How to Measure What AI Is Actually Doing to Your Pipeline
A working guide for marketing leaders who cannot find AI anywhere in their analytics
Two marketing leaders told me the same thing last week. They cannot tell if AEO is working.
AEO is Answer Engine Optimization. It is the work of getting your brand to show up inside AI answers like ChatGPT, Perplexity, and Google’s AI Overviews.
Both had been at it for months. Both opened their analytics, saw close to nothing, and started building the case to cut the budget. They were reading a real number. It was the wrong one.
This guide explains why that number is wrong, what your number probably looks like, and how to measure it yourself without buying anything. The audit at the end takes about an hour the first time and fifteen minutes every month after that.
Part 1. Why you cannot see it in your reports
Before anything else, understand that your analytics is not lying to you on purpose. It is missing AI traffic for a specific mechanical reason, and once you know the reason, every number downstream starts making sense.
Most of the AI traffic you get arrives unlabeled.
When someone clicks a link inside an AI app on their phone, the app does not hand them off to Safari or Chrome. It opens the page in a small built-in browser window inside the app itself. That window usually does not pass along where the visit came from. Your analytics sees a person arriving with no history at all, so it files the session under Direct.
Controlled testing on one major assistant found that only about 9% of visits from its mobile app were identified correctly. The rest landed in Direct.
There are four common ways this happens.
1. In-app browsers. This is the big one. Every major AI assistant has a mobile app, and most of them open cited links inside the app rather than handing off to a real browser.
2. Copy and paste. Someone reads an answer, copies the link, and pastes it into a new tab. There is no referrer to pass, because there was no click. This is very common in AI search right now, because people want to keep the answer open while they check the source.
3. Redirect chains. If a link passes through a redirect, especially one that touches an insecure page along the way, the referrer can be dropped in transit.
4. Privacy settings. Some platforms and some browsers strip the referrer deliberately.
Here is the part that explains most of the variation between companies.
AI used in a browser is mostly trackable. AI used in an app is close to invisible.
Which means your reported AI number is not really measuring AI at all. It is measuring how your buyers happen to hold their phones. Two companies with identical AI visibility will report wildly different numbers if one has a desktop audience and the other has a mobile audience.
There is a second gap worth knowing about. Google does not separate AI Overview and AI Mode clicks from regular organic search. Those visits arrive tagged as google / organic, mixed in with everything else. So even the trackable portion is incomplete.
Part 2. The number everyone is quoting is not a stat
You are going to hear a big number about AI traffic this year. Someone will put it on a slide. Before you plan against it, understand where it came from and why it cannot be applied to your business.
You have probably seen the claim that AI now drives 30% to 60% of traffic. It shows up in decks, on stage, and in vendor pitches. It gets repeated because it sounds urgent.
It is not a stat. It is guessing.
Most of those figures are modeled estimates. Someone looked at unexplained Direct traffic, made an assumption about how much of it came from AI, and multiplied. Often the company doing the modeling also sells the tool that measures it. That does not make the estimate useless, but it does mean you should not take it into a budget meeting as fact.
The deeper problem is that there is nothing to pin down. The number changes by vertical. It changes inside a vertical. It changes for the same company depending on where the buyer sits in the cycle.
Anyone who hands you one number for the entire market is selling you something.
Part 3. Two numbers, both true, measuring different things
This is the confusion at the center of every bad AI conversation happening in marketing right now. Two numbers get quoted as if they contradict each other. They do not. They count different things.
The first is AI referral traffic. That is the share of sessions your analytics can trace back to an AI tool. Across enterprise sites it sits around 1.08%. In information technology it reaches 2.80%. Those are small numbers and they are accurate.
The second is AI influence. That is the share of buyers who used AI somewhere in the process. In B2B that runs around 60%, and inside an active procurement cycle it climbs much higher.
One percent and sixty percent. Both are correct. They are not in conflict, because they are not measuring the same thing.
Referral traffic measures the last click. Influence measures the journey. Your analytics only sees the first one.
When a marketing leader looks at a dashboard and sees 1%, they are not seeing a failed channel. They are seeing the last step of a process that mostly happened somewhere else.
Keep these two words separate in every conversation you have about this. When someone says AI is 60% of the market, ask whether they mean traffic or behavior. Almost always they mean behavior, and almost always the slide says traffic.
Part 4. There is no one number. There is your number.
Your exposure depends on what you sell. The spread between categories is wide enough that an industry average will mislead you in one direction or the other. Find yourself in the list below.
B2B sits at the high end. Around 60% of B2B buyers report using AI tools somewhere in the purchase journey. In B2B software, 71% say they rely on AI chatbots for research and 51% say they start there. Deep inside an active procurement cycle it matters more. Recent research puts AI’s influence on vendor shortlists at 92% and its influence on final decisions at 83%.
Read those two numbers again. If a buyer is building a shortlist, there is a very good chance an AI helped build it. You were either in that answer or you were not, and you will never see the moment it happened.
Automotive is the one most people get wrong, myself included. It is one of the most researched purchases anyone makes, so you would expect it near the top. Consumers are actually less inclined to use AI across the car-buying journey than they are in other categories. That does not mean AI is absent. It shows up first in the research-heavy parts, which are vehicle research, pricing, valuation, and payment calculations. The test drive and the dealership visit are still doing work that an answer cannot do.
Ecommerce runs higher than the referral data suggests. Nearly 60% of consumers say they have used AI to help them shop. Be careful with that figure. It is 60% of consumers reporting AI-assisted shopping behavior. It is not 60% of ecommerce traffic. Those get mixed up constantly, usually by someone quoting it at you in a meeting.
Luxury resists a single figure. Two-thirds of luxury consumers say they use AI features when shopping for fashion online. That sounds high until you look at what they are using it for. Checking fit, visualizing a product, and narrowing a set are all real uses. None of them is the same as asking an AI which handbag expresses who you are. The tool is present. The decision is not being made there.
Clothing splits inside itself. Shoes give buyers something to ask about. Size, width, arch support, surface, mileage, durability. Pants still have questions, including rise and inseam and fabric, but habit and feel and a brand you already trust usually matter more.
Same customer. Same cart. Two different behaviors.
Part 5. The rule underneath all of it
If your category is not in the list above, you can still work out your exposure. There is one rule that explains every number in this guide, and you can apply it to any business in about two minutes.
Once you see the shoes and pants split, the pattern behind every other category becomes clear.
AI use rises when the decision can be written down.
Shoes have specifications, so people ask, because there is something to ask. Pants are mostly fit and habit. You know your size and you know the brand that fits you, so there is less reason to type the question.
The same rule runs everywhere. B2B procurement has a great deal that can be specified, including requirements, integrations, security, pricing, implementation, and trade-offs. That makes AI useful early and often. Luxury holds more feeling, identity, and physical experience in the decision, so AI supports the purchase without driving it. Everything else lands in between based on how much of the choice a buyer can put into words.
That is your diagnostic and it costs nothing to run. Ask how much of your buyer’s decision could be typed into a box. That is roughly where AI can enter the journey.
Run it properly and write the answers down. List the ten questions a buyer asks before choosing you. Mark each one as answerable in text or not answerable in text. If seven of the ten have written answers, assume AI is already sitting in the middle of your funnel whether you can see it or not. If three of ten do, your exposure is real but limited, and it lives early in the journey rather than at the decision.
The numbers, side by side
| Category | Figure | What It Measures |
|---|---|---|
| B2B buyers using AI in buying journey | 60% | AI use somewhere in the purchase process |
| B2B software buyers relying on AI for research | 71% | AI-chatbot use for software research |
| B2B software buyers starting research in AI | 51% | Research begins in an AI chatbot |
| B2B buyers reporting AI influence on vendor shortlist | 92% | AI-influenced vendor consideration |
| B2B buyers reporting AI influence on final decision | 83% | AI-influenced final purchase decision |
| Consumers using AI to shop | ~60% | Self-reported AI-assisted shopping |
| Luxury-fashion shoppers using AI features | ~67% | AI use while shopping online for fashion |
| AI referrals across enterprise sites | 1.08% | Share of all measured website sessions from AI referrals |
| Information Technology AI referrals | 2.80% | Share of IT-site sessions from AI referrals |
| Black Friday AI-driven retail traffic growth | 805% YoY | Growth in AI-driven traffic, not traffic share |
Read that table from the bottom up and the story gets easier to see. The measured numbers are tiny. The reported behavior is enormous. The gap between them is not an error. It is the part of the journey your tools were never built to watch.
60% of B2B buyers use AI somewhere in the journey. 92% say it influenced their vendor shortlist. And 1.08% of measured sessions come from AI referrals. All three are true at the same time.
Part 6. What to report instead
You cannot report a number you cannot measure. So stop trying. Replace the AI traffic line in your deck with three signals you already own, all of which point at the same thing from different angles.
You are not going to buy an enterprise tracking platform and you do not need one. You have first-party data and it is enough.
Start with Direct traffic sorted by landing page. Real direct visitors go to your homepage because they typed an address they remember. Anything landing deep on a product page or a blog post did not come from memory.
Put branded search next to it in Search Console. People who get recommended by an AI go and look you up. When both lines move together with no campaign behind either, something is telling buyers about you.
Then add “How did you hear about us” to your demo form. AI shows up in those answers long before it shows up in analytics.
None of it costs money and all of it survives a conversation with your CFO.
One warning before you start. None of these three is proof on its own. Direct traffic rises for a dozen reasons. Branded search rises after any PR hit. Self-reported attribution is famously unreliable. What makes them useful is agreement. When all three move in the same direction at the same time and no campaign explains it, you have something worth acting on.
Part 7. The audit. Run it once, then monthly.
Everything above is context. This is the work. Five steps, about an hour the first time, fifteen minutes a month after that. You need access to your analytics, Search Console, and your own website. Nothing else.
Step 1. Size your unexplained traffic
This will not tell you your AI traffic. It tells you how much traffic you cannot explain, and whether that pile is growing. That distinction matters, and you should say it out loud when you present the number.
Open your analytics and pull Direct traffic for the last 90 days. In GA4 this is Reports, then Acquisition, then Traffic acquisition, filtered to Direct.
Add landing page as a secondary dimension so you can see where these sessions arrived.
Set the homepage aside, along with any short vanity URLs you print on things. Those are real direct visits.
Everything left is landing on deep pages. Blog posts, product pages, comparison pages, documentation. Nobody types those from memory.
Subtract what you can account for. Email sends, untagged campaigns, links your team shared internally, anything from a printed asset or a QR code.
Write down what remains. That is your unexplained deep-page traffic.
Now repeat the whole thing for the 90 days before that, and compare.
What good looks like. A stable number is fine. What you are hunting for is a trend. If unexplained deep-page traffic grew 40% while nothing else about your marketing changed, that growth came from somewhere you are not measuring.
What to watch out for. Do not call this figure your AI traffic. It is not. Email, dark social, in-app shares, and untagged campaigns all live in here too. Calling it AI traffic is the fastest way to lose credibility with a CFO who knows better.
Step 2. Cross-check against branded search
This is the step that turns a curiosity into evidence. When an AI recommends you, a large share of people do not click the citation. They go and search your name instead. That behavior leaves a mark you can see for free.
Open Google Search Console, then Performance, then Search results.
Filter queries to ones containing your brand name, including common misspellings and any product names people use instead of the company name.
Set the date range to the last 90 days and turn on comparison against the previous period.
Record the change in impressions and in clicks. Impressions matter more here, because they show people searching for you whether or not they clicked.
Put that change next to the change from Step 1.
What good looks like. Both lines rising together, with no campaign, launch, funding announcement, or press hit behind either one. That pattern is the closest thing to proof you can get without paying for a tool.
What to watch out for. Branded search rises after anything that puts your name in front of people. Check your PR calendar and your paid brand campaigns before you claim the lift. If you ran a podcast sponsorship in the window, that is your explanation, not AI.
Step 3. Ask an AI what you cost
Two minutes, and it is the fastest way to find out whether the machines can actually read your most important page. This is where most companies discover a problem they did not know they had.
We ran this with a client and the numbers came back completely made up. The pricing page used a comparison table built out of graphics, and the AI could not read it.
This is the part people miss. AI does not stop and say it does not understand. It fills the gap with whatever it can find and whatever it can reasonably assume. That number is usually wrong.
Buyers were arriving believing a price the company never set.
Nothing in the analytics would have caught this. The page ranked fine the entire time. Traffic looked normal. The only symptom was sales conversations that started from the wrong place.
Open ChatGPT, Perplexity, and Google AI Mode. Use all three, because they source differently.
Ask what your product costs, what plans you offer, and what is included in each.
Then ask it to compare you against your three closest competitors.
Screenshot every answer. These become your baseline.
Mark anything wrong, out of date, or invented.
What good looks like. Prices match your page. Plan names are right. The description of what you do sounds like something you would say about yourself.
What to watch out for. Silence is worse than error. If you do not appear at all in the comparison question, that is a bigger problem than a wrong price, and it will not show up in any traffic report ever.
Step 4. Check whether your key pages can be read
If Step 3 came back wrong, this is where you find out why. Machines read pages differently than people do. A page that looks perfect to a buyer can be almost blank to an AI.
Open your pricing page and view the page source. Search for the word schema, or for the text application slash ld plus json. If nothing comes back, you have no structured data on the page an AI most wants to understand.
Look at your comparison table. If the checkmarks are images, check whether they have alt text. If they do not, a machine sees a grid of unlabeled graphics and cannot tell which plan includes what.
Check your heading order. H1, then H2, then H3, in sequence. Skipped or scrambled headings make a page harder to parse.
Repeat all of this for your three highest-value pages, not just pricing. Usually that means your main product page and your top comparison page.
Run the same check monthly. Structured data disappears silently during CMS updates and redesigns, and nothing in your reporting will flag it.
What good looks like. Structured data present, tables in real text, headings in order, and every image carrying a description that says what it means rather than what it is.
What to watch out for. The page still ranking is not evidence that it is fine. Ranking and citation have come apart. A page can hold position one and be invisible inside an answer.
Step 5. Build the one-page report
Everything above is useless if it lives in a spreadsheet nobody opens. Turn it into a single page you bring to the same meeting every month, so the trend does the arguing for you.
Line one: unexplained deep-page traffic, this period against last.
Line two: branded search impressions, this period against last.
Line three: what people wrote in your “how did you hear about us” field, counted by category.
Line four: whether you appeared in the AI comparison question, and whether the description was accurate.
Line five: any page that lost structured data since last month.
What good looks like. Three months of this and you have a trend line instead of an anecdote. That is the difference between defending a budget and losing one.
What to watch out for. Do not present any single line as proof. Present the pattern. The argument is that several independent signals are moving together, not that one number went up.
What to fix first
The audit will usually surface more problems than you have time for. Work in this order.
First, anything factually wrong in an AI answer about you. A wrong price or a wrong feature list is actively costing you deals right now, and it is usually fixable in an afternoon by making the source page readable.
Second, missing structured data on money pages. Pricing, product, and comparison pages. These are the pages an AI reaches for when a buyer is close to deciding.
Third, absence from comparison answers. This is the slowest to fix and the most valuable. It usually means you are missing from the third-party sources these systems trust, which is a longer piece of work than editing your own site.
Last, everything else. Heading order and alt text on secondary pages matter, but they will not change a quarter.
The AEO Invisibility Score runs the check.