For 20 years, brands paid their audience to intermediaries who referred the customer to them. Conversational assistants do not change the price of this toll. They change the place.
“Which coffee machine for an office of ten people, less than 400 euros, available this week?” The question is ordinary. Asked to a conversational assistant, it receives a reasoned response, with models and prices, which does not go through any brand’s website. The customer will then show up at a merchant with their decision already made, or not show up.
We readily attribute this movement to theartificial intelligence. I believe that this is a diagnostic error, and that it will be costly for the brands that cling to it.
An arrangement accepted for twenty years
The distribution buys part of its audience and obtains the other through natural referencing. In both cases, a third party decides the display order. The arrangement remained bearable for a simple reason: this third party referred the customer to the brand. The passage was expensive, but the comparison ended on the pages of the distributor, with its merchandising and its arguments. What the conversational layer changes isn’t the price of the toll, it’s where the comparison ends. It now ends in the response, from sources that the brand does not provide.
This space is already selling. OpenAI announced advertising formats in ChatGPT, Google confirmed sponsored links in its AI Overviews, Perplexity markets native placements. In France, Google launched AI Overview on July 22. The 36th SRI e-pub Observatory, published thirteen days earlier, still gave it a chance. The paid formats are not yet open, but the screen that will accommodate them is in place. The rarity will be more severe there than on a results page, since a response cites two or three brands where a page displayed ten. In September 2025, Seer Interactive measured an organic click rate of 7.9% when the brand appeared in an AI Overview, compared to 4.1% when it was absent. The measurement relates to the click and not to the sale. It nevertheless indicates the direction of movement.
We still need to be able to follow this movement. Nobody knows how to count it. The customer who compared in a conversation before typing the name of a brand in his browser goes back to direct traffic, on the most flattering line on the dashboards. As for the tools that claim to measure citations in AI responses, two of them can display, for the same brand scope, volumes separated by three orders of magnitude.
What the brands still hold
An assistant knows nothing that has not been given to him. He guesses neither a price, nor a stock, nor the compatibility of a cartridge with an old printer. He reads sources. In this area, brands have three things that neither media budgets nor computing power can produce.
Compatibility, equivalence, real use, that is to say what sellers know and what product sheets say wrong.
The availability of a reference in a nearby store, decisive information in current purchases and that no one else has.
The transaction history, which is better than the declared intentions sold by the platforms and which makes targeting reliable.
These three advantages are consummated, and this is the part of the reasoning that we hear the least. Less traffic on sites means fewer behaviors observed, fewer accounts opened, less data to cross-reference with receipts. The short-term calculation seems good, since rarer and better qualified traffic converts better. It masks a slow decapitalization: brands gain in immediate profitability what they lose in capacity to document their customers. What would make them indispensable to assistants is precisely what mediation by assistants impoverishes. The window therefore does not close because competitors would move forward more quickly. It closes by itself.
Ten questions, only one mention
In July, I asked ten shopping questions to ChatGPT, Gemini and Perplexity, in an offline session and without personalization. Which cartridge for this printer, which ribbon for this label maker, where to buy it. Questions that our sellers deal with every day.
Bureau Vallée comes out in one response out of ten, mentioned as a possible reseller among others. Never in source.
The distinction weighs more than the number. Being named in a sentence means that the model has somewhere in memory of an office automation brand. Being cited as a source means that the answer was made with your data. We have the memory.
The absence is not general, and that is what makes it instructive. On subjects where we have published content, the organization of a move for example, or the binding of a document, the brand is cited, including as a source. We exist for these systems where we have written articles. We no longer exist where it would be necessary to exhibit a catalog. Product advice is the number one reason for visiting our stores and our biggest absence in the responses.
Three projects, including one thankless
The first is to make catalogs machine readable. Structured data, clean and refreshed feeds, access to stock via API. The subject does not present itself well to the management committee and does not produce any visible effect for several months. However, it conditions everything else, including advertising, since the same enriched flows determine the free quote in a response and the eligibility for the paid formats that will be installed there. A brand that neglects its product data will pay more for the same presence, assuming that it remains purchasable.
The second consists of arbitrating on retail media. The market is not asking for better: leverage increased by 18% in the first half of 2026 according to the same edition, and 85% of advertisers surveyed by Publicis Commerce announce a new budget increase in 2026. The growth comes almost entirely from retail search, up 24%. The display that brands sell on their own inventories only gains 3%. The budget follows the request. The fact remains that building an agency requires an audience and a depth of catalog that most French brands do not achieve, in a market which will be concentrated. For the majority of brands, the realistic objective is not to become an agency, but to be the source that neither the agencies nor the assistants can bypass.
The third is to reconstruct the measure on incrementality, because a recommendation formed in a conversation leaves no path to assign. Geographic testing, control groups, incremental sales. These methods are not new and remain in the minority.
The risk weighs first on the current purchase
The threat is not distributed equally, and this must be said. Where the purchase carries an emotional charge, in fashion, luxury or interior design, the brand’s interface retains a reason to exist: the customer wants to see, compare, sometimes make a mistake. On current purchases, an agent can decide alone on the price, availability and reliability of data, then trigger restocking. The platforms are already working on purchasing protocols that will allow this.
This is my brand’s land, and the most exposed. An ink cartridge cannot really be chosen, it must be replaced. The day this decision is delegated, the brand ceases to be a criterion to become an adjustment variable.
The limits of this test
Ten questions asked one morning in July do not constitute a study. The test only says that on our core business, on that date, we were not in the response. It was enough to open an internal discussion, which was its purpose.
I also don’t know if the aides will continue to cite their sources. Nothing obliges them to do so and the interest of the platforms is not obvious. If this practice were to decline, the work described above would lose part of its visible effect. There would remain the condition of access to the advertising devices which are installed in these responses, which is enough to justify it.
What will remain to negotiate
The movement is engaged and will not reverse. Most brands will pay to be featured in the responses, and the real question is what they will have to show that day. Exposed product data, readable inventory and credible measurement enable negotiation. Without this, all that remains is to bid.
It’s a general management decision and it can be summed up in a few words: remain a distribution channel, or become the reference for the product.
