Chapter 8: AI Traffic and Business Results

Last updated: November 1, 2026.

In a budget meeting, a screenshot that says “ChatGPT recommended us” works for a while. Sooner or later someone asks, “What did this actually get us?” The visibility measurement from Chapter 7 doesn’t answer that. You have to look separately at whether visibility turns into visitors, and whether those visitors turn into sales, quotes or customers.

That’s harder than it was in classic SEO. Part of AI’s impact shows up in Analytics under its own name, part of it disappears inside other channels, and part of it leaves no trace in any report. The job is to separate the visible part correctly, estimate the invisible part with a reasonable assumption, and keep the two apart in your reporting.

1. Traffic and business results are not the same thing

One of our e-commerce clients, a store selling volleyball and basketball shoes, once saw traffic fall 10% while revenue rose 180%. Anyone looking only at the traffic chart would have reported it as a bad period.

That example has nothing to do with AI traffic, but it sums up the point of this chapter well: traffic is the means, business results are the goal.

In AI search the distinction gets sharper. As I described in Chapter 1, some of the in-between informational visits now stay inside the chat. Fewer people may reach your site, but the ones who do are closer to a decision.

In B2B the gap is even wider. One of our clients in the machinery sector gets 30–40 visits a day, and the machine it sells costs around €20,000. On a site like that, a few hundred visits more or less in a month doesn’t tell you much. A single qualified lead from AI can be worth more than every swing in the traffic chart.

It helps to think about measurement in three layers. The first is visibility: whether AI answers mention you, recommend you and cite you, which was the subject of Chapter 7. The second is visits: how many people come to your site from those answers, which pages they land on and what they do there. The third is business results: how many sales, quotes, calls and new customers come from those visits and from the impact that never shows up in a report.

This chapter covers the last two layers. However good your visibility measurement is, until it’s tied to business results it’s only half an answer for management.

2. Observed, estimated and modeled impact

In Chapter 1 I said we would split AI’s impact into these three types. In my view, the biggest mistake in AI reporting is showing all three in the same table as if they were equally certain.

Observed impact is the most reliable, but it isn’t the whole picture. If a user clicked through from AI to your site and the browser passed that information along, you see it; if they didn’t click, or the information got lost on the way, you don’t. So read observed impact as a floor.

Estimated impact tries to fill in the hidden part, and it has one rule: write down the assumption. If you say “we attributed part of the increase in Direct traffic to AI,” also write which pages, which period and why. The reader may disagree with the assumption, but at least they’ll know what they disagree with.

Modeled impact is the most ambitious. It tries to separate the effect from other channels, seasonality and campaigns. Done right it’s very valuable, but for most businesses it isn’t the first step; we come back to it in section 9.

Below, I note which type of impact each method measures. The same goes for reports: every number in a report should say whether it’s observed, estimated or modeled.

3. Separating AI traffic in Analytics

The first source of observed impact is web analytics. Since most sites in Turkey use Google Analytics 4, the examples here use it; the logic is the same in other tools.

On May 13, 2026, Google Analytics introduced a separate classification for traffic from AI assistants (Google Analytics). If the referrer is a recognized AI assistant, the session’s medium is automatically set to “ai-assistant,” these visits are grouped under the “AI Assistant” channel in the default channel group, and the campaign name shows as “(ai-assistant).”

So without any setup, you can see traffic from AI assistants as its own row in your reports. On June 11, 2026, Google also added a new field called Source Group, with ready-made groupings for newer traffic sources like ChatGPT and Perplexity.

Still, don’t rely on this channel alone. According to Search Engine Journal, Google lists ChatGPT, Gemini and Claude as examples but doesn’t publish the full list of recognized sources (Search Engine Journal). Traffic from a tool that isn’t on the list can still end up in Referral.

That’s why it pays to set up your own custom channel group as well. Google’s help page gives a ready-made example: create a channel called “AI assistants,” match the source with a regular expression, and move that channel above “Referral.” According to the same page, custom channel groups can be applied to reports retroactively (Google Analytics Help), so you can compare the period before May 2026 using the same definition.

Something like this is enough to start with:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai

Review the list once a month. If you spot a new AI tool you don’t recognize in the Referral report, add it.

Once the channel is separated, four things are worth checking every month:

• Which pages they land on. Filter the landing page report by the AI channel; instead of the pages you’d expect, you may find a comparison post, a spec page or an old guide.
• Which platform they come from. Look at ChatGPT, Gemini, Perplexity and Copilot separately; as we saw in Chapter 4, each platform behaves differently.
• What they do on the site: conversion rate, revenue, average order value, form fills, phone clicks.
• How it changes over time. One month’s number says little, so look at the three-month trend.

If a filling machine manufacturer sees that most visitors from the AI channel land on the technical page about filling foaming products, it knows that adding the quote form and service information to that page should come first. Conclusions like that are worth more than reporting the number.

Clicks from AI Overviews and AI Mode, on the other hand, don’t show up separately in Analytics. They sit inside Google organic traffic, and today there’s no way to separate them. We’ll look at what Search Console can show in section 5.

4. Why AI visits show up as “Direct”

Analytics works out where a visit came from using two things: the referrer the browser sends and the campaign tags added to the end of the link (UTM parameters). If neither is there, the visit is recorded as “Direct.”

With AI traffic, that information often gets lost. Mobile apps, in-app browsers and links that are copied and pasted may not carry a referrer. That’s why it makes a big difference when platforms add their own tags to links. ChatGPT adds “utm_source=chatgpt.com” to links in web-sourced answers. In October 2026 Google also started adding UTM parameters to outbound links from Gemini, but according to Search Engine Journal this isn’t documented and it isn’t known when it applies (Search Engine Journal).

October 2026 showed how much we depend on these tags. According to data compiled by Search Engine Journal, around the time GPT-6 rolled out, ChatGPT stopped adding “utm_source=chatgpt.com” to links in its answers (Search Engine Journal). From October 9, ChatGPT sessions on sites tracked by Graphite fell by between 22% and 57%. Across roughly 500 pages examined by Adobe’s Cedric Huesler, ChatGPT visits dropped by about 31,000 a day while Direct traffic rose by about 26,000 a day. According to Huesler, OpenAI confirmed on October 14 that this was a bug, and a fix began rolling out on October 15. OpenAI has made no public statement about it.

A sudden drop in AI traffic doesn’t always mean lost visitors. Sometimes the visitors still come, just untagged. The same report quotes Huesler as saying most ChatGPT clicks arrive without a referrer, which is why the UTM tag has to be counted.

Note incidents like this in your reports. Add an annotation in Analytics for October 8–15, 2026; six months from now, someone looking at that period’s chart without knowing what happened could draw the wrong conclusion. Sessions that landed in Direct that week won’t be corrected retroactively either.

That’s also why it’s useful to track the AI channel side by side with Direct traffic. If, on landing pages that regularly get AI traffic, Direct suddenly rises while AI traffic falls at the same time, there’s most likely a measurement problem.

There’s no way to fully separate the AI visits hidden in Direct traffic, but you can estimate them. Someone who lands directly on a deep technical page or a blog post, rather than the home page or a campaign page, usually didn’t type the address from memory. Some of those visits come from AI apps, emails or messaging apps. A number you arrive at this way is estimated impact, not observed, and the report should say so.

5. AI Overviews and AI Mode in Search Console

For traffic from Google’s own AI features, the place to look is Search Console, but it has its limits too.

Google states that visibility in AI Overviews and AI Mode is counted under the “Web” search type in the Search Console performance report (Google Search Central). Those clicks and impressions are mixed into your overall search data, and there’s no filter in the performance report to separate them.

A new report was added in 2026. On June 3, 2026, Google announced that Search Console had started showing new information about how pages appear in generative AI features, and on August 31, 2026 it opened this to all sites (Google). According to Search Engine Journal, the report shows impression counts for generative AI features in AI Overviews, AI Mode and Discover, broken down by page, country and date, plus device for Search. There’s no click data, and sites without enough AI impressions may not see the report (Search Engine Journal).

Without clicks, this report says nothing about business results. What it does tell you is which of your pages are used in Google’s AI answers, and in which countries. You can use it in this order:

1. List the pages with the most AI impressions.
2. Look at the Web performance of the same pages: how have clicks, impressions and click-through rate changed over the last six months?
3. Check the conversions for those pages in Analytics.

If a page’s AI impressions are rising while its click-through rate falls, users may be finding what they need in the answer. If conversions are falling too, there’s a problem. If conversions hold steady or rise, the clicks you lost were most likely clicks that weren’t turning into sales anyway. That inference also counts as estimated impact.

After AI Overviews launched in Turkey, none of our clients saw a serious drop in clicks; we saw only small declines. To see where a decline is concentrated, the split by page type I suggested in Chapter 1 helps here too: track blog, category, product and brand queries separately.

Google makes a claim on this as well. In the same document, it writes that it has seen clicks from results pages with AI Overviews be “higher quality,” meaning users tend to spend more time on the site. That’s Google’s own observation; treat it as a claim to check on your own site.

Outside Google, the AI performance report in Bing Webmaster Tools, mentioned in Chapter 2, gives you citation data for Copilot. It has no clicks either. Together, the two reports show which of your pages make it into AI answers on Google and Bing; whether that shows up in the business is something we look for in Analytics and the CRM.

6. Branded search: the trail of the user who didn’t click

Remember the scenario from Chapter 1. A purchasing manager sees your brand in ChatGPT but doesn’t click the link. That evening they type your brand name into Google and reach your site from there. Analytics records the visit as branded organic search; the AI answer that started the decision appears in no report.

So it makes sense to start looking for AI’s indirect impact in branded queries. If AI mentions and recommends your brand more often, you’d expect the number of people searching for it to rise over time.

Tracking this got easier in 2026. In November 2025 Google added a branded queries filter to Search Console (Google Search Central), and in March 2026 it was announced that the filter had opened to all eligible sites. Search Engine Journal compiled the details (Search Engine Journal):

• The filter is in the performance report and splits queries into branded and non-branded.
• Google decides automatically which queries count as branded. The brand name, misspellings and product or service names associated with the brand can count as branded; you can’t add your own list.
• The filter doesn’t appear for subproperties (for example site.com/blog/) or for sites with low impressions.
• On some properties the branded/non-branded split only shows from around February 21, 2026; the data isn’t applied retroactively.

That last point is an annoying limit for Turkey. AI Overviews and AI Mode started rolling out in Turkey on February 18, 2026, so on some sites you may not be able to compare “before and after AI Overviews” with this filter. For a comparison that goes further back, you’ll need to separate branded queries yourself with a query filter.

With branded search, look at three things. First, the trend in branded impressions and clicks over three and six months, not one. Then the words people search alongside your brand: queries like “X machinery foaming product,” “X machinery service” or “X machinery price” can show that someone saw you somewhere and is looking for details. And finally, how this data moves together with the visibility you measured in Chapter 7. If branded search rises as visibility rises, that’s a meaningful sign.

But be careful. Attributing the entire rise in branded search to AI is the easiest and the most wrong way to report it. Ads, trade shows, press coverage, social media, seasonality and your sales team’s work all affect branded search too. Branded search data alone isn’t proof; it’s one part of estimated impact. Write down what else changed in the same period next to the report.

7. The CRM and “Where did you hear about us?”

What Analytics can’t see, the customer can tell you. In Chapter 1 I said this is the simplest method and, in my view, the most valuable.

This approach, based on what the customer tells you, is called self-reported attribution. Analytics follows the trail of a click; this method asks the person what they remember. The two measure different things, so read them together. To set it up:

• Add the question “Where did you first hear about us?” to your quote and contact forms. Alongside Google search, include ChatGPT, Gemini, Google AI Mode, Copilot, Perplexity and Claude as options, and leave a free-text field at the end.
• Don’t forget phone leads. In B2B a large share of leads come in by phone; have the sales team ask the same question on the first call and log it in the CRM with the same options.
• When someone says “I searched on Google,” ask one more question: “What did you search for, and where did you first see us?” Anyone searching for your brand has already heard of you somewhere, and that follow-up often reveals the real source.
• Use hidden form fields to pass the visitor’s first source and first landing page into the CRM. That way the source Analytics saw and the source the customer named sit side by side in the same record.
• Follow the lead all the way to the sale. Only the CRM can tell you how many AI-sourced leads turned into quotes and how many into sales.

In the filling machine example, it works like this. Say a purchasing manager asks ChatGPT about a machine for filling 3,000 bottles a day of a foaming product, with a budget of around €25,000. They see your brand in the answer, search for it on Google two days later, land on your site and fill in the quote form. Analytics attributes the lead to branded organic search. Without the question on the form, you’d never know the sale started with AI.

The method has limits. People misremember sources or tick the first option on the form and move on. That’s why self-reported data goes next to Analytics data, not in place of it. Still, in businesses like B2B, where leads are few and high in value, I think it’s the strongest way to see AI’s impact.

8. Reading conversion rates correctly

There’s no single answer to whether visitors from AI convert better.

Across our clients, traffic from AI platforms is usually under 10% of total traffic, but those visitors convert at around 6–7%. Given that e-commerce conversion rates generally sit around 1%, that’s a big difference.

Published studies give different results. In data Ahrefs shared for its own site in June 2025, visitors from AI search made up 0.5% of traffic over the previous 30 days but brought in 12.1% of signups. Ahrefs read that as a conversion rate 23 times higher than traditional organic search, while noting that the data came from a single site and that it doubted the ratio would hold as AI search grows (Ahrefs).

In an analysis 180 Marketing published in February 2026, covering 2025 GA4 data from 94 e-commerce sites, visits from ChatGPT converted at 1.81% and visits from non-branded organic search at 1.39% (180 Marketing).

An academic study by Kaiser and Schulze, published in Marketing Science in 2026, looked at 973 e-commerce sites between August 2024 and July 2025. ChatGPT traffic was under 0.2% of total traffic. ChatGPT’s conversion rate and revenue per session were higher than paid social and lower than every other established channel. The researchers also note that ChatGPT conversion rates rose over time and that AI may add more value in information-heavy, complex purchases (Frankfurt School).

These results look contradictory, but they measure different things. A software company’s signup conversion and an e-commerce site’s sales conversion aren’t the same. Comparing ChatGPT with non-branded organic search gives a different result than comparing it with all channels. And in a low-volume channel, a handful of sales can swing the rate a lot.

Instead of borrowing someone else’s multiplier, look at your own data. When you do:

• Compare like with like. Benchmark AI traffic against non-branded organic traffic, not total organic; branded searchers already know you and naturally convert very well.
• Look at the same page types. Blog traffic and product page traffic convert differently to begin with, so compare by landing page type.
• Wait for enough data. For a channel that brings 5 sales a month, look at at least three months of combined data before reading anything into the rate.
• Measure value, not just the rate: revenue per session, average order value, and in B2B, quote value and the value of closed deals.

9. When you need modeled impact

For most businesses, observed and estimated impact give a good enough picture. But if a serious budget is going to AI, management will ask, “Wouldn’t this increase have happened anyway?”

Answering that takes methods that separate the effect from other factors. In a controlled comparison, you compare a group of pages you improved for AI with a similar group you left alone; the difference between the two is one of the cleanest signs of your work’s effect. In time series analysis, you build a model from past data that predicts what would have happened if you’d done nothing, then compare it with what actually happened. Marketing mix modeling models spend and results across all channels together to estimate each channel’s contribution. That last one usually suits brands with large budgets and a long data history.

A controlled comparison could work like this for a filling machine manufacturer. On 5 of its 10 product pages it adds a capacity table, suitable product types, service regions and delivery times, and leaves the other 5 untouched. Three months later it compares AI visibility, AI traffic and quotes between the two groups. That’s much stronger evidence than “we improved the pages and demand went up.”

None of these methods gives a definitive result. When you report a modeled effect, give a range instead of a single number and spell out the method. Get observed and estimated impact right first; think about modeling once that foundation is in place and you have enough data.

10. A one-page monthly report

The monthly report that pulls all of this together shouldn’t run past one page. Put the type of impact in square brackets next to each line:

• Sessions, conversions and revenue from the AI channel [observed]. GA4’s AI Assistant channel and your own custom channel, broken down by platform.
• The 10 pages that get the most AI traffic [observed]. Their conversion rates and the change from the previous month.
• Pages with AI impressions in Search Console [observed]. The impression trend and those pages’ Web click-through rate.
• Branded searches [estimated]. Impression and click trends, along with anything else in the same period that could have affected branded search.
• Direct traffic to deep pages [estimated], side by side with the AI channel.
• “Where did you hear about us?” answers [observed, self-reported]. The number of leads that chose an AI tool and how many of them turned into sales.
• Notes for the month: tagging bugs, platform changes, campaigns, trade shows. Anything the person reading the numbers needs to know.

When you present the report to management, say clearly that the observed number is a floor, not a ceiling. AI’s real impact is probably larger than what you can see, but when you say how much larger, you have to show the assumption too.

This report will be the business-results part of the measure, diagnose, fix cycle we build in Chapter 17.

11. What to bring to the budget meeting

AI’s impact shows up in three places at once: in its own channel, inside other channels and in the customer’s own words. You can separate the first in Analytics. You estimate the second through branded search and Direct traffic, writing down your assumption. The third you only learn by asking the customer.

Bring numbers from all three sources to the budget meeting, and say how certain each one is. Presenting observed impact as the floor, showing estimated impact together with its assumption, and comparing your conversion rate with your own non-branded organic traffic rather than someone else’s multiplier is far more convincing than a screenshot.

Checklist

Sources

• Google Analytics: What’s new in Google Analytics (AI Assistant traffic note May 13, 2026; Source Group note June 11, 2026)
• Google Analytics Help: Traffic-source dimensions (custom channel groups and the AI assistants example)
• Search Engine Journal: Google Analytics adds AI Assistant as default channel group (May 14, 2026)
• Search Engine Journal: Google Gemini adds UTM parameters for referral attribution (October 1, 2026)
• Search Engine Journal: ChatGPT traffic drop appears to be a tracking problem (October 15, 2026)
• Google Search Central: AI features and your website (last updated December 10, 2025)
• Google: New opportunities, control and insights for website owners (June 3, 2026, updated August 31, 2026)
• Search Engine Journal: Google Search Console AI reports rolled out worldwide (August 31, 2026)
• Google Search Central Blog: Introducing the branded queries filter in Search Console (November 2025)
• Search Engine Journal: Google answers questions about Search Console’s branded queries filter (March 12, 2026)
• Ahrefs: Does AI search traffic convert better than traditional search? (June 16, 2025)
• 180 Marketing: ChatGPT traffic converts 31% better than non-branded organic search (February 23, 2026)
• Frankfurt School: New study, where ChatGPT already works in online shopping (June 22, 2026; paper: Kaiser and Schulze, Marketing Science)

Previous: Chapter 7, How to Measure AI Visibility · Table of contents · Next: Chapter 9, How AI Bots See Your Site