AI search referrals: what publishers should measure now

AI-generated conceptual illustration of changing routes to a publication, not measured referral traffic or a forecast.

A fall in search traffic is worth investigating. On its own, though, it doesn't tell you whether you're losing future members, casual visitors, or both.

Follow those visits through to subscriptions, return visits and payments, while keeping enough search detail to work out where the change began. It's tempting to take comfort in a higher conversion rate. But a smaller audience that buys more often can still leave you with fewer customers and less revenue.

The worksheet below is for your own records. It won't tell you what your results ought to be.

What the research measured

Pew Research Center's July 2025 analysis used browsing data from 900 US adults during March 2025. Visits to Google results with an AI summary led to a traditional-result click 8% of the time, compared with 15% for visits without a summary. Links inside the summaries received clicks in 1% of visits to pages containing a summary.

The dates and methodology matter. Pew reconstructed the searches in April, after the March browsing period, and acknowledged that summaries could change over time. This was an observational comparison, not a randomized test of what those same people would have done without the summary. It doesn't show that every publisher lost the same share of traffic.

Pew's June 17, 2026 report on Americans and AI adds evidence about use, but measures something different. It draws on a February 17–23 survey of 5,119 US adults, in which 60% said they ever read AI summaries at the top of search results. That's self-reported reading, not a 2026 click-through study. It cannot update the earlier click percentages or tell you how many referrals your site lost.

These findings are a reason to check your exposure. Decisions about your publication still need your own records.

Don't put visibility, visits and customer value in the same bucket

Google's guidance on AI features and websites says traffic from AI Overviews and AI Mode is included in Search Console's overall Performance report under the Web search type. Calling that whole total "AI traffic" would mislabel search activity beyond those features.

When your analytics captures an identifiable referral from an AI service, record it separately as observed AI referrals. Leave unidentified visits unidentified, even if AI would be a convenient explanation.

Direct traffic needs the same care. In Google Analytics, (direct) / (none) means traffic without a clear referral source. It can include people typing your URL and visits with missing source information. An increase isn't enough to establish that more readers are loyal to the publication.

View full-size image

Set a baseline before changing anything. Compare similar periods and note publication volume, seasonality, campaigns, major site changes and measurement outages. Use the same content groups where possible. A launch month and a quiet month aren't necessarily comparable just because they contain the same number of days.

A worksheet for the audience meeting

Give each row an owner, then add your baseline, current result and uncertainty. These are comparisons to make within your own publication, not industry benchmarks.

Question: How dependent are we on one source?

Record and compare: Each source's share of observed new visitors, alongside its absolute count

What it can help you decide: Whether acquisition depends too heavily on one route

Check before acting: Keep unknown sources visible; a share can rise while counts fall

Question: Where did search change?

Record and compare: Search Console clicks and impressions by page group, country and device

What it can help you decide: Which section needs investigation first

Check before acting: Review indexing and site changes before attributing the fall to AI

Question: Do arrivals join the newsletter?

Record and compare: Confirmed signups and eligible landing sessions by source and content group

What it can help you decide: Whether the next step fits the reader's reason for visiting

Check before acting: Exclude existing subscribers where identifiable; keep the denominator consistent

Question: Do identifiable readers return?

Record and compare: Registered readers returning within a stated follow-up window

What it can help you decide: Whether a new cohort develops a reading habit

Check before acting: Compare cohorts old enough to complete that window

Question: Do new readers become customers?

Record and compare: New paid customers and the defined acquisition cohort

What it can help you decide: Whether discovery contributes to paid demand

Check before acting: Separate first observed source from later converting visits

Question: What are those customers worth so far?

Record and compare: Collected revenue, refunds and acquisition costs at the same cohort age

What it can help you decide: Whether the channel deserves more investment

Check before acting: Don't compare mature annual customers with last week's arrivals

Show counts beside rates. If conversion improves while paid signups fall, you may have a more selective audience without a better business result. Advertising-led publications also need monetized pageviews and net advertising revenue. Newsletter growth doesn't replace lost advertising income unless the business case explains how.

Keep the meeting report short enough to use. If nobody can say what decision a row affects, move it to the supporting diagnostic detail.

Look at the content groups separately

Group content by what readers use it for: quick reference, detailed instruction, original reporting, product evaluation, or recurring member material. Use labels that fit your publication. There's no need to force every article into those particular categories.

For each group, compare acquisition with what readers do afterwards. A reference page might bring in many people and few immediate signups. Before deciding to stop maintaining it, check whether readers who first arrive there return through another route.

Suppose a specialist publisher sees fewer arrivals on glossary pages while registrations from detailed tutorials remain steady. It could inspect the two groups separately and test a more relevant newsletter invitation on the glossary. This is an example of how to investigate, not evidence that tutorials are safe from referral losses.

Keep other explanations in the same note. The content might be outdated, a template change might have hidden the invitation, or the analytics setup might have changed. A before-and-after chart won't distinguish those causes on its own.

View full-size image

Test a change you can evaluate

Choose one content group, one change and a review date. You might test a newsletter invitation promising the next useful lesson instead of asking readers to "stay updated." Write down the result you expect and what would make you stop the test.

Be careful about attribution. If your records support it, you can say, "Readers first observed on these articles later subscribed." Saying the articles caused those subscriptions requires a stronger design. Use a controlled test where practical; otherwise record the competing explanations.

Crawler policy needs a separate decision. AI crawler controls are not one switch explains why restricting training and preserving search discovery require different checks. Blocking a crawler doesn't tell you whether the visits you lost were commercially useful.

At the next review, bring the worksheet and a written decision to maintain, investigate, test or stop. Note which missing data could change your mind. You don't need perfect attribution to run a useful publishing business, but the budget needs more than an assumption that every missing visit is an AI loss or every remaining visitor is a future member.