The hidden call journey: how LLMs and AI Overviews are reshaping the marketing funnel

For years, the marketing funnel started with a search bar. Today, it increasingly starts with a conversation. AI assistants and AI Overviews (AIO) – powered by Large Language Models (LLMs) – are answering your customers’ questions before they ever reach your website and, in a growing number of cases, before they ever run a traditional search at all.

For senior marketers, this raises an uncomfortable question: if someone finds you through an AI-generated answer and then picks up the phone, can you actually prove it?

What’s happening in the customer journey

The research journey has quietly split in two. Some customers still search, click, and browse in the way analytics platforms were built to measure. Others now ask an LLM a question, receive a synthesised answer, and act on it directly, sometimes without visiting a single website.

This matters most at the top of the funnel, where high-consideration decisions are made such as choosing a car or comparing care homes. These are exactly the moments customers used to spend on Google, working through multiple sites and comparing offers. Increasingly, an LLM does that comparison for them and hands over a shortlist, complete with contact details.

The result is a journey with some touchpoints not visible, more zero-click answers, and a phone call that can happen at almost any stage, and sometimes as the very first interaction with your brand. That call is now one of the clearest signals left of what’s working, and so analysing it is more important than ever.

Why it’s changing

Two forces are driving this shift.

Conversational AI assistants such as ChatGPT, Claude, and Gemini use Large Language Models (LLMs) within stand-alone chatbots. Customers use them like a knowledgeable assistant, asking follow-up questions and receiving direct, synthesised answers. Crucially, this usually happens entirely outside Google, on a separate platform with its own interface.

AI Overviews is an LLM-powered search experience integrated into Google Search. They are Google’s AI-generated summaries that sit at the top of standard search results. The customer never leaves the search engine, but the traditional list of blue links is pushed further down the page, beneath an answer Google has already written for them.

The similarity is what matters commercially: both AI assistants and AIO synthesise information from multiple sources, both cite those sources, and both can surface a phone number directly in their answer. The difference is where that happens and what data you’re left with afterwards.

In practice, this creates four distinct ways a phone call can originate from AI-generated search:

• A number is displayed in the results, pulled directly from your website. The LLM has read your site, found your phone number, and presented it to the customer without a click.
• The customer clicks a citation, lands on your website, and calls from there either during that visit or in a later visit.
• Your number is displayed because a third-party source was cited. For care providers, this is often the Care Quality Commission (CQC); for other sectors, it might be an industry directory or review site the AI model trusts.
• The customer sees your brand name within an LLM and then starts a new query using your name and then calls you.

Each of these behaves differently in your analytics, and each needs a different approach to measurement.

How to deal with it

Build authority, then optimise for it

Being cited starts with being trustworthy in the eyes of the model, not just the search engine. Source authority is now a core ranking factor for AI-generated answers: a citation from a recognised, trusted source (such as CQC for care providers) often carries more weight than your own website copy.

Alongside authority, a few practical steps help LLMs understand and cite your business correctly:

• Generative Engine Optimisation (GEO) practices: structuring content to directly answer likely questions, rather than optimising purely for keywords. Do this by:
  • Using short paragraphs, lists, and clear heading hierarchies with each section covering one distinct topic or question.
  • Leading with answers at the beginning of each section. Don’t bury the answer within lots of context. Instead, give AI systems a direct, extractable answer to use.
  • Adding authority signals such as including real data, quotes, and naming experts.
  • Building mentions across third-party discussion sites.
• Visibility and optimisation tools: Consider trialling optimisation tools such as Loved by AI, which can adjust content specifically for AI platforms. Also use visibility tools like Semrush to track mentions, research prompts, find content gaps, and audit your web pages for structural readability.
• An llms.txt file: Give AI crawlers a clear, structured summary of your site and offering. However, it’s worth noting these files aren’t actively used by major platforms; a study early this year by Otterly.AI showed that just 0.1% of AI bot traffic accessed llms.txt files. There’s a debate on whether it is worth doing. Any gains are likely to be marginal, and so it’s worth treating it as a secondary priority. The more crucial thing to do is check your robots.txt file isn’t blocking AI crawlers.

Rethink your top-of-funnel metrics

Traditional funnel metrics (clicks*, sessions, rankings etc) don’t capture AI-driven visibility. Marketers need to start tracking an additional layer of metrics to understand what’s happening before someone ever reaches your site.

Four measures to get on top of:

1. Citations: how often your website, or a source that mentions you, is referenced within LLM answers for your target queries. If it’s another source that mentions you, it’s a good way to identify who are the trusted sources.
2. Brand mentions: whether your business name is surfaced in the answer at all, even without a link or citation attached.
3. Visibility score: how often your brand shows up across the range of questions your customers are likely to ask, tracked over time.
4. Share of voice: how your citation and mention frequency compares with named competitors for the same set of queries. This is a strong proxy for who’s winning this new top of funnel.

These matter because if you can’t see whether you’re being cited, mentioned, or losing share of voice to a competitor, you have no way of connecting a change in call volume back to a cause, or knowing where to invest to improve it.

*It’s worth noting that GA4 have recently created a new default medium called AI Assistant medium e.g. chatgpt.com/ai-assistant, in place of using the referral medium.

Know what you can and can’t track directly

Where a customer clicks a citation from an AI assistant and lands on a page, Mediahawk uses Dynamic Number Insertion (DNI) technology to attribute the visit and any resulting call in the same way as any other digital source. This is the easiest scenario to measure.

What is difficult to track directly is:

• Where a customer clicks on a citation from an AI Overview. This is attributed to Google Organic, and so there is no current way of knowing it came from AIO.
• A call made from a number read straight out of an AI answer, with no click involved, or a call sourced from a third-party citation such as CQC. To be able to track these, investigate which phone numbers are being displayed across LLMs, and which pages they’re being pulled in from.

Build a feel for the indirect impact

Direct attribution has limits, but three approaches help build a reliable picture:

• Baseline and compare: Look at journey and call data from before LLMs were prevalent, then track how patterns have shifted since. Establishing a baseline lets you measure genuine uplift, rather than guessing at cause and effect. For example, if your business appears on a directory site and you were not previously receiving a high volume of calls from that number, but now you are. It shows that directory has a high level of authority and is heavily cited. It’s still the directory that has got you the positive volume of calls, but it’s being found using this new channel.
• Audit which numbers are getting calls: If calls to your untracked, static phone number start climbing without a corresponding change elsewhere, that’s a strong signal your business is being surfaced using that number and called directly from AI answers.
• Let your customers tell you: Use AI-driven Speech Analytics that can automatically identify when a caller references finding you through an AI tool or a specific third-party source, giving you qualitative evidence to sit alongside the quantitative uplift.

Connect it to revenue

The final step is linking this activity to sales. By capturing the call or email address and combining it with your sales data, Sales Matching together with User Journey allows you to follow the first click and/or call through to closed revenue and beyond. This helps you build a credible, data-backed picture of how AI-driven search is contributing to the bottom line, even without a clean, single-touch attribution path.

Where to start

The customer journey has changed, and it will keep changing as LLMs and AI Overviews evolve. The businesses that adapt fastest will be the ones who stop expecting a single, clean attribution line and start building a broader evidence base instead – with baselines, citations, phone trends, and call insights, all working together.

The harder the research stage becomes to see, the more value there is in capturing the moments where it resurfaces. A conversation with an LLM might leave little trace in your analytics, but the website visitor or phone call that follows it does. That moment, where invisible research turns into a real enquiry, is now one of the few reliable signals left of what’s actually working. Measuring it well and linking it back to the marketing that made it possible matters more now than it did when the customer journey was easier to see in full. Getting it right isn’t optional anymore; it’s the difference between guessing and knowing.