❌

Vue lecture

The AI visibility gap: What happens before the website visit?

Someone asks ChatGPT to compare web analytics platforms. They read the answer, close the chat and come back a day later by searching for one of the vendors by name. After reading a few pages, they book a demo. 

Web analytics can capture the organic search visit, the pages viewed and the conversion. What it can’t show is whether an AI assistant influenced the decision before the visitor ever reached the site. That missing context creates an AI visibility gap: part of AI’s influence never becomes visible in the analytics data. 

Here, visibility refers to measurement visibility, not how often a brand appears in AI-generated answers. 

In Matomo’s Future of Web Analytics report, based on a survey of 300 web analytics experts in the United States, Germany and France, 92.7% say traffic from AI platforms such as ChatGPT and Perplexity is important to their business. Another 85.7% say AI-generated traffic has a strong or moderate impact on analytics data. 

Visibility is uneven. 53.7% report clear visibility into AI traffic, while 40.7% say their visibility is partial. AI is also changing how teams approach measurement: 42.7% say it is forcing them to rethink how they measure performance, and 41% say it is increasing uncertainty in their data. 

What the AI visibility gap means for analytics 

AI search tracking covers identifiable visits from AI assistants. AI traffic attribution looks at how those recorded AI touchpoints contributed to a conversion. 

The AI visibility gap begins where that evidence stops. If someone researches a product in ChatGPT and later arrives through organic search or Direct, analytics can measure the visit and the conversion. It can’t connect the earlier AI interaction to that person unless another measurable signal links the two. 

That distinction affects how channel performance should be interpreted. An organic search visit is still an organic search visit, and Direct traffic shouldn’t be reclassified as AI simply because AI may have influenced an earlier stage of the journey. 

Other sources can add context. CRM data can show whether a lead became revenue. Branded search trends can show changes in demand. Customer surveys can capture how people remember discovering a brand. AI citation monitoring tools can show where a brand appears in AI-generated answers. Each source contributes a different piece of evidence, and none should be treated as proof of an interaction it did not actually record. 

Use each source for what it actually records, and don’t treat supporting data as proof of an AI interaction that wasn’t measured. 

AI search tracking captures the visible part 

When someone clicks a source link from an AI assistant and referrer information is available, analytics can identify the visit. Teams can then compare AI-referred visits with other acquisition channels and analyse landing pages, visit behaviour, goal conversions and other on-site outcomes. 

An AI-influenced conversion may still belong to another channel 

Traffic source and AI influence aren’t always the same thing. Consider three common journeys: 

  • A person compares vendors in ChatGPT, later searches for one brand and converts through organic search. 
  • A person sees a recommendation in Perplexity, returns later by typing the domain and converts during a Direct visit. 
  • A person discovers a product through an AI assistant, then clicks a paid campaign while continuing their research and eventually converts. 

In each case, the visible acquisition channel describes the visit that actually reached the website. Reclassifying organic search, Direct or paid traffic as AI without evidence would create false precision and make the channel report less reliable. 

The same limit applies to attribution models. Multi-touch attribution can distribute credit across touchpoints that were recorded, but it can’t create a touchpoint that was never observed. 

More AI referrals improve visibility into measurable visits. They still don’t explain AI influence that happened off-site, so teams need to separate recorded evidence from supporting context and inference. 

What the data can and can’t tell you 

A practical way to work with the gap is to look at each journey in layers. Start with what web analytics actually recorded, then add other evidence without presenting it as direct attribution. 

JourneyWhat web analytics can see What remains unknown Other context 
AI referral, then conversion AI referrer, landing page, on-site behaviour and conversion. Any earlier research or AI exposure before the click. CRM data can add pipeline or revenue after the conversion. 
AI research, then branded search and conversion Organic visit, on-site behaviour and conversion. Whether AI caused or influenced the organic search. Self-reported discovery, search demand, CRM and AI citation monitoring can add context. 
AI answer, no website visit No on-site signal from that interaction. The whole AI interaction and whether it influenced a later decision. AI citation monitoring can show that the brand or page appeared in AI-generated answers, while brand-search trends can add further context. Neither provides direct attribution. 
AI referral, then lead and closed deal in a multi-step enterprise sales process AI referral and the on-site lead conversion. The commercial outcome after the lead leaves the website. CRM opportunity and revenue data connect the measurable visit to downstream value. 

The final column helps you interpret the signals you already have. You probably won’t be able to reconstruct every journey end to end. Comparing data across systems can still show whether different parts of the journey support the same interpretation, without turning that context into a measured touchpoint. Self-reported discovery data can help too, with the limitation that it reflects what people remember and choose to report. 

AI attribution needs context from more than one system 

Different systems record different parts of the customer journey. Web analytics shows what happened on the website, while CRM data shows what happened to a lead or account afterwards. Search and campaign tools can show changes in demand and acquisition. AI citation monitoring can show whether, for example, a product page appears in ChatGPT or Perplexity responses. On its own, though, it doesn’t tell you whether those appearances led to a website visit or conversion. Web analytics provides the separate on-site view of identifiable visits and conversions. BI or a data warehouse can then bring several of those signals together under shared definitions. 

The Future of Web Analytics report found that 49% of respondents expect hybrid approaches combining multiple systems to define the future of web analytics. Another 38% see foundational data layers connected to BI tools as part of that future, while 41% select integration with existing tools as a buying factor. 

To connect AI-referred visits with pipeline or revenue, teams need a way to use analytics data alongside CRM, BI or other business data. Analysts also need ways to retrieve the data without being confined to one dashboard. 

APIs and structured exports are established routes for that work. As AI becomes part of analysis itself, protocols such as MCP provide another option for compatible AI applications to retrieve analytics context. Matomo’s MCP Server, for example, can connect authorised tools such as ChatGPT, Claude and Codex to Matomo data. If the AI application is also connected to other systems, it can use Matomo data alongside that context without teams having to assemble everything manually first. 

What to look for in an analytics setup 

When assessing an analytics setup for AI attribution, look at whether it can support the following: 

What to check Why this matters 
Can you identify human visits referred by recognised AI assistants and analyse their conversions separately? This lets you compare AI-referred visits with other acquisition sources and see how those visitors behave and convert once they reach the website. 
Can you separate automated AI agent or chatbot activity from human behaviour where the available signals allow it? Automated activity can affect traffic and engagement figures. Keeping it separate makes human behaviour easier to interpret and allows AI optimisation as well. 
Can analysts retrieve data through APIs or structured exports? APIs and structured exports let teams use analytics data alongside CRM, BI or other datasets to analyse performance beyond the dashboard. 
Can ecommerce teams compare AI-referred visitors with other human visitors across the on-site conversion funnel, while keeping automated AI activity separate? This shows whether people arriving from AI tools behave differently across product, cart and checkout steps, without automated requests distorting conversion rates or drop-off analysis. 
Can analytics data move into CRM, BI, warehouse and reporting workflows? Conversion data becomes more useful when teams can connect on-site behaviour with leads, pipeline, revenue and other business outcomes. 
Can authorised AI tools work with analytics data under clear access controls? This keeps AI-assisted analysis within the access boundaries set for the underlying analytics data. 
Can teams assess the quality of the underlying data, and is it unsampled and auditable? AI tools work from the data they receive. Teams need to understand its limitations and trace how the numbers were collected and reported.  
Can the team trace how sources are classified and how attribution logic is applied? When a channel suddenly gains or loses conversions, analysts need to understand whether the change came from visitor behaviour, source classification or reporting logic. 

These checks show whether teams can trace AI-referred visits, understand how they were classified and connect them with downstream business outcomes. 

How Matomo helps with the measurable part 

Matomo separates human visits referred by AI assistants from automated AI activity, so teams can analyse each on its own terms. The AI Assistant acquisition channel identifies human visits from recognised AI assistants when referrer information is available. Those visits can be compared with other acquisition channels and analysed across landing pages, behaviour, goal conversions and other on-site outcomes. 

Dedicated AI reports cover activity that doesn’t represent a human referral. The AI Agent Overview compares recognised agent visits with human traffic, while AI Chatbots reports show how supported chatbots retrieve website content through server-side tracking. Content Requests reports can also show which pages and documents chatbots request, including content they favour or fail to retrieve. AI Chatbot requests themselves are processed separately from visits, so they don’t create visit or attribution data. 

Teams can also work with Matomo data through AI-assisted workflows. The Matomo MCP Server gives authorised tools such as ChatGPT, Claude and Codex structured access to Matomo analytics data, so analysts can query it in natural language. If the AI application is also connected to other systems, it can use Matomo data alongside that context. 

Ask Matomo, scheduled for the second half of October 2026, will bring that type of natural-language analysis into Matomo itself. Signed-in users will be able to ask questions about supported reports, analytics data and configuration, subject to their existing permissions. 

Build AI attribution around the evidence you have 

AI can influence research before a visitor reaches your website, but attribution still needs to follow the touchpoints you actually record. 

If a recognised AI referral is recorded, you can analyse the visit and include that touchpoint in attribution. If someone arrives through organic search and there’s no measurable AI touchpoint, keep the visit under organic search. Customer surveys and AI citation monitoring can add context about earlier discovery, while CRM data can connect the recorded visit or conversion with downstream outcomes. None of those sources turns an unrecorded AI interaction into a measured touchpoint. 

Use directly measured interactions for attribution and bring in other data where it adds useful context. When a connection is inferred rather than recorded, make that clear. 

Explore the wider findings on AI, trust and connected analytics in The Future of Web Analytics. 

See how Matomo handles AI traffic in practice 

Separate AI referrals, agents and chatbots, then connect analytics data to the wider workflows your team uses. 

Start your free Matomo trial  |  Book a demo 

The post The AI visibility gap: What happens before the website visit? appeared first on Analytics Platform - Matomo.

  •  

EU: Commission proposes a new KIDS Act to protect children online

On 17 September 2026, the European Commission unveiled its proposal for the EU KIDS Act (EU Keeping Internet Digital Spaces Accountable and Trustworthy), a harmonised EU framework designed to reshape how children engage with online services. The proposal goes well beyond age verification requirements and seeks to fundamentally change the design of digital services used by minors, placing greater responsibility on technology providers to create age-appropriate digital environments. It builds on existing legislation such as the EU Digital Services Act (DSA) and the EU AI Act but introduces more specific rules focused on child safety.

Which Services Are Covered?

The KIDS Act imposes extensive obligations on a broad range of online services. Its scope extends beyond social media platforms to encompass a wide range of online services, including:

  • online social networking services;
  • video-sharing platform services;
  • app stores;
  • online games;
  • operating systems;
  • AI companions; and
  • general conversational chatbots.

Summary of key requirements

The proposal is built around four core themes:

  1. EU-wide minimum age for social media accounts

A key feature of the proposal is the establishment of common age-based rules governing children’s access to social media and certain video-sharing platforms across the EU. The Commission’s objective is to introduce a gradual and age-appropriate pathway into the online environment, reflecting growing concerns about the impact that social media and algorithm-driven content can have on children’s wellbeing and development. The Commission emphasises the need for a harmonised framework to avoid a fragmented landscape of national age thresholds and access requirements.

Under the proposal:

  • Children under 13 will be prohibited from accessing social media services. They would, however, be able to access specially designed child-friendly video-sharing services through an account managed by a parent or guardian.
  • Children aged 13 and 14 will only be able access social media and certain video-sharing services through parent-managed “mini accounts”. These accounts would offer limited functionality, enhanced parental controls and restrictions on use, including a daily screen-time limit.
  • From age 15, minors will be permitted to open and manage their own social media accounts and access covered video-sharing services independently.

In addition, under the proposal, providers would be required to determine whether existing account holders are under the age of 15. Accounts belonging to users identified as under 15, or whose age cannot be verified, would need to be disabled. Providers designated as Very Large Online Platforms (VLOPs) under the DSA would face additional obligations, including submitting a compliance plan demonstrating, through robust technical and evidentiary measures, a “high degree of confidence” that users have reached the age of 15.

  1. Safety-by-design obligations

While the age-based access rules have attracted the greatest public attention, the KIDS Act’s broader ambition is to address the design choices that can encourage excessive use, amplify harmful content, or expose minors to inappropriate interactions. In doing so, it reflects a growing regulatory view that children’s online safety cannot be achieved through parental controls and age verification alone, but must be embedded into the architecture of the services themselves.

To achieve this, the proposed KIDS Act requires a wide range of online services used by minors, including social media platforms, video-sharing platforms, online games, AI companions and chatbots, to implement extensive safety-by-design measures. These include:

  • Restrictions on addictive features such as infinite scrolling, autoplay functionality, reward mechanisms and engagement-driven notifications, particularly during sleeping hours.
  • Safer recommender systems that provide minors with greater control over how content is suggested and reduce the risk of being drawn into harmful “rabbit holes” through profiling-based recommendation algorithms.
  • Privacy-protective default settings, including private accounts by default and limits on the collection and use of children’s personal data.
  • Strong protections against unwanted contact from strangers, cyberbullying and other forms of harmful interactions.
  • Child-friendly reporting, blocking and support tools that are easy for young users to access and understand.

The proposal also introduces specific safeguards for AI companions and chatbots. This includes requiring AI companions to be disabled by default for minors and avoid design features and system behaviours that simulate interpersonal relations that are likely to create emotional dependencies, manipulate behaviour or otherwise exploit children’s vulnerabilities.

  1. Age verification and parental controls

The proposal introduces specific age assurance obligations requiring providers to establish whether users are above or below relevant age thresholds and to ensure that children are provided with experiences, features and protections that are appropriate for their age. The proposal also introduces a framework for verifying parental responsibility where an adult seeks to create or manage an account on a child’s behalf.

The proposal acknowledges that effective age assurance must be balanced with users’ privacy and data protection rights. As a result, it emphasises that age verification and age estimation mechanisms should be designed in a privacy-preserving manner, minimising the collection and retention of personal data and avoiding the introduction of de facto identity verification for access to online services. The proposal requires age to be verified through certified solutions independent of the platforms, including a free EU age verification app. The proposal requires Member States to offer at least one free way to prove age, including for people without digital ID.

To strengthen the effectiveness of these safeguards, social media services and video-sharing platforms are required to verify users’ ages at the point of account creation.

Parents and guardians will also be given enhanced control tools, including screen-time management, contact supervision and account-setting controls.

The proposal goes beyond the GDPR’s specific rules on children’s consent. Its age assurance and age-appropriate design obligations are not limited to processing based on consent but may also affect services that rely on other legal bases, such as contractual necessity or legitimate interests.

  1. Compliance, Oversight and Penalties

    Primary responsibility for supervising the largest services would rest with the European Commission, building on the model established under the DSA. In-scope providers could face substantial penalties for non-compliance, including fines of up to 6% of their worldwide annual turnover and, in the most serious cases, temporary restrictions on the provision of services within the EU.

    The proposal also introduces proactive compliance obligations, particularly for VLOPs. The KIDS Act shifts the burden from regulators to providers, requiring VLOPs to proactively demonstrate, through a Commission-facing compliance plan and independent audits at their own expense, that they can meet obligations before their services are made available to children. The KIDS Act also introduces a supervisory fee, capped at 0.03% of a provider’s worldwide annual net income, to fund oversight and enforcement activities. The fee would apply not only to in-scope VLOPs, but also to certain providers of AI companions, general conversational chatbots and video gaming platforms.

    Timing

    The EU KIDS Act must now pass through the EU legislative process before it can be adopted. Several aspects of the proposal are likely to attract significant debate, particularly the age assurance requirements and the proposed age threshold of 15. While the European Commission has called for the legislation to be prioritised, adoption is unlikely to be immediate. Once adopted, most of the Act’s obligations would begin to apply six months after coming into force, giving providers a relatively short window to implement what are likely to be substantial compliance changes.

    What does this mean for organisations?

    Although the proposal is primarily aimed at online platforms and digital services used by children, its implications are likely to be far-reaching. Organisations operating consumer-facing digital services will need to assess whether their services fall within the scope of the Act and, if so, whether existing product designs, recommendation systems, engagement features and age assurance mechanisms are compliant.

    The proposal also reinforces a growing regulatory expectation that children’s safety should be embedded into product development from the outset. Businesses may therefore need to revisit governance structures, risk assessments, design processes and compliance programmes to ensure that child safety considerations are integrated across the product lifecycle.

    •  
    ❌