Combined monthly visits to the ten biggest news websites in the U.S. fell 32% between July 2024 and July 2026, according to Press Gazette's analysis of Similarweb data. For publishers, the problem is immediate: the reporting still costs money, but fewer people arrive on the page that pays for it.
The data establishes the loss of visits. Working out how much AI caused, and what a publisher should do about it, takes different evidence. Getting cited by an answer engine does not settle either question.
What the traffic decline measures
Published on August 21, 2026, the analysis separates total visits from organic search referrals:
| Measure | Sites included | July 2026 versus July 2024 |
|---|---|---|
| Combined monthly visits | Ten biggest news sites in the July 2026 U.S. ranking | Down 32% |
| Combined monthly visits | Fifty biggest sites in that ranking | Down 37% |
| Organic search referrals | Those fifty sites | Down 42% |
The list starts with the biggest sites in July 2026 and looks backward. Publishers that fell out of that group are excluded. These are visits to a selected set of websites, not a count of everyone consuming news. The comparison also starts near a traffic peak. Across the top ten, the decline relative to July 2023 was 18%.
That leaves a substantial loss under either comparison. It also leaves the cause unresolved. An aggregate before-and-after comparison cannot isolate an AI interface from changes in audience demand, rankings, publishing output, or other distribution channels.
Another dataset helps put the channel question in context. Chartbeat reports that average weekly pageviews across its network fell about 6% between 2024 and 2025, while pageviews from Google Search fell 34% between December 2024 and December 2025. Its population, metric, and dates differ from the Similarweb analysis, so the percentages cannot be combined. Within Chartbeat's data, however, search deteriorated much faster than total pageviews.
For a publisher, that distinction determines the investigation. A decline concentrated in search needs a different response from a decline across every source. A decline concentrated in explainers needs a different response from readers abandoning an entire publication.
AI can remove the reason to click
Consider a reader asking why a city changed its parking rules. A search result can point them to a local reporter's explanation. An AI answer can put the explanation directly in front of them, with the reporter's article linked underneath. If the summary answers the question, opening that link becomes optional.
This is an illustrative example, but there is evidence for the behavior. Pew Research Center's July 2025 analysis used browsing data from 900 U.S. adults. Traditional results received a click in 8% of visits to Google pages with an AI summary, versus 15% without one. Links inside the summary received clicks in 1% of visits with a summary.
Pew observed March browsing and collected corresponding search results in April. It did not randomly assign summaries to identical searches. The figures show an association in that sample, rather than a universal percentage of traffic lost to AI.
A newer August 2026 preprint by Stephanie T. Wang and colleagues tests the mechanism experimentally. The researchers used a browser extension to assign different Google experiences in a preregistered field experiment with 1,100 participants. They report that removing AI features increased publisher click-through, while routing searches to AI Mode reduced it.
The experiment ran for seven days of treatment. Its sample was not representative of the whole U.S. population, and a Google interface change weakened the intervention that hid AI Overviews. The forced AI Mode condition also differs from gradual adoption. Those limits matter when interpreting the size of the effects.
The evidence supports a practical concern: an answer interface can reduce outbound clicks. It does not assign the entire two-year news traffic decline to that mechanism.
A reader asks what changed and why.
The system finds relevant reporting.
It presents a summary with a source link.
The reader opens the article only if they need more.
The publisher can succeed at being the source and still lose the visit. That is the part an AI visibility report has to confront.
Citations and revenue can move in opposite directions
In the Press Gazette analysis, generative AI supplied an average of 0.4% of total traffic across the fifty sites. Chartbeat also reports that AI chatbots accounted for less than 1% of total pageviews across its network, despite strong growth in ChatGPT referrals. Neither dataset supports budgeting on the assumption that chatbot visits have already replaced lost search traffic.
A citation still has potential value. A reader may recognize the publication, investigate its work later, or decide to subscribe. But each of those is a separate outcome that needs evidence. Counting a source link establishes that the answer attributed something to the publisher. It does not establish that the reader noticed the attribution or that the publisher earned anything from it.
Here is a hypothetical subscription example. A publication gets 100,000 search visits in a month, and 1% become paid subscribers: 1,000 subscriptions. Traffic falls to 70,000 visits. Holding acquisition steady would require the subscription rate to rise to about 1.43%. Better conversion could compensate for fewer arrivals, but the improvement has to appear in the subscription data.
Meanwhile, citations could double without changing that calculation. They might contribute to later demand; they might mostly accompany answers that end the reader's search. The citation count alone cannot distinguish those cases.
The economics also differ by business. A service company may benefit when an answer recommends it and the buyer contacts it later. A publication selling advertising around an article needs someone to load the article. A subscription publication needs evidence that exposure helps acquire or retain paying readers. The same citation metric cannot stand in for all three outcomes.
I would want a publisher's AI report to show what happened after discovery, even when the answer is that the connection remains unknown. An honest gap is more useful than a conversion estimate built from citation counts.
Give the reader something to do at the source
Return to the parking example. A short summary can cover the rule change and its stated rationale. A local publication could also maintain a street-level lookup, publish the underlying documents, track corrections, or let residents subscribe to updates about their neighborhood.
Those are reasons to visit that survive a useful summary. Their value comes from helping the reader complete a task or follow a developing story. They also cost money, so the publication needs to test whether readers actually use them.
I would start by reviewing the pages losing visits. For each one, ask what a reader still needs after receiving a competent paragraph about the subject. Sometimes the answer is very little. Sometimes it is substantial: inspect the evidence, compare alternatives, search a dataset, ask a question, or keep up with a beat over time.
That review should change the editorial investment. A generic explanation that duplicates widely available information has fewer obvious reasons for a separate visit. Original reporting, useful archives, and maintained tools give readers something specific to return to. None is immune to summarization, and originality alone does not guarantee revenue. The point is to make an explicit bet about what the audience values, then measure it.
The destination should fulfill the promise. If an answer cites a housing data page, a reader should land on the relevant data and methodology. If they want ongoing coverage, the newsletter offer should say what coverage they will receive. An unrelated popup or a generic homepage makes it harder to learn whether the underlying reporting persuaded them to stay.
Publishers also have to decide which material they want available for retrieval and under what terms. That is an editorial and commercial choice. A measurement program should reflect it. Treating every blocked fetch as a defect would erase the publisher's own distribution policy.
Keep the measurements separate
A useful report preserves the distinction between a machine requesting a page, an answer citing it, and a person arriving. These observations come from different systems and prove different things.
| Observation | Evidence to retain | What it establishes |
|---|---|---|
| A crawler or fetcher requested a page | Server request, timestamp, URL, status, and verified identity where available | A machine requested that resource |
| An answer cited the publication | Saved answer, source URL, question, provider, collection method, and date | The citation appeared in that collected response |
| A reader arrived from an AI service | Recorded referral and landing page | An attributable visit occurred |
| A reader subscribed or returned | Subscription event or defined return-visit measure | The measured audience outcome occurred |
A request log alone does not prove that the page appeared in an answer. A saved answer does not reveal how many real users saw equivalent responses. Referral tracking does not connect every later direct visit to an earlier citation. Keeping these limits visible prevents a plausible story from becoming an invented funnel.
For answer monitoring, choose a stable set of questions related to the publication's actual coverage. A local newsroom might track questions about recurring civic issues; a specialist publisher might track the decisions its reporting helps readers make. Record the wording and market, then repeat the same panel. If the panel changes, show the change alongside the results.
Our query-selection method at Canonry treats that panel as part of the measurement instrument. It describes the questions sampled. It does not claim to reveal every question the audience asked. A publisher should be able to inspect the answers behind its visibility percentage and see which reporting received credit.
Alongside that panel, compare visits and outcomes for a fixed group of relevant pages. Keep organic search, identifiable AI referrals, and other sources separate. Record publication changes that could affect the comparison. If a story breaks or a major page disappears, the reader of the report needs that context before attributing the movement to AI.
Start with the pages where the economics changed
I would begin with one coverage area where search visits have fallen and where the publication has something readers cannot easily get elsewhere. Establish the traffic and subscription baseline, inspect the answers to a small set of relevant questions, and review the pages receiving citations. That is enough to choose a bounded experiment.
The experiment might be a better route from an explainer to a specialist newsletter. It might be a maintained data page replacing several overlapping articles. State the expected outcome before shipping: more relevant arrivals, more subscriptions per arrival, or more returning readers. Keep the observation window and comparison group consistent enough to interpret the result.
If citations increase while subscriptions remain flat, the result is increased visibility with no measured subscription gain. If traffic falls but retained subscribers increase, report both. The next editorial decision depends on those differences.
The traffic decline makes distribution harder to take for granted. Publishers need to understand how their reporting appears in AI answers, and they need a reason for readers to form a relationship with the source. Fund the work that strengthens that relationship, with evidence that reaches beyond the citation.