To prepare this post-summer review, we started with the announcements and numbers that had been circulating over the past few months. And honestly, LinkedIn made the picture look rather dramatic. AI referral traffic was supposedly converting four times better. Assistants had become shopping advisors. Most of all, agentic commerce had apparently already arrived.

We wanted to verify those claims, so we traced each one back to its original source. Once the more sensational language was stripped away, the picture became less spectacular, but still genuinely interesting.

The market has structured itself faster than we expected. That does not mean agents have suddenly started shopping on everyone’s behalf. What took shape over the summer was more practical: new advertising formats, direct connections to merchant catalogs, early measurement tools, and research that gives us a better view of what happens after an AI-generated answer appears.

This review is current through September 4, 2026. It focuses primarily on the United States, where many of the commercial applications are already further along, before looking at what can reasonably be inferred for France.

Laptop and phone showing conversational interfaces, surrounded by charts about the US and French markets
United States, France, organic visibility, paid media, measurement: the market’s pieces are starting to connect.Visual generated for J2A Digital.

In the United States, assistants are becoming a new kind of shopping interface

U.S. ecosystemOne conversation, several levels of integration
CompareProducts, features, and offers
TrackPricing and availability
Prepare a purchaseCart, checkout, or merchant handoff
BuildCommerce-agent architecture

These capabilities are not equivalent. Availability varies by country, merchant, subscription, and experimental status.

Since March, ChatGPT has been enhancing product-related answers through the Agentic Commerce Protocol, a technical standard designed to connect assistants with merchant catalogs. OpenAI has also announced integrations with Shopify and several major U.S. retailers, including household names such as Target, Sephora, Best Buy, and Home Depot.

Products can now be compared within a conversation using information about pricing, availability, and features. (OpenAI, March 24, 2026)

Amazon has followed a similar path with Alexa for Shopping. Since May, U.S. customers have been able to use the assistant to compare products, create buying guides, view up to a year of price history, and ask it to perform certain shopping actions. (Amazon, May 13, 2026)

Those are established facts. ChatGPT and Amazon, however, are not the only companies bringing conversation closer to a transaction.

In the United States, Gemini already displays products and comparisons in its answers, while Google is gradually rolling out agentic checkout across Search and AI Mode. Perplexity offers an integrated purchasing flow for eligible products, and Microsoft Copilot can compare offers, display price history, and direct users to a merchant’s website. (Google, Perplexity, Microsoft)

Based on the official sources we reviewed, Claude is currently following a different path. Anthropic provides an architecture that companies can use to build their own commerce agents, rather than an equivalent consumer shopping experience. Its public repository is a reference implementation and does not include an integrated payment system. (Anthropic, GitHub)

The general direction spans several platforms, but the level of integration varies widely. Some assistants recommend or compare products. Others can track prices, build a cart, or prepare a purchase. Availability still depends heavily on the country, subscription, merchant, and experimental status of each feature.

For a company selling products, the change is easy to picture. Someone might ask, “What vacuum cleaner would work in an apartment with a cat for under $300?” The assistant has to understand those criteria, identify suitable products, verify their features, and assemble a useful selection.

A brand is no longer simply trying to earn a position on a search results page. Its information has to make it through that entire process without losing the correct price, market availability, compatibility, or feature that actually makes the product relevant.

For a product to be found, compared, and potentially recommended correctly, the underlying information must be clear, consistent, and current. Missing or conflicting data can cause an assistant to describe the product incorrectly or leave it out of the selection altogether.

Service businesses may not have products or inventory, but they face a similar problem when public sources disagree about their services, expertise, or geographic reach. We will come back to that shortly.

The first numbers show a direction, but not a universal rule

Shopify observationFour signals, not a universal benchmark
+197%AI-attributed sessions year over year
≈ 3×corresponding orders
≈ 1/2of visits land on a product page
+80%conversion vs. organic traffic on those pages

Published by Shopify on August 11, 2026. The public report does not fully detail countries, categories, or cohort composition.

In August, Shopify published findings based on activity across its own platform. According to the company, AI-attributed sessions increased 197% year over year, while the corresponding number of orders nearly tripled. About half of those sessions landed directly on a product detail page.

Shopify also reported that AI-referred visits reaching a product page converted about 80% better than organic traffic. (Shopify, August 11, 2026)

These numbers are useful because they reflect commercial behavior rather than stated intentions in a survey. They still come from Shopify, which does not provide a complete breakdown of the sample, countries, and product categories in the public report.

We therefore treat them as a platform observation, not as proof that “AI traffic converts 80% better” for every brand. The early evidence suggests that some visitors arriving from assistants may be more qualified.

That makes Shopify’s analysis worth following. Over the next several months, we will be able to see whether the direction holds and how much it varies across industries, markets, and platforms. For now, the populations and conversion events being measured are still too different to produce a meaningful benchmark for every company.

Google can now show part of the journey, but not the business outcome

MeasurementFragments of the journey, not complete attribution
Search ConsoleImpressions by page, country, date, and device
Attribution gapNo direct link between appearance and outcome
GA4 and CRMObservable visits, leads, or sales

Traffic from AI Overviews and AI Mode remains classified as organic Google Search traffic.

On August 31, Google added a Search Console report covering AI Overviews and AI Mode, its two main forms of AI-generated search results.

Companies can now see which pages received impressions in those interfaces and review the data by country, date, or device. Previously, those appearances were mixed into Google’s broader search data. (Google Search Console)

The report answers one useful question: which pages appeared in generative search? It does not reveal the queries people entered, the conversions that may have followed, or the revenue involved. Features still running as experiments are also excluded.

Google Analytics 4 picks up another part of the journey once someone reaches the website. Since May, its “AI Assistants” channel has been able to separate some visits coming from ChatGPT, Claude, Gemini, or Perplexity. Visits from AI Overviews and AI Mode, however, remain classified as organic Google Search traffic. (Google Analytics, attribution rules)

The two tools therefore reveal different fragments of the same path. Search Console may show that a page appeared in an AI-generated result. GA4 may later show a change in visits, leads, or sales, but it cannot establish a direct connection between that appearance and the business outcome.

Measurement is moving forward, and that is already valuable. It still cannot provide a clean return-on-investment calculation for brand visibility in AI. Any dashboard displaying a highly precise figure today is either measuring only part of the phenomenon or relying on estimates that should be clearly explained.

When Google answers directly, users click less

U.S. experiment, March 2026Change in clicks to external websites

Percentage-point change versus the regular Google experience. The study does not isolate brand websites.

A preprint posted on August 18 provides a more recent view of this behavior. It reports on an experiment conducted in March 2026 with 1,100 participants in the United States.

For seven days, participants were assigned to one of three versions of Google Search: the current experience, a version in which AI Overviews and AI Mode were hidden, or a version that automatically redirected every search to AI Mode.

When the AI features were hidden, the click-through rate to external websites increased by 8.8 percentage points compared with the regular Google experience. When every search was redirected to AI Mode, the external click-through rate fell by 18.8 percentage points.

The authors describe these external destinations as “publishers,” but the category includes websites outside Google more broadly. It is not limited to news or informational media, and the study does not isolate traffic going specifically to brand websites. (preprint, August 18, 2026)

The experiment also found that forced use of AI Mode slightly reduced participants’ trust in information found on Google, along with their satisfaction and their sense of control over the search process. It did not measure a loss of trust in the brands mentioned in those results.

For brands, the practical lesson is straightforward: appearing in an AI-generated answer, even with a link, does not guarantee a website visit.

Two separate questions now need to be measured. The first concerns the answer itself: Is the brand present? Is it described accurately? Is it actually recommended? The second concerns what happens after a possible click: a visit, a lead, or a sale.

This study documents a decline in outbound clicks. It does not measure the commercial value of a citation or the influence the generated answer may have had on someone’s final decision.

ChatGPT Ads: A six-month head start in the U.S. and a more regulated launch in France

ChatGPT AdsOne channel, different personalization rules
United StatesFrance / EEA
Available sinceFebruary 9, 2026August 24, 2026
Eligible plansFree and GoFree and Go
Past chats, memory, and ad interactionsWith user permissionNot available at launch
Current conversationMay provide contextMay provide context
Organic answerSeparate from the adSeparate from the ad

Advertisers do not receive users’ conversations. Paid campaigns do not improve organic recommendations.

The United States has been testing advertising in ChatGPT since February 9, 2026. The initial pilot was limited to a small group of advertisers and primarily sold on an impression basis.

Within a few months, OpenAI added cost-per-click buying, a self-service campaign manager, pixel-based measurement, and a Conversions API. The platform later gained conversion optimization, geographic targeting, and custom audiences. (OpenAI, May 5, 2026)

When ChatGPT Ads reached France on August 24, French advertisers gained access to an infrastructure that was already more developed than the original U.S. pilot. Self-service access through Ads Manager became available on August 31.

Ads are shown only to users on the Free and Go plans and remain separate from ChatGPT’s answers. Purchasing an ad does not improve a brand’s position in organic recommendations. (OpenAI, August 18, 2026)

The two markets also differ in how ad personalization works. In the United States, users who enable personalized ads may allow ChatGPT to consider past conversations, saved memories, and previous interactions with ads. Personalization based on that history is not available at launch in France or elsewhere in the European Economic Area.

An ad may still be relevant to the subject of the current conversation. In all cases, OpenAI says advertisers do not receive access to users’ conversations. (OpenAI Help Center)

For French brands, this creates a new paid media channel with more capabilities than the original U.S. pilot offered in February. There is not yet enough evidence to know the size of its audience, its performance, or the role it may play in purchase decisions.

Whether a company sells products or services, assistants need a consistent story

Public consistencyTwo business models, the same requirement
ProductsNames, prices, features, variants, inventory, and markets
One clear version of the company
ServicesServices, expertise, coverage, evidence, profiles, and directories

Working hypothesis: reducing contradictions better prepares a brand, without guaranteeing inclusion in an answer.

Google states that no special markup or model-specific file is required to appear in AI Overviews or AI Mode. The fundamentals of search visibility still matter: content must be accessible, the structured data must match what people can see on the page, and the information itself must be clear. (Google Search Central)

For a retailer, consistency includes product names, prices, specifications, variants, and availability. Google recommends combining the information shown on the website with structured product data and a current Merchant Center feed. OpenAI also allows merchants to submit product information directly to improve freshness.

None of these systems guarantees that a product will appear in an organic answer. (Google, OpenAI)

A service business has a different set of information to maintain: its current services, the language used to describe them, the markets it serves, the expertise it claims, and the evidence supporting those claims. Those details should remain consistent across the company website, professional profiles, directories, partner pages, and outside publications.

Take our own case. J2A Digital now offers services focused on brand visibility in AI and no longer lists some of its earlier work, including events and print. If an old directory, professional profile, or partner page still carries that previous description, an assistant may continue to classify us according to what we used to do rather than what we offer today.

There is no inventory or product detail page in this example. The underlying problem is still the same: public sources are telling different versions of the company’s story. An assistant may describe an outdated offer, miss a newer area of expertise, or leave the company out of a recommendation it could otherwise have matched.

The inconsistency does not have to be dramatic to become a problem. An old service left in a directory or a new offer documented on only one recent page may be enough to blur the company’s identity. Visibility work increasingly involves finding those traces, correcting them where possible, and making the current offer easier to identify.

Our working hypothesis for this fall is that brands maintaining consistent information across their public sources should be better prepared for these interfaces. “Should” matters here. None of the platforms promises that accurate information will guarantee an appearance.

What this summer changes for brands

Several layers of the market moved forward at the same time. In the United States, assistants are moving closer to the transaction by comparing products, drawing on fresher catalog data, and in some cases preparing or executing part of a purchase. Conversational advertising is becoming a paid media channel and has now reached France. Measurement tools are also beginning to reveal generative search appearances, even though they cannot reconstruct the full path from an answer to a business result.

The range of information brands need to monitor has expanded. For a retailer, a conflict in price, inventory, or market availability can distort an answer. For a service business, the same problem may come from an outdated directory listing, a professional profile that was never revised, or a newer area of expertise that has not been documented clearly enough.

Brands can already take several practical steps:

  • Check what ChatGPT, Gemini, Claude, Perplexity, and Google’s generative search interfaces say about the company in every priority market and language.
  • Make sure important pages remain accessible to search engines and relevant AI crawlers, with key information clearly visible in the content.
  • Separate citations, brand mentions, accurate descriptions, and genuine recommendations. They do not represent the same kind of visibility.
  • Align information across the website and other public sources: products, pricing, and inventory for retailers; services, expertise, markets, and supporting evidence for service businesses.
  • Update professional profiles, directories, partner pages, and other sources that still describe an older version of the company.
  • Track generative impressions in Search Console and isolate AI referral visits where possible, while recognizing that part of the influence will remain invisible.
  • Test ChatGPT Ads as a separate paid channel, with its own budget and performance measures, without assigning it an effect on organic visibility.
  • Keep a record of the conditions behind every observation: platform, country, language, date, web access, and subscription level.

Summer 2026 did not produce a formula for controlling what AI systems say. It made a familiar kind of work more urgent: publishing clear information, maintaining reliable data, monitoring several platforms, and correcting the public traces that no longer describe the company accurately.

What about your brand?

J2A helps companies examine what assistants understand about them, align their public information, and build measurement that does not claim more than the evidence can support.

Discuss your AI Visibility
View sources and evidence status

Sources reviewed for this article