92.8% of dialogues in ChatGPT end without a transition to the website — this is Profound data that Kevin Indig analyzed in the digest on July 21. Measured on live traffic of a real product, the figure can be trusted. All search promotion relied on a single deal: you get to the top of search results, a person clicks the link to your site, and converting that visit into a lead is your concern. The link was the point where a business first touches a customer. In the AI response, this deal breaks at the very first step: the click most often simply doesn’t happen.
Why 92.8% of dialogues in ChatGPT end without a transition to the website
For several months, I’ve been measuring how ChatGPT, Perplexity, Alisa AQI, and GigaChat name brands in their responses: I run control prompts and see which model mentioned whom and which didn’t. Behind these runs, a figure surfaced that breaks the usual SEO logic — and made me change what I track for clients altogether. Let me break it down in order.
Three weeks ago, we ran prompts for a client in the food delivery niche through ChatGPT. Usual routine: you type a question like «which delivery service to choose in Moscow,» and look at the response. At that point, the client had far more links from external sites than the nearest competitor. In Ahrefs — a service that calculates a site’s link weight — this is immediately visible: both the overall authority score is higher, and the number of different sites linking to it is one and a half times greater. ChatGPT named the competitor. The client — no. Didn’t mention it at all, not once in twenty prompts.
I double-checked three times, changed the wording of the question, the date of the dialogue, and the region. The result was the same. For an hour and a half, I honestly couldn’t understand what was going on. Twenty years of website promotion taught a simple rule: the more links point to you, the higher you rank in Google and Yandex results, and therefore the more readily neural networks should name you. Here, the rule didn’t work. Only there was no search results page in the usual sense. There was a model that decided on its own whom to name and whom not to. And it decided not by a link counter.
A brand mention cannot be bought with links — it accumulates like reputation
At the same time, AI traffic itself is tiny in volume — Ahrefs reports that ChatGPT accounts for about 0.19% of all link transitions on the internet, compared to 42% for Google. Ridiculously small if you compare head-on. But here’s what’s more interesting: those few who do click from ChatGPT convert far more readily — conversion is about 7.1%, and the visit itself is valued 4.4 times higher than a regular one. Fewer people reach the site, but those who do — came ready to make a purchase or any other target action.
And here’s an important twist. If the decision is made BEFORE the click, inside the model’s response itself, then you need to fight not for the click. You need to fight for how you’re named at the moment when the click hasn’t happened yet and, most likely, won’t happen at all. A mention in a ChatGPT response is not structured like a spot in search results. A spot can be earned by working on the site — links, loading speed, technical cleanup. A mention can’t be bought that way. The model gathers it from two sources: from the texts it was trained on and from pages it pulls up right during the response. And what decides is not the number of links, but how the brand is generally written about — in reviews, on forums, in industry articles, in comparisons.
Search promotion used to be about taking top positions, because a position led to a click on the site. Promotion in AI answer engines works differently: you have to fight for the mention itself, because a click may not happen at all, and the person will still make a decision — based on how the model named the brand. A link is a one-time asset. It works as long as the page holds in the search results, as long as the site is alive, as long as the algorithm values it. A mention accumulates differently: it is reputation that has already settled into the texts on which the model was trained and which it reads now. Remove one article — the mention will not go anywhere; it is spread across dozens of sources.
This is where my mistake with that food delivery client was. His link profile was strong, but almost all of it came from aggregators and directories, where the brand is simply listed in a list. The competitor had fewer links, but had three detailed editorial reviews describing how the service differs from others. The model was trained on them, among other things. It did not count links — it read context. The first place in Google, by the way, does not at all guarantee a mention in ChatGPT. We checked this on several clients separately: search and AI engines have different selection mechanics, and the first place in regular search results does not by itself turn into a line inside the model’s answer.
Where a link still brings a client, and where the question is closed right in the answer
The link has not died entirely — it is important not to confuse this. In regular Google, 68% of searches in the US by early 2026 also end without a click, according to SparkToro citing Search Engine Land. An answer without a click to the site has also come to good old search, it is just that it was visible there earlier — through blocks with ready answers, knowledge cards, and direct answers at the top of the results.
The link still works where the decision is made AFTER the first touch: price comparison, reading reviews, technical documentation, everything that requires going into details that simply do not fit in the model’s short answer. It works in niches with a long decision-making cycle — there, a person will still reach the site, just later. The link does not work where the question is closed with one phrase in the answer. «Which delivery service to choose» — closed with a phrase. «How much does delivery from a specific restaurant in a specific area cost tomorrow evening» — not closed, a click is needed there. The same business lives on both sides of this boundary — it all depends on which specific query the user asked.
What I now measure for clients instead of search positions
I stopped looking only at positions in the search results and added a separate measurement: how often and in what context models mention the client’s brand directly, without a click, without a transition. We run control prompts — real audience questions — through ChatGPT, Perplexity, Alisa AI, and GigaChat, and see whether the model names the brand, in what order relative to competitors, and with what wording. ChatGPT is not the only one here. Perplexity builds its answer almost entirely from cited sources and shows them as a list — the model readily takes from articles on third-party platforms. Alisa AI mixes in fresh Yandex results, GigaChat answers in its own way. A brand that sounds in one model may be silent in another — so one ChatGPT is not enough for the measurement.
From this, three things emerged that I now advise any marketer to track:
- Share of mentions among top competitors for your own key questions: a specific percentage for a specific set of prompts, without general ratings.
- Tone of phrasing: the model may mention the brand neutrally, or it may add an evaluative word like «trusted» or «expensive» — this influences choice more than the mere fact of mention.
- Sources from which the model apparently draws this phrasing: often it’s one or two articles — these should be prioritized for strengthening, instead of spreading the budget across dozens of platforms.
With that food delivery client, we ultimately didn’t chase new links. We wrote several detailed materials about the product on platforms that the model clearly reads and cites — and after a month and a half, the mention appeared, and visibility in AI responses grew. The client had more links, but ChatGPT named a competitor. Now we know why.
Frequently Asked Questions
How to measure brand mentions in ChatGPT?
Run control prompts — real audience questions — through ChatGPT, Perplexity, Alice AI, and GigaChat. Calculate the share of mentions among top competitors, record the tone and the sources from which the model draws information.
Can you buy a mention in neural networks with links?
No. Links work in classic search, but neural networks don’t count them directly. A mention depends on how the brand is written about in reviews, articles, and forums — it’s reputation that accumulates in the texts on which the model is trained.
What to do if the brand is not mentioned in ChatGPT?
Find the platforms that the model cites most often and place detailed materials about the product there. Strengthen editorial reviews, comparisons, and expert articles — they influence mentions more than link mass.
Should you abandon SEO in favor of neural networks?
No. Links still work in niches with a long decision-making cycle and for transactional queries. The optimal strategy is to combine classic SEO with work on mentions in AI responses.
Conclusion
The bottom line is simple: 92.8% of dialogues in ChatGPT end without a click, but the decision to choose a brand has already been made. A link is a one-time asset; a mention is cumulative reputation. To win in AI search results, stop chasing links and start managing how you are written about in the texts that models read. Run control prompts, track the share of mentions and tone — and you’ll see where you’re really losing customers. Start with an audit of your mentions today.
