% · 256 of 259Components 100.0% · 105 of 105Platforms 97.2% · 137 of 141Retailers 95.0% · 229 of 24120 sessions runningFewer than 50 sessions runningNova turn 11Manufacturers 99.2% · 257 of 259Components 100.0% · 105 of 105Platforms 97.9% · 138 of 141Retailers 97.9% · 236 of 24114 sessions runningFewer than 50 sessions running

First appearances: manufacturers 259, components 105, platforms 141, retailers 241. Cumulative share by turn. Shaded: fewer than 50 sessions still running.
Source: Nova Laptop Pilot Study, UK, Aug-Sept 2026. 75 conversations, Claude Sonnet 4.6 with live web search.

Looking at how these numbers change as the conversation progresses gives us a clear picture with distinct conversation phases. If you've qualified for visibility in AI, that visibility is spread across a shortlist, and that shortlist forms early in the conversation.

Most of what Nova says about laptops after that returns to the same few brands in the shortlist. So we see 1,208 laptop brand mentions across the 75 sessions, against 259 first appearances of a brand in a conversation. Most of that is repetition. 

As such, the question for a brand isn't just whether it gets mentioned, but whether it makes it to the shortlist in the AI's first few turns, and who else they’re sharing it with. The early shelf is where competitive substitution risk is highest. 

The shape of the AI conversation presented here is relevant to this product type, in the UK, for this LLM model and this system prompt. This may change depending on how complex information needs are for a given product, the cultural and linguistic norms among the people using the AI, and of course which LLM we use, how it's told to behave, and its underlying configurations. This is something we'll continue to explore in future studies.

The shelf, the lobby and new entrants give us a top down view of a conversation's shape, but they tell us nothing about what happens to those brands by the end of the conversation.

Silence.

A mention's real enemy is silence, not rejection. We call that the Silent Fade, where Nova brings a brand to the shopper's attention, but is met with no reaction.

When we analysed the final fate of each of the 229 manufacturer brands that first appeared in a Nova message, we looked at the best thing that happened to each of them by the end of the conversation. We then tried to bucket these fates into predictable categories.

Some of these outcomes were easy to catch, such as when a brand is explicitly selected, or engaged with but not chosen, or outright dismissed. What proves much harder is distinguishing between an ignored brand and an implicitly dismissed one. Sometimes, what looked like a rejected brand when a user left it out of a follow up, turned out to be one which just fell aside to be resurfaced later.

No amount of response coding can cleanly separate the two, because non-reaction isn't always an active signal in these exchanges. It's just how real people navigate the attrition of a natural, information-dense, and heavily asymmetric AI conversation.

Outcome ladder

What shoppers did with brands Nova introduced

Share of 229 Nova-introduced brands by outcome.

Selected

26 of 229

11.4%

SelectedShare11.4%Brands26 of 229

Engaged

68 of 229

29.7%

EngagedShare29.7%Brands68 of 229

Dismissed by shopper

17 of 229

7.4%

Dismissed by shopperShare7.4%Brands17 of 229

Dismissed by AI

1 of 229

0.4%

Dismissed by AIShare0.4%Brands1 of 229

Ignored (explicit)

43 of 229

18.8%

Ignored (explicit)Share18.8%Brands43 of 229

No explicit reaction

74 of 229

32.3%

No explicit reactionShare32.3%Brands74 of 229

0%10%20%30%40%50%

Share of 229 Nova-introduced brands by outcome.
Source: Nova Laptop Pilot Study, UK, Aug-Sept 2026. 75 conversations, Claude Sonnet 4.6 with live web search.

Outright dismissal is rare, only 18 of 229 (8%) were explicitly rejected by the shopper or Nova. Silence takes two forms in the outcome ladder. In 51% of cases the brand mention was met with silence, and in almost 4 in 10 of those cases, the brand was mentioned again and still encountered no reaction.

That scarcity of attention and the lopsided information exchange should redefine how we interpret the outcome ladder shown above:

• Silence is the most common outcome, and it should not be read as an explicit remark or sentiment toward a brand. Silence goes hand-in-hand with shoppers passively pruning the AI's information-dense replies.

• Engaging with a brand is expensive because the shopper expends a turn on a follow up against information-dense AI messages.

• Selection is uncommon (26 of 229, 11%), probably for a variety of reasons. We know from this pilot some shoppers didn't fully trust the AI's output so wanted to verify on other specific platforms afterwards, and there's plenty of other factors such as timing and budget, to name a few, standing in the way of final decision.

• Explicit rejections are also rare because they cost turns. A shopper is more likely to skip past an option rather than spend a turn saying "no", so any instances of explicit dismissals carry weight.

Every message costs the shopper

We talk about costs in the outcome ladder, but what do we mean by that? When a purchase journey moves into a linear AI conversation, it behaves differently from other channels. Each direction the message moves in leaves something behind, and requires cognitive effort.

Compare that to a traditional web browser, where options live in persistent tabs, and returning to earlier options costs a single click.

In a linear AI chat, the spotlight in the conversation drives forwards in stages, and every new message changes the state of play.

A shopper may be interested in every option in a given lobby, but following up on one can, willingly or unwillingly, be a decision to leave others behind. Returning to them later takes re-prompting or scanning the previous messages.

This is why silence is the default and not the verdict. A linear conversation has room for one focal point at a time, and everything else scrolls out of sight with each passing turn. 

So maybe you've optimised your brand into the AI conversation, but maybe your mention is met with silence and you're left behind. That changes what's worth measuring.

Implications

Being named by the AI is a very different outcome to being considered by the person making the purchase.

You need to be visible to be considered in this environment, but to see what happens next, you need four measures:

• Visibility (are you in the lobby?). GEO tools let you see this at scale.

• Hook rate (did someone walk toward it?). Whether the shopper follows up on a brand the AI names, or chooses it. In the pilot study, 41% of brand mentions were engaged or selected.

• Survival (did you avoid the silent fade?). Whether your brand stays in play as the shopper adds real-world constraints and buying needs.

• Share of decision (did they leave with you?). How often your brand is the final choice. In the pilot, 11% (26 of 229) of brands mentioned were eventually selected by the shopper. 

And, because Nova captures the full multi-turn conversation together with the model’s search, citation and recommendation behaviour, we can track those signals in granularity and see how they actually relate to real shopper behaviour.

Questions we still have

For us, this study raises as questions as it answers. These are just some of the questions we have in mind and plan to explore.

• Does timing change a brand's chances? Does it matter if a brand arrives later in the conversation? If it does, what determines timing or entry?

• To what extent can the AI influence a shopper’s prior consideration set? How often does a shopper engage with or choose a Nova-introduced brand that was not on their starting list?

• How does the AI's own behaviour shape the shelf? That includes what it searches for and which sources it draws on.

• Does the AI conversation level the playing field between larger and smaller brands, once they’ve qualified?

• Does the conversation shape which emerged in this study hold elsewhere? That means other product categories, other markets and other models.