The Death of Analytics: How AI is Breaking the Buyer Journey We Thought We Could Measure

Marketing attribution is the practice of assigning credit for a conversion or sale to the marketing touchpoints a buyer was observed interacting with. The key word is observed. Attribution can only distribute credit across what your analytics, CRM, and ad platforms recorded, and it cannot prove that any one touchpoint caused the purchase.

Every analytics platform I have ever implemented was built on one simple premise: a person finds you, a person visits you, and that visit is captured as an event. Search, click, land, convert. Count it, chart it, credit it.

That premise is dying. I am not sure it will ever recover.

To be fair, it was never entirely true. Dark social, consent banners, ad blockers, cross-device behavior, phone calls, word of mouth, and offline conversations have always punched holes in our data. We learned to live with those blind spots because the core of the journey, the search and the click, still happened where we could see it. AI assistants have moved that core somewhere else. The research, the comparison, and often the decision now happen inside a conversation we will never see.

Let me be precise about what is dying. Measurement itself is not useless. What is dying is our confidence that we can reconstruct how a buyer discovered us, what influenced them, and when they decided. That confidence underpinned how most marketing teams justified budgets.

A Buyer Journey Your Analytics Will Never See

Consider a hypothetical example. It is an illustration, not a measured case, and not every assistant works this way.

A marketing director at a mid-sized software company opens an AI assistant and types a long, specific question: which customer data platforms integrate with her existing CRM, support her industry’s privacy requirements, and can be implemented by a small team this quarter. Only the assistant sees that prompt.

The assistant may start answering based on what its large language model (LLM) learned during training. If your brand was well represented in that training data, you may already be on her shortlist. If it was not, you may already be off it. Either way, no request reached your site.

If the assistant decides to search, it may break her question into a series of machine-written queries covering integrations, compliance, implementation time, reviews, and alternatives. It reads the results, may fetch a handful of pages, and extracts passages. It writes an answer. She asks a follow-up about one vendor’s onboarding, and another round of generated searches and fetches begins. She asks for a comparison table. Another round. In a long session, a single buyer could plausibly trigger dozens of generated searches across several rounds of conversation.

Somewhere in that conversation she decides. Maybe two vendors make the cut. Three days later she types one vendor’s name into a browser, lands on its homepage, and requests a demo.

Now answer the questions every marketing report is supposed to answer. How many keywords were queried? You will never know, because you have no access to the queries the assistant generated on her behalf. How many visits occurred? One human session, plus whatever automated fetches hit your server and never ran your tags. Was she engaged? Analytics defines an engaged visitor as someone who spends time, scrolls, or interacts with your site. In this hypothetical, she may have spent many minutes reading, comparing, and thinking about your content inside the assistant, while the fetch that supplied it lasted a split second on your server. None of that attention registers as time on page, scroll depth, or an interaction. At what point was the decision made? Inside the assistant, days before your analytics recorded anything. Your dashboard will report this deal as a branded search or a direct visit that converted on its first session. That is the madness we are living in.

How AI Assistants Fan Out a Single Question

In its Search Central documentation, AI Features and Your Website, Google says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response. Google has not published how many searches a given question produces, and it notes the two features may use different models and techniques.

Other assistants behave differently, and the same assistant can behave differently from one question to the next. Some answers come entirely from learned knowledge with no web access. Some use a search index. Some fetch live pages when a user asks. That variability is part of the problem: there is no single pipeline to instrument.

Separate five things marketers routinely lump together:

• The user prompt: what the human actually asked. You never receive it.
• Generated queries: the searches the system writes on the human’s behalf. They are not your keywords, and they are not reported to you as individual queries tied to a buyer.
• Retrieval, fetch, or crawl: an automated request for your content. OpenAI, for example, documents separate agents for search crawling (OAI-SearchBot) and for pages fetched during a user’s request (ChatGPT-User) in its Overview of OpenAI Crawlers. These requests can appear in server logs, but they typically do not execute your JavaScript analytics tags.
• Agent tool calls: an agent invoking a function your site or service exposes, rather than reading a page.
• The human session: a person in a browser, on your site, running your tags. This is the activity browser-based analytics tags are best positioned to measure.

Do not treat the first four as the fifth. A generated query is not a human visit. A fetch is not necessarily a human session, even if an agent-driven browser happens to execute your tags. A tool call is not a page view. And server logs are not the same as JavaScript analytics: one records requests to your server, the other records behavior in a browser that agreed to run your scripts. A raw request that arrives, grabs a page, and leaves can also look a lot like a security scanner probing for holes. Verified crawler identities, published IP ranges, user agents, and request patterns help separate the two, but intent, benign or malicious, cannot be reliably judged from a single request, and a bot’s brief visit is never a measure of human engagement.

Where WebMCP Fits, and Where It Does Not

WebMCP adds another layer worth understanding accurately. It is a proposed browser API that lets a web page expose structured tools that an AI agent can call, such as searching inventory or submitting a support request, instead of having the agent click through the page like a person. The specification is a Draft Community Group Report from the W3C Web Machine Learning Community Group, not a W3C standard. The Chrome team announced an early preview in February 2026 (WebMCP Early Preview Program, Chrome for Developers), and the WebMCP project’s implementation status page lists origin trials in Chrome and Edge.

Two distinctions matter for measurement. First, Chrome’s WebMCP guide explains that tools run client-side in a browser tab, so they generally require a page from your site to be open, which is different from a remote MCP server your company hosts, and different again from a crawler or a search index. Second, availability depends on whether a site implements tools at all and whether the browser and agent support them. Support and implementation vary widely. Where it is used, an agent completing a task through a tool may never produce the scroll depth, clicks, and form events your analytics expects, unless you deliberately log those tool calls yourself. Some tool executions may still fire your browser tags, but what they record is an agent completing a task, not a person engaging with your content.

Referrals Help a Little, and Mislead a Lot

Some of the late click is measurable. OpenAI’s Publishers and Developers FAQ states that ChatGPT adds utm_source=chatgpt.com to referral URLs from its search results, so those sessions can be segmented in analytics platforms like GA4. Google states in its Search Central documentation that traffic from AI Overviews and AI Mode is included in the Web search type of the Search Console Performance report, not broken out on its own.

Those signals describe the last step, not the journey. Links copied out of an app, opened in another browser, or revisited days later frequently arrive without a referrer or UTM at all and land in direct. The buyer who decided inside an assistant and then searched your brand name gets credited to branded search or direct, as if that session were the beginning of the relationship rather than the end. Last-click models will happily hand all the credit to the one moment you could see.

Why I Do Not Expect Analytics to Rebound

For years, vendors responded to measurement losses with partial workarounds. Cookies broke, so we built first-party tracking. Referrers vanished, so we leaned on UTMs. None restored a complete journey. AI makes that incompleteness structural. The activity that matters is happening on platforms that have no obligation to report it to you, and in a form- a private conversation- that should not be reported to you. No tag, server-side container, or AI visibility tool can recover a prompt you never received or a decision that happened in someone else’s interface. Vendors who promise to reconstruct the hidden journey are selling a model of it, not a recording.

So the goal is no longer to see everything. The goal is to measure what you can observe honestly, connect it to revenue, and stop pretending the dashboard is the journey. Search engine optimization (SEO) still matters; how we prove it matters has to change.

Takeaways: How to Minimize the Impact

• Treat attribution as probabilistic, not factual. Attribution allocates credit across observed touchpoints. It is not proof of causation. Present it with ranges and caveats, compare multiple models, and stop treating last-click branded search or direct as the source of demand.
• Preserve what you can observe. Capture UTMs, referrers, landing pages, and form source fields on conversions where they are available and consented, and pass them into the CRM touch history. Keep assistant referrals in their own channel grouping. Do not try to reconstruct queries you never saw.
• Ask buyers how they found you. Add a free-text self-reported attribution field to demo and contact forms, and have sales confirm it on the first call. It is imperfect, but it can surface an influence absent from your clickstream, like an AI assistant recommending you.
• Define engaged users consistently. Pick a documented definition, such as return visits, pricing or documentation views, or product actions, and keep it stable. It remains useful for observable human sessions, but it does not measure total brand or content engagement: a brief visit, or none at all, is not a lack of interest when the reading happened inside an assistant. Pair it with self-reported attribution and qualified pipeline and revenue in your CRM, and use it as a quality indicator, not as proof that a channel caused a sale.
• Qualify leads with discipline. A marketing qualified lead (MQL) should reflect fit with your ideal customer profile (ICP) and real intent, and a sales qualified lead (SQL) should require sales acceptance and a confirmed opportunity, not just a handoff. Never make MQL volume the north star; it is the easiest number to inflate.
• Monitor sales over visits. Prioritize closed-won revenue, pipeline value and quality, stage conversion rates with their denominators, sales cycle length, and retention and repeat purchase. Watch these trends alongside traffic, never instead of them.
• Join your systems. Connect site conversion events to CRM records using lawful, consented first-party identity, such as a form submission or login, so you can trace opportunities and revenue back to observed touchpoints.
• Measure outcome cohorts and incrementality. Compare cohorts by first observed source and by self-reported source, and run holdout or geographic tests where feasible. Account for sales-cycle lag before declaring any channel a success or failure.
• Sample AI visibility directionally. Periodically run the questions your buyers ask in the major assistants and record whether and how you are cited. Treat it as a directional sample, not complete coverage, since answers vary by user, phrasing, and time.
• Separate bots from humans. Filter automated agents out of human analytics, and use server logs as a complementary view of which pages crawlers and user-triggered fetchers request. Never add the two together as visits.
• Reset expectations with leadership. Lower traffic is not, by itself, a decline; it may mean buyers arrive later and better informed. Higher traffic is not, by itself, success. Report the business outcomes and be candid about what the data can no longer show.

Analytics as we knew it, the confident reconstruction of every step from first query to purchase, is gone. Companies that adapt will stop chasing the journey they cannot see and start measuring the revenue they can.