Signal Harbor Weekly
Imagine a software buyer looking for a new competitive intelligence platform.
Instead of opening Google and clicking through ten results, they ask ChatGPT:
What are the best competitive intelligence platforms for a mid-sized B2B marketing team?
ChatGPT recommends three companies and explains the strengths and weaknesses of each.
The buyer reads the answer, closes the tab, and moves on with their day.
Two days later, they remember one of the companies. They search the brand’s name on Google, visit the website, read a few reviews, and request a demo.
In the company’s analytics, that buyer will probably be credited to organic search.
Google gets the attribution.
The AI platform that introduced the company, explained the category, and helped place the brand on the buyer’s shortlist receives none.
This exposes a growing gap in digital measurement.
Companies are beginning to track how often they appear in AI answers. But when they look at their analytics, direct referral traffic from ChatGPT, Gemini, Claude, and other AI platforms often appears small.
That can lead to a simple conclusion:
If AI is not sending much traffic, it must not be having much influence.
Emerging research suggests the reality may be more complicated.
Direct AI referral traffic may capture only part of AI’s role in discovery and consideration. To understand that role, companies have to look beyond the immediate click and examine what happens afterward.
AI often answers without sending the user anywhere
Traditional search was largely built around referrals.
A user entered a query, reviewed the results, and clicked through to another website.
AI search changes that relationship.
The platform can now explain the topic, compare alternatives, summarize reviews, and recommend a company without requiring the user to leave the conversation.
A recent academic preprint analyzing U.S. desktop clickstream behavior found that ChatGPT produced a clean outbound referral click in only 5.2% of conversation sessions. Traditional Google queries produced clean referrals at a much higher rate.
The same research found that most active ChatGPT households in the dataset did not generate a single clean outbound referral during the study period.
That does not mean every AI interaction influences a future purchase.
It does mean direct referral traffic is an incomplete measure of what happened inside the conversation.
A company could appear repeatedly in relevant AI answers without receiving many direct clicks. The platform may have already answered enough of the user’s question for the session to end there.
The influence may appear later
The harder question is what the user does after the AI conversation.
A recent observational working paper connected opt-in users’ AI conversations with their later browsing behavior.
The researchers examined what happened when an assistant recommended a brand to someone who had not recently engaged with that company.
In the following days, those users became more likely to:
search for the brand by name on Google;
visit the brand’s own website;
visit a brand-specific page on a third-party retailer.
The study did not observe completed transactions, and it cannot prove that the AI recommendation caused every later action.
But the pattern is important.
It suggests that an AI recommendation can act as an upstream touchpoint, while the later behavior appears under another channel.
A buyer may discover the company through ChatGPT and then return through:
branded search;
direct traffic;
a review platform;
a marketplace;
a separate device;
a conversation with another decision-maker.
Traditional analytics may record the final step while missing the interaction that helped create the interest.
Similarweb has reported a related pattern in its own proprietary research, finding that much of the traffic associated with AI-influenced discovery later arrives through traditional search rather than a direct AI referral.
Similarweb sells digital and AI intelligence products, so its findings should be treated as directional industry research rather than independent academic proof.
Still, the broader point is consistent across the sources:
The channel that receives the visit may not be the channel that introduced the brand.
A mention is not the same as a recommendation
This is also why raw mention counts can be misleading.
An AI system may mention a company because it integrates with another platform, appears in a list, or is relevant to one small part of the answer.
That is different from actively recommending the company as one of the best options for the buyer.
The observational research found that stronger recommendations were associated with more downstream activity than incidental brand mentions.
This matters for measurement.
A company should not only ask:
Did we appear?
It should also ask:
Were we recommended?
Were we presented as a serious option?
Were we described accurately?
Were we positioned for the right customer?
Which competitors appeared beside us?
What reasons did the AI give for choosing one company over another?
A brand can have a high mention rate while still losing the recommendation.
That is why AI visibility cannot be reduced to one share-of-voice number.
Companies also need to avoid false attribution
The opposite mistake is just as dangerous.
AI visibility may influence buyer behavior, but that does not mean every increase in traffic, leads, or revenue should be credited to AI.
One working paper analyzed an Answer Engine Optimization experiment in which ChatGPT referral traffic grew significantly after changes were made to selected pages.
At first glance, the result looked extremely strong.
But the untreated pages on the same website also experienced substantial growth during the same period.
The reason was simple: ChatGPT itself was growing.
A company may see more AI referrals because the platform has more users, not because its optimization strategy worked.
Other factors can also affect branded search, direct traffic, and pipeline:
seasonality;
paid advertising;
public relations;
product launches;
pricing changes;
review coverage;
broader category growth;
changes in competitor activity.
This is why before-and-after comparisons are not enough.
If AI visibility rises and revenue rises, the two may be related.
But that relationship still needs to be tested.
Good measurement should protect companies from both errors:
underestimating AI because direct traffic appears small;
overstating AI because business performance improved at the same time.
Why this still needs to be measured
The absence of perfect attribution does not make AI visibility less important.
It makes disciplined measurement more important.
Companies already have systems for tracking website traffic, paid media, rankings, conversions, leads, and revenue.
What they often lack is a way to measure the AI conversations that may shape which companies buyers investigate in the first place.
Before someone reaches a website, an AI system may have already:
defined the category;
introduced the available options;
explained the differences;
raised concerns about certain providers;
recommended which companies deserve further consideration.
Those decisions happen before the visit.
Traditional analytics usually begin after it.
This is the layer Signal Harbor is built to measure.
We look at whether a company appears in relevant buyer conversations, whether it is recommended or displaced, how accurately it is represented, which competitors appear instead, and which sources shape the answer.
That does not replace web analytics or revenue attribution.
It gives those systems context they currently do not have.
A company may not be able to prove that one AI answer created one sale.
But it can still determine whether it is consistently entering consideration, losing recommendations to competitors, or being described inaccurately across the platforms buyers are using.
Those are meaningful business risks even when the final click is attributed somewhere else.
A more useful way to measure AI influence
A stronger measurement system connects four layers.
1. Exposure
First, determine whether the company appears in relevant buyer conversations.
This includes:
inclusion frequency;
recommendation frequency;
position within the answer;
consistency across AI platforms;
consistency across repeated tests.
One answer is not enough. AI outputs vary across prompts, platforms, time, and individual runs.
2. Representation
Next, determine how the company is being described.
This includes:
factual accuracy;
strengths and weaknesses;
target-customer fit;
sentiment and framing;
competitor comparisons;
the sources shaping the answer.
A company can be visible and still be misrepresented.
It can also be recommended for the wrong reason or to the wrong audience.
3. Downstream behavior
Then look for changes in behavior that may follow AI exposure.
Possible signals include:
branded search volume;
direct website visits;
review-site activity;
comparison-page visits;
demo requests;
detectable AI referral traffic;
increases in high-intent inbound activity.
None of these signals proves causation on its own.
Together, they can help build a stronger evidence chain.
4. Commercial outcomes
Finally, companies should examine whether those behavioral changes connect with:
qualified leads;
pipeline;
purchases;
revenue;
retention;
customer acquisition.
The closer the analysis gets to revenue, the more careful it must become.
AI visibility is an upstream signal.
Traffic, leads, and revenue are downstream outcomes.
They should be connected carefully rather than treated as interchangeable.
What companies should ask
The question is no longer only:
How much traffic did ChatGPT send us?
Companies should also ask:
Are we appearing in the buyer conversations that matter?
Are we being recommended or merely mentioned?
Are buyers being given accurate reasons to consider us?
Which competitors are entering the conversation?
What sources are shaping the recommendation?
Do changes in AI visibility move alongside branded demand or inbound activity?
What other factors could explain those changes?
These questions will not always produce a clean attribution number.
That does not make the measurement useless.
It means the goal should be a credible chain of evidence rather than a convenient story.
The measurement problem is changing
Traffic still matters.
A direct referral from an AI platform can be a valuable signal of intent.
But traffic begins after many of the most important decisions have already started to form.
Before a buyer visits a website, AI may have already defined the category, narrowed the options, explained the tradeoffs, and decided which companies deserve further consideration.
Traditional analytics measure what happens after the visit.
Signal Harbor measures what is happening before it.
That will not create perfect attribution.
But it gives companies something they currently lack: a repeatable way to understand whether AI systems are helping them enter the buyer’s consideration set or quietly handing that opportunity to someone else.
Sebastian Miller
Co-Founder, Signal Harbor
signalharborconsulting.com
sebastian.miller@signalharborconsulting.com
402-306-2213
Research Notes
This newsletter is based on emerging research examining AI recommendations, referral traffic, downstream buyer behavior, and the limits of traditional attribution.
The goal is not to claim that every AI mention produces traffic, leads, or revenue.
The narrower argument is this:
An AI platform may influence which companies a buyer investigates without receiving credit for the later website visit or conversion. At the same time, increases in AI referral traffic should not automatically be attributed to successful optimization.
AI recommendations and downstream behavior
Iannelli, M., & Ai, A. (2026). From prompt to purchase: How AI brand recommendations move consumers on the open web. Scrunch AI. [Preprint]
This observational study connected users’ AI conversations with their later browsing activity.
The researchers found that when an AI assistant recommended a previously unengaged brand, users became more likely to search for that brand by name, visit its website, and visit brand-specific pages on third-party retailer sites.
The study did not observe completed transactions and cannot prove that every recommendation caused the later activity.
It supports the more limited conclusion that AI recommendations may influence downstream behavior that traditional attribution records under another channel.
AI answers and outbound referral traffic
Shi, Q., Zhu, K., & Gu, K. (2026). Answering without referring: How AI search rewrites the web’s economic bargain. Bocconi University. [Preprint]
This research examined how often AI and traditional search platforms send users to external websites.
The authors found that ChatGPT produced a clean outbound referral in only 5.2% of observed conversation sessions, compared with a much higher rate for Google searches.
This supports the newsletter’s argument that direct referral traffic is an incomplete measure of what occurs inside AI-assisted discovery.
It does not prove that every conversation without a click creates downstream commercial value.
Platform growth and false attribution
Watanabe, K., & Nakayashiki, K. (2026). Disentangling answer engine optimization from platform growth: A log-based natural experiment on ChatGPT referral traffic. Glasp Inc. [Preprint]
This study examined changes in ChatGPT referral traffic following an Answer Engine Optimization experiment.
Although optimized pages experienced substantial referral growth, untreated pages also grew as ChatGPT’s overall user base expanded.
The findings suggest that raw referral growth can exaggerate the apparent impact of optimization.
This supports the newsletter’s warning that companies should account for platform growth, seasonality, marketing activity, and other outside factors before claiming that AI visibility caused a business result.
Industry evidence on downstream traffic
Similarweb. (2026). The downstream impact of AI visibility.
Similarweb’s proprietary research suggests that some users exposed to brands through AI later reach those companies through traditional search or other channels rather than through a direct AI referral.
This supports the possibility that conventional analytics may credit the final channel while missing the earlier AI interaction.
Because Similarweb sells digital and AI intelligence products, its findings should be treated as directional industry research rather than independent academic proof.
Conclusion
Direct AI referral traffic is useful, but it does not show the full role AI may play in discovery and consideration.
Companies should measure where they appear, whether they are recommended, how they are represented, what happens afterward, and whether those patterns correspond with meaningful business outcomes.
The goal is not perfect attribution.
It is a stronger and more credible chain of evidence.
This article first appeared in Signal Harbor Weekly, the Signal Harbor newsletter on AI visibility.