How AI Search Changes Science Writing

Google's AI Overviews and tools like ChatGPT search and Perplexity now answer a growing share of science questions directly on the results page or inside a chat window, often without the reader ever visiting the outlet that originally reported the finding. Global Google search traffic to publishers fell by roughly a third over the year to November 2025, according to Press Gazette's analysis of industry data, and health content, one of the categories science writers work in most, was the single most affected topic vertical. That shift is not just a traffic problem. It is reshaping what a science story is for, who gets credit for original reporting, and whether the economics that fund careful science journalism can keep functioning at all.

The traffic numbers are no longer speculative

For the first two years after generative AI search tools launched, the effect on publisher traffic was mostly argued about in conference panels and anecdotes from individual editors. That is no longer true. Multiple independent analyses, using different methodologies and different sample sets, now point in the same direction.

Press Gazette's 2026 review of publisher data found that search referral traffic fell by 33% globally over the year to November 2025, with a 38% decline in the United States and a smaller but still substantial 17% decline in Europe. A separate Authoritas study, also reported by Press Gazette, found that pages losing their normal top organic position to an AI Overview saw clickthrough rates fall by roughly half. Chartbeat data covering 2,500 sites, cited by Nieman Lab in 2026, put the US decline even higher, at 38%, with a global figure near 33%. These are not the same study reaching the same number by coincidence; they are several different measurement approaches converging on the same order of magnitude.

The mechanism behind the drop is measurable at the level of individual search results. Press Gazette reported that when an AI Overview appears above the traditional list of blue links, the clickthrough rate on the first organic result drops from around 27% to about 11%, a decline of nearly 60%. In October 2025, the Daily Mail's SEO director publicly stated that when an AI Overview activates on a search results page, the site's average clickthrough rate falls between 80% and 90% compared to searches where no AI Overview appears. That is not a modest dip in a secondary metric. It is the loss of most of the traffic a page would previously have earned for ranking near the top of a results page.

Press Gazette's analysis found that health-related search results are supplanted by an AI-written summary about 72% of the time, the highest rate of any topic category studied, ahead of technology (49%), travel (39%), lifestyle (37%), and personal finance (35%). Health and biology explainers, a core beat for science writers, sit at the center of the category AI Overviews displace most often.

Zero-click search turns the reported fact into the product, not the article

The term "zero-click search" describes a query where the user gets a complete, satisfying answer directly on the results page or inside a chat response and never clicks through to any source at all. For most of search engine history, the search results page was a directory pointing traffic toward publishers. AI Overviews and conversational search tools increasingly make the results page, or the chat window, the final destination.

This is a structural change, not a cosmetic one, for how a science article earns its value. Historically, a science writer's article was the unit that got read, shared, and monetized through ad impressions or subscriptions. If an AI system extracts the one sentence stating a study's headline finding and never sends the reader to the piece explaining the methodology, sample size, or limitations that made that finding credible in the first place, the article's informational content has been separated from the economic vehicle, the actual page view, that used to carry it.

Nilay Patel, editor-in-chief of The Verge, described this endpoint in 2024 as "Google Zero," the moment search stops sending meaningful traffic to third-party sites at all. Nieman Lab's continued 2026 coverage of publisher traffic shows that framing has moved from a warning to a documented trend rather than a prediction. Some publishers, per Nieman Lab's July 2026 reporting, are now weighing whether to block Google's crawlers from indexing their content entirely, a step that would have been unthinkable a few years earlier, given that it would also remove the site from traditional search results.

Does zero-click search mean fewer people are learning about science?

Not necessarily. It means fewer people are learning about science by visiting a publisher directly. The information itself may still reach a reader through an AI summary, but stripped of the surrounding methodology, caveats, and named sourcing that distinguish careful science reporting from a flattened one-line takeaway.

The AI system still needs the reporting it displaces

The tension at the center of this shift is that AI answer engines depend on the exact reporting infrastructure their summaries are drawing traffic away from. A language model does not independently interview a study's lead author, obtain an early copy of a paper under embargo, or notice that a press release's framing overstates what the underlying data actually shows. Those steps require a human reporter with editorial judgment and, usually, a salary funded by the traffic and subscriptions that AI summaries are now diverting.

Columbia Journalism Review's Tow Center for Digital Journalism has documented this dependency directly. Its analysis, published as "Journalism Zero: How Platforms and Publishers are Navigating AI," frames the core risk plainly: AI systems generate polished, confident-sounding summaries built on top of professionally reported journalism, but if the revenue that funds that reporting keeps eroding, the pipeline of original, verified information feeding those same AI systems eventually thins out too. It is a closed loop with a weak link, and science journalism, which depends on reporters who can read a primary paper, contact its authors, and contextualize its limitations, is more exposed to that weak link than commodity news that simply restates a press release.

This has already produced concrete business responses rather than just editorial hand-wringing. Some large publishers, including the Associated Press, have negotiated direct licensing deals with AI companies. Others, including The New York Times, have pursued litigation, arguing that AI systems are trained on and now compete directly with the reporting that produced their answers, according to Poynter's 2026 coverage of the licensing debate. A separate group of publishers is exploring collective bargaining approaches, negotiating jointly rather than individually, per reporting from the Reuters Institute for the Study of Journalism. None of these responses has settled the underlying economic question, but their existence confirms that publishers now treat this as a structural business problem, not a temporary traffic dip that recovers on its own.

Accuracy failures compound the traffic problem

A separate line of research complicates the picture further: even when AI systems do send a reader an answer instead of driving them to the original article, that answer is frequently wrong or misattributed. The Tow Center's March 2025 study, run by Klaudia Jaźwińska and Aisvarya Chandrasekar, tested eight AI search tools by feeding each one an excerpt from a real news article and asking it to identify the correct headline, publisher, date, and URL. Collectively, the tools answered incorrectly more than 60% of the time across 1,600 test queries, and Perplexity, the best performer among the eight tools tested, still answered roughly 37% of queries incorrectly.

The BBC ran a parallel investigation into how AI tools handle its own journalism and reported, per Nieman Lab's coverage, that 51% of AI-generated answers built on BBC reporting had significant issues, 19% introduced outright factual errors not present in the original story, and 13% of quotes attributed to BBC articles were either altered or did not appear in the cited piece at all. For science stories specifically, where a single misquoted number, say, a confused correlation for a causal claim, or a mangled sample size, can flip a study's actual finding into something the researchers never claimed, that error rate is not a minor inconvenience. It is a mechanism for spreading exactly the kind of overstated or distorted science claim that careful science writing exists to prevent.

The BBC's internal audit, reported by Nieman Lab in February 2025, found that 13% of quotes AI tools attributed to BBC articles were either altered from the original wording or did not appear in the cited article at all, a distortion rate that undercuts the idea that AI summaries are simply a faster, equally reliable version of the reporting underneath them.

Science writing now has to survive being extracted, not just being read

Given both the traffic reality and the accuracy risk, the practical response for a science writer is not to abandon careful reporting for AI-friendly formatting tricks. It is to write in a way that survives extraction: if an AI system is going to lift one sentence out of an article regardless of what the writer intended, that sentence needs to be the accurate one, stated with enough precision and named sourcing that a model pulling it out of context still represents the finding correctly.

This means naming the actual study, its sample size, its methodology, and its named authors close to the specific claim, rather than burying that context in a later paragraph an AI system is less likely to retrieve. It means stating explicitly what a finding does not show, correlation instead of causation, an animal model instead of confirmed human effect, a preprint instead of a peer-reviewed result, in the same sentence as the claim itself, not as a caveat three paragraphs later that a retrieval system might never surface. A vague, hedged sentence like "some research suggests a link" gives both a human reader and an AI system almost nothing to check against; a precise sentence like "a 2024 randomized trial of 240 participants published in JAMA found X, though the effect size was small and the study did not track outcomes past six months" gives both a fact worth extracting and the guardrail that keeps the extraction honest.

Science writers who cover embargoed research face an added wrinkle here. Newsrooms that ran through the Tow Center's and Nieman Lab's newsroom-AI reporting have flagged press-release-driven science coverage, where a study's own institutional press office frames the "groundbreaking" hook, as especially prone to distortion once an AI system compresses it further. A press release already tends to overstate significance; an AI summary compressing that press release, or a news article based on it, compounds the overstatement rather than correcting it. The corrective is the same discipline good science writers already practice: read the actual paper, not just its press release, and write the sentence that would still be true if someone quoted it completely out of context.

Traditional search discovery versus AI-answer zero-click discovery, for science content

Dimension	Traditional organic search click	AI-answer zero-click summary
Where the reader lands	The publisher's own article page	The search results page or a chat window
Who controls surrounding context	The original writer and editor	The AI system's synthesis step
Typical clickthrough on a top result	Roughly 27% before an AI Overview appears (Press Gazette, 2026)	Roughly 11% on the same query once an AI Overview activates
Revenue signal to the publisher	Page view, ad impression, potential subscription conversion	Often none; the query is resolved without a visit
Risk of factual distortion	Errors are the original outlet's own, subject to correction on that page	Compounded risk from a separate synthesis layer, per Tow Center and BBC findings
Incentive for original reporting	Directly rewarded by traffic tied to the specific article	Indirectly dependent on reporting the AI system did not itself produce

The clickthrough figures come from Press Gazette's 2026 reporting on Google AI Overviews; the pattern illustrates why zero-click discovery weakens the traffic-to-revenue link that funded original science reporting.

What people get wrong about this shift

A common reaction is to treat this as a simple traffic story that will resolve itself once publishers adapt their SEO strategy. That undersells the size of the problem. Nieman Lab's 2026 reporting found that even as Google referral traffic collapsed, ChatGPT referral traffic to publishers grew by more than 200% year over year, but AI-tool referrals still accounted for less than 1% of total publisher pageviews. The replacement traffic is not remotely close to filling the gap left by the decline in traditional search referrals, at least as of mid-2026.

A second misconception is that AI Overviews and chat-based search summaries are simply a faster version of what a human reader would have found by clicking through and reading the actual article. The Tow Center's and BBC's error-rate findings argue against that framing directly. A summary that gets the source, the quote, or the finding wrong is not a compressed version of accurate reporting; it is a different, less reliable product that happens to look authoritative because it is delivered in confident, complete sentences.

A third misconception, more specific to science coverage, is that this pressure only affects breaking-news outlets and not explainer or evergreen science writing. In practice, evergreen science explainers are exactly the content most likely to get summarized rather than clicked through, precisely because the underlying facts (how a vaccine works, what a p-value means, why a study used a control group) are stable enough that an AI system can extract and restate them with relatively little risk of the answer going stale. That makes evergreen explainer writing, not investigative scoops, the category most structurally exposed to zero-click substitution.

Licensing deals help the largest publishers, not specialist science outlets

The headline AI licensing deals, agreements between companies like OpenAI and large newsrooms, get covered as if they solve the underlying revenue problem for journalism generally. They don't, and the reason matters specifically for science writing, most of which is produced by outlets that are small or mid-sized relative to the wire services and major dailies signing those deals.

A large commercial publisher can negotiate a licensing agreement because it has millions of cleared, well-tagged articles, a dedicated legal and licensing team, and enough archive volume to be worth an AI company's negotiating time. A specialist science publication, a nonprofit newsroom, or a single-topic health outlet typically has none of those three things. Nieman Lab's December 2025 analysis put this plainly: publishers overall will see no meaningful AI licensing revenue, because the market pays for scale and negotiating leverage, and most science and specialist outlets structurally have neither. A follow-up Nieman Lab report in May 2026 described the emerging licensing market as putting publishers in a "double bind": sign a deal and hand over training material that helps AI systems answer more queries without a click-through, or refuse and get scraped anyway with no compensation at all, since large language models are trained on the open web whether or not a publisher consents.

Some collective responses are emerging specifically to address this gap. Publishers' Licensing Services (PLS), a UK collective licensing body, opened a scheme in 2026 aimed at small and mid-sized publishers, including magazines and specialist digital outlets, that lets them license content to AI platforms as a group rather than negotiating alone, according to Press Gazette's coverage of the scheme. The logic mirrors how music-licensing collectives work for songwriters who individually have no leverage with a streaming platform: pooled negotiating power substitutes for scale that no single small outlet has on its own. Whether collective licensing produces revenue that actually matters to a specialist science newsroom's budget remains unproven as of mid-2026, but it is the clearest structural response so far to a problem that direct one-on-one licensing deals, by design, only solve for outlets large enough to be worth an AI company's time.

There is one narrow advantage specialist and science-focused outlets may hold in this market that general-interest publishers don't. AI companies training on the open web already have broad access to mainstream news coverage; what they have less of is well-reported, primary-source science journalism grounded in named researchers, specific study data, and technical accuracy that a model cannot easily reconstruct from a press release alone. That scarcity value is unproven as a source of actual licensing revenue for any specific outlet, but it is the only lever a small science publication has that a large wire service, whose archive AI companies can already approximate from thousands of other sources, does not.

Why don't AI licensing deals help most science publishers the way they help major newsrooms?

Licensing deals require negotiating leverage that comes from archive scale and legal resources most science-focused outlets don't have individually. Nieman Lab's analysis found the broader publisher market will see no meaningful AI licensing revenue for this reason, leaving collective schemes like the UK's PLS licensing pool, aimed specifically at small and mid-sized publishers, as the clearest available alternative so far.

The takeaway

AI search has not eliminated the need for careful science journalism; if anything, the documented accuracy gaps in AI-generated summaries make that reporting more necessary, not less. What it has broken is the traffic mechanism that used to fund it: a reader used to have to visit a publisher's page to get an answer, and now, a growing share of the time, they don't. Science writers who want their reporting to survive that shift need to write claims precise and self-contained enough to stay accurate even when a machine extracts one sentence out of context, while the business side of journalism keeps negotiating, through licensing deals, lawsuits, and collective bargaining, over who gets paid for the reporting an AI system relies on but no longer has to send anyone to read.

---

Read next:

• Why AI search tools cite sources differently
• How to spot AI-written science fluff
• Why headlines overstate breakthrough science

FAQ

Q: How much has AI search actually reduced traffic to publishers?

Press Gazette's 2026 analysis found global search referral traffic to publishers fell roughly a third over the year to November 2025, with a 38% decline in the US. Chartbeat data covering 2,500 sites found a similar 38% US decline. Multiple independent measurements now converge on the same magnitude.

Q: Which science content is most affected by AI Overviews?

Health content is the most affected topic category, with AI-written summaries supplanting the traditional search result about 72% of the time, according to Press Gazette's reporting, ahead of technology, travel, lifestyle, and personal finance content.

Q: Are AI search summaries as accurate as the original reporting they're built on?

Not reliably. A March 2025 Tow Center for Digital Journalism study found eight AI search tools answered source-identification queries incorrectly more than 60% of the time across 1,600 tests, and a separate BBC audit found 13% of quotes attributed to its articles were altered or did not appear in the cited piece.

Q: Is AI referral traffic replacing what publishers lose from declining search traffic?

Not yet, as of Nieman Lab's 2026 reporting. ChatGPT referral traffic to publishers grew over 200% year over year, but AI tools still accounted for less than 1% of total publisher pageviews, far short of offsetting the traffic lost from traditional search.

Q: How should science writers adjust their writing given AI extraction?

By naming the specific study, sample size, and methodology close to the claim itself, and stating limitations (correlation versus causation, preprint versus peer-reviewed, animal model versus human trial) in the same sentence as the finding, so the claim stays accurate even if an AI system extracts it without the surrounding paragraph.

Q: Do AI licensing deals solve the revenue problem for science publishers?

Mostly not for small or specialist outlets. Nieman Lab's December 2025 analysis found the broader publisher market will see no meaningful AI licensing revenue, because deals reward archive scale and negotiating leverage that most science-focused newsrooms lack individually. Collective licensing schemes, such as the UK's PLS pool for small and mid-sized publishers, are an emerging alternative, but unproven as a meaningful revenue source as of mid-2026.

Q: Can a science publisher just block AI crawlers instead of negotiating?

Some are considering it, per Nieman Lab's 2026 reporting, but it carries a tradeoff: blocking an AI company's crawler also typically means opting out of the same company's search or answer products entirely, and it does nothing to stop scraping by AI systems that don't respect crawler-exclusion signals in the first place.